阅读说明：本技术 基于虚拟数据和真实数据的机器学习 (Machine learning based on virtual and real data ) 是由 A·J·M·范德威登 毕婧 萨布汉姆·塞特 于 2020-04-23 设计创作，主要内容包括：一种方法,包括：采用基于计算机的软件来模拟过程,以生成关于所述过程的虚拟数据；识别出用于所述过程的真实世界版本的过程参数；提供真实世界传感器以感测与所述过程的所述真实世界版本相关联的参数；在所述真实世界版本被执行时,从所述真实世界传感器接收传感器读数；以及训练机器学习软件模型以基于关于所述过程的虚拟数据、所述过程参数和所述传感器读数来预测所述真实世界传感器的性能。(A method, comprising: simulating a process using computer-based software to generate virtual data about the process; identifying process parameters for a real-world version of the process; providing a real-world sensor to sense a parameter associated with the real-world version of the process; receiving sensor readings from the real-world sensor while the real-world version is executed; and training a machine learning software model to predict performance of the real-world sensor based on the virtual data about the process, the process parameters, and the sensor readings.)

1. A method, comprising:

simulating a process using computer-based software to generate virtual data about the process;

identifying process parameters for a real-world version of the process;

providing a real-world sensor to sense a parameter associated with the real-world version of the process;

receiving sensor readings from the real-world sensor while the real-world version is executed; and

training a machine learning software model to predict performance of the real-world sensor based on virtual data about the process, the process parameters, and the sensor readings.

2. The method of claim 1, wherein simulating a process with a computer-based software application to generate virtual data about the process comprises:

providing a computer-aided design (CAD) model of a part associated with the process; and

virtual data is generated for the process with respect to a machine tool path based on the CAD model using computer-based software.

3. The method of claim 2, wherein the process is a manufacturing process and the machine tool path is used by a machine to manufacture a part represented in the CAD model.

4. The method of claim 3, wherein the manufacturing process is an additive manufacturing process and the machine comprises: constructing a platform; a powder bed located above the build platform; and a heat source configured to deliver heat to various portions of the powder bed to generate the part.

5. The method of claim 4, wherein the data about the machine tool path comprises: powder data, speed data, and pattern data for the heat source during the additive manufacturing process.

6. The method of claim 4, wherein the process parameters used by the machine to perform the real-world version of the manufacturing process comprise: powder data, speed data, and pattern data for the heat source during the additive manufacturing process.

7. The method of claim 4, wherein the real-world sensor is a photodiode sensor configured to sense light in the machine generated by an additive manufacturing process.

8. The method of claim 2, wherein the process parameters are used by a machine to execute the real-world version of the process based on virtual data about a machine tool path.

9. The method of claim 1, further comprising:

the process parameters are preferred using a trained version of the machine learning software application.

10. The method of claim 9, wherein using the trained version of the machine learning software model to prefer the process parameters comprises:

adjusting the process parameters in the process based on information from the machine learning software model to generate desired sensor readings at the real-world sensor.

11. The method of claim 1, further comprising:

employing the machine learning software model to predict performance of the real-world sensor in the process;

comparing the predicted performance to the actual performance of the real-world sensor in the process; and

an anomaly in sensor performance is detected based on the comparison.

12. A system, comprising:

a computer, the computer comprising:

a computer-based processor; and

a computer-based memory storing computer-executable instructions that, when executed by a computer-based processor, cause the computer-based processor to simulate a process to generate virtual data about the process;

one or more machines configured to execute a real-world version of the process based on process parameters associated with the process;

a real-world sensor to sense a parameter associated with a real-world version of the process; and

a machine learning software model trained to predict performance of the real-world sensor based on: virtual data about the process, the process parameters, and sensor readings from the real-world sensors in a real-world version of the process.

13. The system of claim 12, further comprising:

a CAD software reference for generating a CAD model of a part associated with the process,

wherein the computer-based processor simulates the process to generate the virtual data based at least in part on the CAD model.

14. The system of claim 13, wherein the process is a manufacturing process and the virtual data about the process comprises: virtual data for the one or more machines regarding a machine tool path.

15. The system of claim 14, wherein the manufacturing process is an additive manufacturing process and the machine is a 3D printer comprising:

constructing a platform;

a powder bed located above the build platform; and

a heat source configured to deliver heat to various portions of the powder bed to generate the part.

16. The system of claim 15, wherein the data regarding the machine tool path comprises: powder data, velocity data, and for the heat source in the additive manufacturing processPattern data。

17. The system of claim 15, wherein the process parameters used by the machine to perform the real-world version of the manufacturing process comprise: powder data, speed data, and pattern data for the heat source during the additive manufacturing process.

18. The system of claim 15, wherein the real-world sensor is a photodiode sensor configured to sense light in the 3D printer produced by an additive manufacturing process.

19. The system of claim 18, wherein the computer-based processor is further configured to optimize the process parameters for subsequent iterations of the process using a trained version of the machine learning software application.

20. The system of claim 1, wherein the computer-based processor is further configured to:

employing the machine learning software model to predict performance of the real-world sensor in the process;

comparing the predicted performance to the actual performance of the real-world sensor in the process; and

an anomaly in sensor performance is detected based on the comparison.

Technical Field

The present application relates to machine learning, and more particularly, to machine learning based on virtual data and real data.

Background

On the hardware, the sensor values are used for control (e.g., aircraft flight control or manufacturing control). As of today, cyber physical control systems are typically designed using virtual models. Therefore, it is important to accurately predict the sensor value. Furthermore, the manufacturing industry is challenged to produce product parts of acceptable quality early (e.g., for the first time).

In particular, the complexities associated with the additive manufacturing process include a large number of printing parameters and rapidly evolving physical properties that occur at different scales, such as material phase changes and porosity at the melt pool level, as well as distortion and residual stress development at the part level. Typically, it is costly to use physical simulations to predict all quality metrics.

Disclosure of Invention

In one aspect, a method comprises: simulating a process using computer-based software to generate virtual data about the process; identifying process parameters for a real-world version of the process; providing a real-world sensor to sense a parameter associated with the real-world version of the process; receiving sensor readings from the real-world sensor while the real-world version is executed; and training a machine learning software model to predict performance of the real-world sensor based on the virtual data about the process, the process parameters, and the sensor readings. In typical embodiments, the training technique includes supervised training.

In another aspect, a system includes: a computer; one or more machines configured to perform a process; at least one real world sensor to sense a performance parameter of the process; and a machine learning software model. The computer includes: at least one computer-based processor; and a computer-based memory storing computer-executable instructions that, when executed by the computer-based processor, cause the computer-based processor to simulate a process to generate virtual data about the process. The one or more machines are configured to execute a real-world version of the process based on process parameters associated with the process. The real-world sensor senses a parameter (feature) associated with a real-world version of the process. A machine learning software model is trained to predict performance of the real-world sensor based on: virtual data about the process, the process parameters, and sensor readings from the real-world sensors in a real-world version of the process.

In certain implementations, one or more of the following advantages exist.

For example, highly accurate sensor prediction becomes possible. Furthermore, predictions can be made without the need to model the specific operation of the sensor using first principles and without extensive calibration efforts. Highly accurate sensor prediction can be efficiently achieved. Once trained, the predictive models can be used in a variety of ways, including: to optimize machine process parameters, to detect sensor anomalies, and/or to support subsequent system simulations.

Furthermore, particularly in certain embodiments, the disclosed technology provides for fast training and fast prediction of sensor performance. This is particularly useful where the available time is limited but accuracy and precision is required.

Furthermore, in certain embodiments, the disclosed systems and techniques provide an easy-to-use tool that can help quickly achieve the print quality of any newly designed part. Further, in certain embodiments, optimized machine parameters in a machine-readable format are generated directly by means of the disclosed systems and techniques without involving complex simulation analysis to model complex processes, such that, for example, optimized print parameters can be obtained.

Other features and advantages will become apparent from the description, the drawings, and the claims.

Drawings

FIG. 1 is a schematic diagram that includes a method of training a machine learning software model 106 with real and virtual data about a process to be able to predict the behavior of a sensor in a real version of the process.

Figure 2 is a schematic diagram of a more detailed representation of the method shown in figure 1.

Fig. 3 is a schematic cross-sectional view of an exemplary three-dimensional (3D) printer.

FIG. 4 is a schematic diagram of an embodiment of the method represented in FIG. 2 specific to a 3D printing environment.

Fig. 5 shows several examples of computer-aided design (CAD) models with different geometries.

FIGS. 6-9 show that provided as Dassault Syst SE

A screenshot of a powder bed manufacturing application that is part of a software platform.

Fig. 10 shows a top view of a powder bed where a part is being printed in a 3D printer and a schematic of the path that a laser tool will travel along the upper surface of the powder bed in a rectangular print layer.

Fig. 11 and 12 are schematic views of the powder bed of fig. 10 showing the laser position and the area of interest circles at the very beginning of the scanning process (in fig. 11) and towards the end of the scanning process (in fig. 12).

FIG. 13 illustrates a single layer test data prediction scatter plot that does not include physics-based features during training.

FIG. 14 illustrates a single layer test data prediction scatter plot including physics-based features during training.

Fig. 15 shows experimental verification of photodiode sensor readings for a single layer before and after the print optimization process.

FIG. 16 is a schematic diagram illustrating an exemplary flow of a machine learning and print optimization framework.

FIG. 17 is a schematic diagram of a process for sensor prediction and optimization of machine parameters layer by layer and in real time in an additive manufacturing environment.

Fig. 18 is a schematic diagram showing an exemplary flow of sensor abnormality detection.

FIG. 19 is a schematic diagram of a computer network or similar digital processing environment.

FIG. 20 is a schematic diagram of an exemplary computer in the computer network or similar digital processing environment of FIG. 19.

Like reference characters refer to like elements.

Detailed Description

The present application relates to systems and processes related to predicting sensor behavior in a process performed by a real-world machine or system (hereinafter "machine"). The process may be a manufacturing process or any other type of process that causes at least one sensor to sense a characteristic associated with the process. The machine may be any machine or system configured to perform the process and including a sensor. For example, in the additive manufacturing industry, three-dimensional (3D) printers typically include photodiode sensors to sense light at or near the molten pool of the 3D printer. The light level measured by the photodiode sensor is related to the amount of heat transferred into the molten bath. The 3D printing industry typically considers data collected by photodiode sensors in a 3D printer as part quality indicators for printed parts. Other examples of processes to which the systems and methods disclosed herein may be applied include autonomous driving of vehicles, spacecraft processes, and many others.

FIG. 1 is a schematic diagram of a method that includes training a machine learning software model 106 with real data 102 and virtual data 104 about a process to be able to predict the behavior of sensors in a real version of the process. Once trained, according to FIG. 1, the machine learning software model 106 can be applied to any of a variety of possible uses, including: for example, the machine is controlled/optimized (at 108) to perform the same type of process or a different type of process, to diagnose the machine and/or sensors in the machine (at 110), and/or to simulate various aspects of the machine/sensor's behavior or process under various operating parameters (at 112). Of course, other uses are possible. Further, in some embodiments, the machine learning software model 106 may be used for one or more of these processes while continuing to be trained. In those cases, continued training will continue to develop the machine learning software model 106 to predict sensor behavior even better.

In typical embodiments, the systems and methods disclosed herein (e.g., as schematically illustrated in fig. 1) result in highly accurate sensor predictions. Furthermore, in some embodiments, these predictions can be made very quickly, which is especially important where a limited time is available for extensive training of the machine learning software model 106. In these cases, a high accuracy of prediction can be achieved despite the limited time available during the real-time process. This makes the systems and methods disclosed herein very valuable in a variety of different applications.

The real data 102 utilized to train the machine learning software model 106 may be any type of data from a real-world version of a process. There are many possible sources for the real data 102.

One possible source of the real data 102 is a real-world version of the sensor for which the machine learning software model 106 is learning to predict. Typically, such real-world sensors will be deployed in a running machine to sense one or more conditions produced by the machine while performing the process in question. For example, in an additive manufacturing setting such as 3D printing, the machine learning software model 106 may be trained to predict the behavior of a photodiode sensor in a 3D printer configured to sense light in the 3D printer (as an indicator of thermal energy transferred into a print-ready object). In those embodiments, the real data may include actual data collected by the photodiode sensor during the 3D printing process.

Another possible source of real data 102 is process parameters or settings used to implement the process on the real machine. In the 3D printing example above, the process parameters may include information or settings on the 3D printer related to one or more of: laser exposure time, point-to-point distance, laser power, laser speed, and laser strategy including profile and shading/filling patterns, radial step, offset, direction, and sequence, etc. Certain process parameters or machine settings may be relevant to training the machine learning software model 106 to predict the behavior of the photodiode sensors in the 3D printer example because they may affect the amount of heat transferred into the printed part. Other process parameters in a 3D printing environment or other type of process may also be relevant.

The real data 102 may also come from other sources.

The virtual data 104 used to train the machine learning software model 106 typically comes from simulation or virtual modeling of one or more aspects of the process or machine in question.

The simulation to produce the virtual data 104 is typically performed by a computer-based simulator and may be relatively simple or refined. Simpler simulations tend to be faster, which may be desirable in some cases, but contain less detail. Complex and refined simulations provide more detail about the interaction of physical phenomena, but tend to take more time. In general, the simulation method should be adapted to the requirements of a particular situation.

For example, in a relatively simple simulation, the simulator may calculate the virtual data 104 in the form of one or more physics-based features (e.g., features related to the underlying physics of the process in question) based on readily available process information, such as one or more of the process parameters mentioned above. The physics-based features are often not directly measurable and may not be easily measurable in an accurate manner by physical sensors. Some examples of physics-based features suitable for use in a 3D printing environment include some measure of the energy introduced to a part being printed within a particular time window, and some measure of the ability of the part to dissipate the introduced energy through thermal conduction.

More complex simulations may be performed by complex computer-based simulators, such as powder bed manufacturing and additive manufacturing scheme software applications, which are of Dassault Syst mes SE

A part provided by the software platform. The powder bed manufacturing software application is capable of generating virtual data 104 in the form of high resolution machine tool path information based on a Computer Aided Design (CAD) model and, in some cases, other input data. The high resolution machine tool path information generated in this regard may include information similar to the process parameters described above, for example, information related to one or more of: laser exposure time, point-to-point distance, laser power, laser speed, laser strategy including profile and shading/filling pattern, radial step size, offset, direction and sequence, etc. The additive manufacturing solution software application is capable of generating virtual data 104 in the form of high fidelity physics information based on Computer Aided Design (CAD) model machine tool path information and, in some cases, other input data. The high fidelity physics information generated in this regard may include virtual data of physical quantities such as temperature field, bath size, material phase change and content, mechanical deformation, residual stress and strain, and the like.

In some implementations, the virtual data 104 can include a combination of physics-based features and information from more complex computer-based simulators.

The real data 102 and virtual data 104 are typically collected in, for example, computer-based memory and used with a trained machine learning software model 106 to enable prediction of the real behavior of sensors (e.g., photodiode sensors in a 3D printer) for various computer-aided design (CAD) models under various operating conditions.

Figure 2 is a schematic diagram of a more detailed representation of the method shown in figure 1.

The method represented in fig. 2 includes training the machine learning software model 106 with both real data (in the form of machine process parameters 102a and real sensor data 102b) and virtual data 104 about the process to be able to predict (at 214) the behavior of real sensors in a real version of the process.

The method includes providing (at 216) machine process parameters 102a and/or model data to facilitate simulation (at 218) of a process to generate virtual data 104. As an example, the model data may be a Computer Aided Design (CAD) model of a product to be manufactured by a 3D printing process. However, the model data may be virtually any other collection of information related to a real-world process that may be used to assist in simulating a process. Generally, if the process to be simulated is to include a machine operating on process parameters, those process parameters 102a may also be provided into the simulation (at 218). In some embodiments, the simulation may be performed based solely on machine process parameters (at 218).

The illustrated method also includes performing a real process (at 220). Sensors are provided to sense certain characteristics of the process. In this regard, the sensors may be deployed inside a machine for performing a process or some aspect of a process. Accordingly, the execution (at 220) of the process produces more real data in the form of real sensor data 102 b.

Real data (including machine process parameters 102a and real sensor data 102b) and virtual data 104 generated by the simulation (at 218) are fed into the machine learning software model 106 to train the machine learning software model 106 to be able to predict the behavior of the sensors.

Once trained, the machine learning software model 106 predicts the true sensor values according to the illustrated embodiment. The machine learning software model 106 may make these predictions based on any combination of real data 102 (as opposed to real sensor data, as real sensor data would not be available to processes that have not yet occurred) and/or virtual data 104. In this regard, the behavior of the machine learning software model 106 is similar to a transfer function that models sensor behavior for any possible combination of model and machine or process input.

Of course, the techniques disclosed herein may be applied to a variety of possible applications. One such application is additive manufacturing, such as 3D printing. The nature of 3D printing is that there may not be enough time to perform extensive training of the machine learning software model 106 or to simulate/model the precise physics of the photodiode sensors in the 3D printing process, the 3D printer and/or the 3D process.

Fig. 3 is a schematic cross-sectional view of an exemplary 3D printer 322.

The illustrated 3D printer 322 is generally configured to generate one layer of functional engineering components (parts) at a time from Computer Aided Design (CAD) data. In a typical embodiment, the 3D printer 322 produces a quantity of material in a layered manner by melting raw material materials (e.g., powder, wire, etc.) using an energy source (e.g., laser, electron beam, arc, etc.) that follows a tool path derived from a CAD model.

The 3D printer 322 in the illustrated embodiment includes: a housing 324 defining a pair of internal cylinders; a build platform 326 movable in a vertical piston-like manner in a first one of the cylinders relative to the housing 324; a new powder storage platform 328 in the cylinder that moves in a vertical piston-like manner in the second cylinder; and a powder roller 330. In an exemplary embodiment, during operation, the new powder magazine platform 328 pushes the new powder magazine 332 upward, the build platform 326 moves in a downward direction, and the powder rollers 330 roll the new powder from the new powder magazine out of the powder bed 331.

The 3D printer in the illustrated embodiment has a heat source 334, which may be a laser, electron beam source, or the like. During operation, the heat source 334 moves through the top of the powder bed to direct thermal energy (e.g., with a laser beam) onto different portions of the powder bed to melt the upper layer of powder in the portions of the powder bed. Heat from the heat source melts the powder as it moves through the powder bed. After the laser passes a particular point on the powder bed, the molten powder cools and solidifies into a solid (non-powder) form, forming a portion of the part being printed. The 3D printer in the illustrated embodiment also has a photodiode sensor 336 near the laser delivery point on the powder bed. The photodiode sensor 336 is configured to sense the light intensity (as an indicator of the thermal energy transferred into the powder bed by the heat source 334 at each point in time). More specifically, in typical embodiments, the photodiode sensor includes the combined effects of delivered energy, material absorption, conduction, radiation, airflow, light reflection, and the like.

FIG. 4 is a schematic diagram of an embodiment of the method represented in FIG. 2 specific to a 3D printing environment. While specific to a 3D printing environment, the method represented in fig. 4 may be readily applied to a variety of applications that may benefit from rapid training and rapid prediction.

The method represented in fig. 4 includes training the machine learning software model 406 with both real data (in the form of machine process parameters 402a and photodiode sensor data 402b) and virtual data to be able to predict (at 414) the real behavior of the photodiode sensor 336. The virtual data in the illustrated embodiment contains physics-based features 404a that represent the energy introduced to the part being printed within a particular time window, as well as the ability of the system to dissipate the introduced energy via thermal conduction. There are a number of possible ways to generate (at 403) these physics-based features. In a typical implementation, the physics-based features are generated (at 403) by software executing on a computer-based processor and may be based on any one or more of a variety of inputs. These inputs may include, for example, machine processing parameters 402a of the 3D printer and/or information from the high resolution machine tool path 404 b. The machine tool path 404b in the illustrated embodiment is generated by a powder bed fabrication simulator 418 based on a CAD model 416 of the part to be fabricated.

Thus, the process represented in FIG. 4 includes generating a CAD model 416 of certain geometries (e.g., parts) to be printed by the 3D printer. Figure (a). FIG. 5 shows several examples of CAD models 501, 503, 505, 507, 509, 511, 513, and 515 having different geometries represented in the CAD models. These geometries are shown in the figure above the virtual build platform 517 of the 3D printer.

The process represented in FIG. 4 includes determining machine processing parameters for a 3D printer to build a part (at 402 a).

There are a number of ways in which the machine processing parameters 402a may be determined. In some embodiments, the technician will look at the geometry of the part and use a set of machine parameters based on experience (e.g., uniform laser power for laser shadow scans, distance between shadow scans, etc.). The technician may then enter the parameters into a real machine (i.e., a real 3D printer) or a virtual machine (i.e., a virtual simulation of a 3D printer/photodiode sensor) in the powder bed manufacturing 418 application. The real machine will generate machine controls to build parts accordingly, or the virtual machine will generate a machine tool path for virtual print simulation. The quality of the real part coming out of the machine is usually checked, and if necessary, the machine settings can be adjusted based on the check, and the part is then printed again. This trial and error process may continue until an acceptable quality level is reached. In the virtual aspect, an engineer may inspect the virtual part and adjust settings in the powder bed manufacturing application 418 as needed.

Fig. 6 and 7 show screenshots of the powder bed manufacturing application (418) that enable a user to input various machine process parameters for a particular printing process.

More specifically, the illustrated screen shot enables a user to specify global parameters for the printing process (in FIG. 6), compose a policy (in FIG. 7), and identify scan path order (not shown in a separate figure).

The global parameters page in the screenshot of FIG. 6 enables the user to specify or enter the name of the scan path rule and identify the slice step size (in millimeters) for printing. The screen shot enables the user to specify overlay (upskin) information including a minimum overlay width, number of layers, and to restrict the core through the overlays. Also, the screen shot enables the user to specify the lower skin (downtkin) information including the minimum skin width, number of layers, and can restrict the core through the lower skin.

Composing the policy page in the screenshot of fig. 7 enables the user to specify the region type, outline information, fill information, definition parameters, and policy information including XY policy information. The contour information includes depth, definition, periodicity, start, range, and motion information. The fill information includes depth, definition, periodicity, start, range, and action information. Defining parameters includes scanning rules. The strategy information includes the filling pattern, the angular rotation between slices, the direction of the hatching, the radial step size and the offset from the last contour.

The scan path order page typically includes a list of definitions created for scan path orders (e.g., outlines, shadows, etc.).

Referring again to FIG. 4, based on the information provided (based on CAD models and/or machine process parameters); the powder bed fabrication application in fig. 4 generates (at 418) the high resolution machine tool path 404 b. An example of this high resolution machine tool path 404b information generated by the powder bed manufacturing application (at 418) is shown in fig. 8 and 9. Fig. 8 and 9 are exemplary screenshots generated by the powder bed manufacturing application 418 illustrating information about the detailed tool/machine path 404b generated by the application 418 for an array of cubic parts on a build platform. The arrows and lines in these screenshots represent the direction and path of travel of the laser head.

The screenshot in FIG. 8 includes an image field 811 having a three-dimensional visual representation of the laser scan path for printing the part. The screenshot in fig. 8 also includes a first user menu 813 on the right side of the screen that provides the user with access to powder bed manufacturing guide functions including functions related to the machine, build tray, powder, layout, support, scan path, analysis and output. A second user menu 815 appears at the bottom of the screenshot and provides the user with access to powder manufacturing functions including criteria, settings, programming, analysis and output, views, AR-VR, tools and touches.

The screenshot in FIG. 9 has an image field 911 with a three-dimensional visual representation showing a portion of the laser scan path for printing a part. The screen shot in fig. 9 also includes a user menu 913 located on the left side of the screen, which user menu 913 provides the user with access to various powder bed manufacturing functions related to displaying the slice and scan path, the trajectory, and the type of scan path.

In general, the information that may be accessed from the high resolution machine tool path 414 (e.g., in a powder bed manufacturing application environment) includes information related to the machine tool path of the part represented by the CAD model 416. This information typically includes laser power information, laser speed information, and laser pattern information. In an exemplary embodiment, the high resolution tool path contains temporal and spatial information of the laser path, pattern and state. In this regard, the tool path segments may be as follows:

< time 1>, < x1>, < y1>, < z1>, < power 1>

< time 2>, < x2>, < y2>, < z2>, < power 2>

< time 3>, < x3>, < y3>, < z3>, < power 2>

The tool path segment information indicates that the laser should travel from point 1 at the time 1, x1, y1, z1 coordinates to point 2 at the time 2, x2, y2, z2 coordinates at a constant speed equal to the distance between point 1 and point 2 divided by the difference between time 1 and time 2 and at a constant power equal to power 1. The laser then travels from point 2 at the time 2, x2, y2, z2 coordinates to point 3 at the time 3, x3, y3, z3 coordinates at a constant speed equal to the distance between point 2 and point 3 divided by the difference between time 2 and time 3, a constant power equal to power 2, and so on.

The illustrated method also includes performing one or more real world prints of the part (at 420), which results in real world photodiode sensor data 402 b. Thus, the photodiode sensor data 402b is generated by one or more real machines during one or more real printing processes. In some embodiments, to begin the machine learning process, some parts are programmed with various machine process parameters 402a and first printed to produce an initial set of photodiode sensor data 402b for training purposes. The data format will depend on the particular machine and sensor, but in typical embodiments, the data format will include at least temporal, spatial location, and sensor amplitude information. The amount of data is typically dependent on part size, machine parameters, and sensor output frequency.

In the illustrated embodiment, the physics-based features 403 are calculated based on information from the high-resolution machine tool path 404 b.

As described above, the high resolution machine tool path 404b specifies laser positions at different times, laser powers at these different times, and laser speeds between these times. In addition, in a typical embodiment, a region of interest is specified around the laser position (either specified by the user or preset as a process) and stored in computer-based memory. The size and shape of the region of interest may vary from application to application. However, in a typical embodiment, the region of interest will be a circular area on the upper surface of the printed powder bed/part, having the same size (+/-10%) as the printed portion of the powder bed/part that the photodiode sensor is configured to sense. Thus, if the photodiode sensor is configured to sense a circular area having a diameter of 4 millimeters, the area of interest may also be a circle having a diameter of 4 millimeters (+/-10%). The region of interest moves with and generally concentrically surrounds the laser position on the surface of the powder bed or part being printed at each point during the printing process.

For example, based on the above information, a computer-based processor can determine how much energy is delivered by the laser to the upper surface of the powder bed and part being printed during printing, and precisely where to deliver the energy anywhere and for any period of time.

In typical embodiments, a computer-based processor calculates an energy flux signature that represents how much energy has been delivered by the laser into a region of interest surrounding the laser location during a particular time period. The time period may include any part or all of the printing process and may have been defined by the user or set by default. A computer-based processor may calculate the energy flux characteristics by integrating the laser power/energy data delivered to the region of interest over a particular time period. The computer-based processor may update its energy flux signature calculations periodically or nearly continuously and/or in near real-time throughout the printing process.

In typical embodiments, the computer-based processor also calculates an effective conductance signature that provides a measure of the ability of the powder bed and/or part to dissipate the energy delivered by the laser at a particular point in time. In an exemplary embodiment, the computer-based processor calculates the effective conduction signature by calculating the area (e.g., in square millimeters) within the region of interest that is not scanned by the laser for a particular period of time. The computer-based processor may update its effective conduction signature calculation to coincide with the update in its energy flux signature calculation. Thus, they may be updated periodically throughout the printing process, or almost continuously and/or almost in real time.

The concept of energy flux and efficient conduction in the environment in physics-based feature computation can be further understood, for example, by referring to fig. 10-12.

Fig. 10 is a top view of a powder bed 1351 showing a part being printed in a 3D printer (e.g., 3D printer 322 in fig. 3), and a schematic of the path that a laser tool will travel along the upper surface of the powder bed in a rectangular print layer.

The outer rectangle 1350 shown as a dotted line represents the boundary of the part at the current level. The two solid rectangles 1352 in the dotted line represent the laser paths of the two contours. The solid and dashed lines within the boundaries of the outline represent the path of the filling laser tool. Specifically, a solid line 1354 within the contour line boundary represents the area within the region that the laser has scanned by the laser up to a particular point in time, and a dashed line 1356 within the contour line boundary represents the area within the region that the laser has not scanned by the laser up to a particular point in time. A point 1358 near the center of fig. 10 represents the laser position at a particular point in time during the scanning process of the laser. A circle 1360 surrounding the black dot represents a region of interest for computing the physics-based features 404a (e.g., energy flux features and effective conductance features). The region of interest is essentially the area within the region of the powder bed/part being printed, and the area within this region is believed to be relevant for the purpose of calculating the physics-based features described above.

In an exemplary embodiment, the boundary (outer dotted line 1350) at which the part is at the current layer may be based on the size of the part being printed and/or the size of the powder bed 328. The two contour laser paths (solid rectangular lines 1352) and the path of the filler laser tool (solid lines 1354 and dashed lines 1356 within the contour boundaries) are a function of the high resolution machine tool path 404 b. In certain embodiments, the region of interest 1360 is the same size and shape as the region on the upper surface of the powder bed 328 that the photodiode sensor 336 is configured to sense (+/-10%). In one exemplary embodiment, the photodiode sensor 336 is configured for sensing the area on the upper surface of the powder bed 328/part being printed within a circular area of approximately 4 millimeters in diameter. In those embodiments, the diameter of the circle identifying the region of interest may be set to about 4 millimeters (+/-10%). In some embodiments, the size of the circle may vary depending on one or more factors (e.g., the speed of laser 334 through powder bed 328, etc.). In some implementations, the size of the circle may also be specified by the user (e.g., at a computer-based interface that enables the user to interact with the physics-based feature generation process). In a typical embodiment, as the laser moves across the powder bed surface, the area of interest circle will also move across the powder bed surface with it, such that the laser position point 1358 moving across the powder bed surface is always centered within the area of interest circle 1360.

The thicker solid line 1362 within the region of interest circle 1360 represents the area within those regions within the region of interest circle 1360 that have been scanned by the laser over a particular period of time. Also, these thicker solid lines 1362 identify areas within the region of interest circle 1360 that are believed to be relevant for the purpose of calculating the energy flux characteristics corresponding to the laser position indicated by point 1358.

The narrower solid lines 1364 within the region of interest circle 1360 identify areas within the region that have been scanned by the laser, but these areas are not relevant for the purpose of calculating the energy flux characteristics of the indicated laser location 1358. One reason why the area in the region identified by these narrower solid lines is not considered relevant in this respect may be: those scans occur outside of a certain period of time that is considered relevant for energy flux calculations. In essence, if a sufficiently long time has elapsed after laser scanning a particular region, the energy dissipation of that region will be sufficiently large that the area within that region can be ignored in calculating the energy flux characteristics.

Dashed line 1366 within region of interest circle 1360 identifies the area within the region that has not been scanned by the laser. And these areas (identified by dashed line 1366) are considered to be irrelevant for the purpose of calculating the energy flux characteristics of the indicated laser location 1358. They are considered independent of the calculation of the energy flux characteristics because the laser energy has not yet been delivered to these regions. However, the area within these regions is considered relevant for calculating the effective conduction characteristics. And further, the area within these regions is considered to be relevant for calculating effective conduction characteristics because these regions have not received direct transfer of laser energy. As such, these regions 1366 are at a much lower temperature than the recently scanned region 1362, and therefore, these regions 1366 provide escape paths for heat/energy from the region 1362 through thermal conduction. In some embodiments, the region 1364 may also be included in the calculation of effective conduction because the region 1364 may also provide an escape path for heat/energy from the region 1362 through thermal conduction if it is sufficiently cooled. Generally, if the area 1366 (and area 1364) is larger, then a greater heat dissipation capability will result; generally, if the area 1366 (and the area 1364) is small, then a weak heat dissipation capability will be generated.

The computer-based processor may calculate the effective conductance features in a number of ways. In one such example, the computer-based processor would subtract the area of region 1362 from the area of the entire region of interest 1360. In another such example, a computer-based processor would calculate the area of region 1366. In yet another such example, the computer-based processor adds the area of region 1366 and the area of region 1364. In some embodiments, these calculations may be made directly from information provided by the high resolution machine tool path 404 b.

Fig. 11 and 12 are schematic views of the powder bed 1351 of fig. 10 showing the laser position 1358 and the area of interest circle 1360 near the very beginning of the scanning process (in fig. 11) and near the end of the scanning process (in fig. 12).

These figures help illustrate the following points. First, the area of interest circle 1360 moves with the laser position 1358 throughout the printing process, and maintains the circle 1360 in a centered position. Second, the calculations discussed above for the physics-based features are applicable and accurate even if the laser is near the geometric boundary of the part being created. The thick lines in fig. 11 represent integration paths for calculating an unscanned region, which is a relatively large portion of the region of interest (i.e., a feature size local patch). The thick lines in fig. 12 represent integration paths for calculating an unscanned area, which is a relatively small portion of a circle. Thus, physics-based features (i.e., energy flux features and effective conduction features) will automatically incorporate geometric boundary effects, eliminating the need to design geometric features into the machine learning process, which can be extremely complex and difficult to meet.

Other various techniques or methods for calculating physics-based features are also possible. Generally, however, these physics-based features 404a can be quickly and easily calculated even without extensive simulation or modeling. In addition, these features have a significant impact on enabling the machine learning software model 406 to accurately predict the behavior of the photodiode sensor 406 b. In addition, it has been found that the physics-based features 404a are included in the training process of the machine learning software model 406 to enable the machine learning software model 406 to better predict the behavior of the photodiode sensor in other processes/machines. In addition, by accessing detailed tool path information (e.g., according to DassaultSystemes SE)Virtual machine data provided by a software platform), these methods can be easily implemented and employed, and because they are relatively simple and efficient, they can generally be extended to part-level prediction.

Referring again to fig. 4, physics-based features 404a (i.e., one or more energy flux features and one or more effective conduction features) are provided to machine learning software model 406 along with machine process parameters 402a and real world photodiode sensor data 402b to train machine learning software model 406 to be able to predict values 414 of the photodiode sensors.

In certain embodiments, particularly those embodiments where training speed and/or predicted speed are of particular interest, the aforementioned physics-based features (i.e., energy flux features and effective conductance features) may be the only form of virtual data used to train machine learning software model 406. In those cases, all of the remaining training data provided to machine learning software model 406 may be real data (e.g., machine process parameters 402a and/or real photodiode sensor data 402b of the 3D printer). In some embodiments, it may be desirable to supplement training with other dummy data, particularly without severely limiting the time available.

It has been found that the phenomena represented by these physics-based specific features (i.e., energy flux features and effective conduction features) can be generated very quickly and are very effective in training the machine learning software model 406 to accurately predict the behavior of the photodiode sensor 336 under various operating conditions. Moreover, high accuracy can be achieved even if these features are the only virtual data types used in the training process, while all other training data are real data.

Once trained, the machine learning software model 406 is able to predict (at 414) the behavior of real photodiode sensors, according to the illustrated embodiment.

It has been found that using machine-learning software models trained with physics-based features (in addition to the real data) can produce more accurate sensor predictions than machine-learning software models trained using only real data. In this regard, fig. 13 shows a single layer test data prediction scatter plot based only on machine process parameters and photodiode sensor data, and fig. 14 shows a single layer test data prediction scatter plot based on machine process parameters, photodiode sensor data, physics-based features, and a CAD model. These figures show a significant improvement in accuracy when physics-based features and CAD models are included.

The trained machine learning software model can then be used to provide a quick prediction of physical sensor data for any given part to guide the machine operator to better machine settings; or machine parameters for optimizing target sensor signals for the entire part, thereby reducing trial and error and achieving a suitable printing effect. For example, based on the initial settings of machine process parameters for a given CAD geometry, the above-described method can be used to generate physics-based features and combined with real data (e.g., initial machine process parameters) to train a machine learning software model to predict sensor readings for the entire part. The optimal machine process parameters including the optimal 3d laser power or laser velocity field of the entire part can then be iteratively derived using, for example, sensitivity-based optimization methods.

Fig. 15 shows experimental verification of photodiode sensor readings for a single layer before and after a print optimization process that relies on this approach. The old photodiode sensor reading with the initial machine process parameter and the new photodiode sensor reading with the best machine process parameter are both marked. The results show that by print optimization, the photodiode sensor can be optimized very close to the target level, thereby preventing overheating or underheating of parts during printing, and significantly reducing the trial and error required to successfully build.

FIG. 16 is a diagram illustrating an exemplary flow of a machine learning and print optimization framework. In the illustrated embodiment, a training data set is first defined (at 1970). Physical experiments were performed to record machine settings and sensor data. With the proposed method of these parts, relevant physics-based features are created (at 1972). The data is then used to train a predictive machine learning software model. Given a new part CAD model (at 1974) and machine operational constraints, the machine settings are then optimized for the target sensor output (at 1976). A machine-readable file format is generated and sent to the hardware (at 1978) for printing at the target part quality for the first time. The machine learning software model may continue to be trained as more parts are printed and data is added as other training data, and the accuracy of the machine learning model will increase as more parts are printed.

Once the machine learning software model is trained, the trained machine learning software model, physics-based feature generation, and optimization methods can be used in conjunction with one or more additive manufacturing machines to make sensor predictions and optimize machine parameters in real-time on a layer-by-layer basis. An exemplary implementation of such a process is shown in fig. 17.

The process represented by fig. 17 includes importing a CAD model into the machine and setting initial machine parameters (at 2080) and optimizing for first-tier deployment (at 2082). Typically, optimization of the first print layer is accomplished according to machine learning predictions using machine parameters and generated physics-based features. The machine then prints (at 2084) the first layer using the optimized parameters. While the first layer is being built, the system deploys the optimizations for the second layer (at 2086). The machine then prints (at 2088) a second layer using the optimized parameters. While the system is optimized for the third tier deployment (at 2090) while building the second tier. This process continues to the last layer and until the building of the last layer is complete (at 2092).

FIG. 18 is a schematic view similar in many respects to that shown in FIG. 16. In fig. 18, sensor and operating condition data are first defined (at 2170). Physical experiments were performed to record machine settings and sensor data. Simulation data is created (at 2172). These data are then used to train a predictive machine learning software model. The in-service sensor prediction is then made using the machine learning software model (at 2174). Next, the system determines (at 2176) whether a sensor anomaly exists (i.e., whether the actual sensor reading differs from the predicted sensor reading by a sufficient amount). In a typical implementation, a system (e.g., a computer-based processor in the system) may receive an actual sensor reading and compare the actual sensor reading to a corresponding predicted sensor reading, and then issue an alert (at 2178) if the difference between the two exceeds a certain threshold (e.g., 1%, 5%, 10%, etc.). In some embodiments, the threshold may be defined by a user. In some embodiments, the threshold may be a preset threshold. The warning may take any of a variety of different forms, including, for example, a message to the operator with an audible and visual element or tactile element. In some embodiments, the warning may be accompanied by a pause in the printing process. And other warnings and associated operations are possible.

FIG. 19 illustrates a computer network or similar digital processing environment in an embodiment in which the invention may be implemented.

The client computer/devices 50 and server computer 60 provide processing, storage, and input/output devices that execute application programs and the like. Client computers/devices 50 may also be connected to other computing devices, including other client devices/processes 50 and server computers 60, via communications network 70. The communication network 70 may be part of: a remote access network, a global network (e.g., the internet), a cloud computing server or service, a worldwide collection of computers, a local or wide area network, and gateways that currently use various protocols (TCP/IP, bluetooth, etc.) to communicate with one another. Other electronic device/computer network architectures are also suitable.

Fig. 20 is an internal block diagram of a computer (e.g., client processor/device 50 or server computer 60) in the computer system of fig. 3. Each computer 50, 60 includes a system bus 79, where a bus is a set of hardware lines used for data transfer between components of a computer or processing system. Bus 79 is essentially a shared channel that connects different elements of a computer system (e.g., processor, disk storage, memory, input/output ports, network ports, etc.) that enables the transfer of information between the elements. Attached to system bus 79 is I/O device interface 82 for connecting various input and output devices (e.g., keyboard, mouse, displays, printers, speakers, etc.) to the computer 50, 60. Network interface 86 enables the computer to connect to various other devices attached to a network (e.g., network 70 of fig. 3). Memory 90 provides volatile storage for computer software instructions 92 and data 94 (e.g., the code detailed above in fig. 1 and 2) used to implement embodiments of the present invention. The memory may include any combination of one or more computer-based memory storage devices, which may be local or distributed. Disk storage 95 provides non-volatile storage for computer software instructions 92 and data 94 used to implement embodiments of the present invention. A Central Processor Unit (CPU) 84 is also attached to system bus 79 and is used to execute computer instructions. In typical embodiments, the techniques disclosed herein as being performed by a computer-based processor or the like may also be performed by: CPU 84, or some combination of processors, which may be local or distributed.

In one embodiment, the processor routines 92 and data 94 are a computer program product (generally referred to as 92) including a computer readable medium (such as one or more removable storage media like a Digital Video Disk-Read Only Memory (DVD-ROM), a compact Disk Read-Only Memory (CD-ROM), a diskette, a tape, etc.) providing at least a portion of the software instructions for the system. The computer program product 92 may be installed by any suitable software installation process, as is well known in the art. In another embodiment, at least a portion of the software instructions may also be downloaded via a cable, communication, and/or wireless connection. In other embodiments, the inventive program is a computer program propagated signal product 107, embodied as a propagated signal on a propagation medium (e.g., a radio wave, an infrared ray, a laser wave, a sound wave, or an electric wave propagated over a global network such as the Internet or other networks). Such carrier medium or signals provide at least portions of the software instructions for the present invention routines/programs 92. In alternative embodiments, the propagated signal is an artificial carrier wave or a digital signal carried on the propagation medium. For example, the propagated signal may be a digitized signal propagated over a global network (e.g., the Internet), a telecommunications network, or other network. In one embodiment, the propagated signal is a signal that is transmitted over a propagation medium over a period of time, such as instructions for a software application sent in packets over a network in milliseconds, seconds, minutes, or longer. In another embodiment, the computer readable medium of the computer program product 92 is a propagation medium that the computer system 50 can receive and read, such as by receiving the propagation medium and identifying a propagated signal contained in the propagation medium for the computer program propagated signal product as described above. In general, the term "carrier medium" or transient carrier encompasses the foregoing transient signals, propagated medium, storage medium, and the like. In other embodiments, the program product 92 may be implemented as a so-called Software as a Service (SaaS) or other device that supports end-user installation or communication.

Virtual simulation data may provide more physical information about the system than using experimental data alone. For example, the simulation may inform not only the temperature of the sensor mounting location, but also the temperature everywhere in the system; the sensors are limited by the operation of the device, for example, may be limited by the measurement range or by the operating parts, yet the simulation data is always accessible; simulations provide additional physical quantities such as thermal radiation, stress, strain, etc., which cannot be measured experimentally in practice, but are critical to predicting system results. The invention provides a machine learning service based on Virtual simulation and Real (Virtual + Real, V + R) data. Simulation data includes high fidelity physics simulation models (including but not limited to finite element modeling) and data, as well as reduced or first order models and data developed from feature importance and intelligence obtained or verified from machine learning processes. The real data includes: design parameters, geometry, materials, manufacturing settings, etc. that are known without knowledge of modeling and simulation.

Various embodiments of the present invention have been described. Nevertheless, it will be understood that various modifications to the described embodiments may be made without departing from the spirit and scope of the invention.

For example, the disclosed systems and methods may be applied to a wide range of processes including all types of manufacturing processes, including additive manufacturing processes and other non-manufacturing processes. Once trained, the machine learning software model may be applied for a variety of different purposes in the process of interest, such as forecasting, diagnostics, simulation, and the like.

The particular combination of real and virtual data used to train machine learning software may vary greatly. In most cases, the training method can be tailored based on the actual considerations involved in a particular procedure. For example, if under time-critical conditions, the physics-based features disclosed herein may be the only (or nearly only) virtual data used in the training process. However, other forms of virtual data may be used when there is more time available. Also, various types of real data can be used.

CAD models may be generated using almost any type of computer-aided drawing application, and may or may not be very detailed. Likewise, the simulation may be generated using any one or combination of a number of different simulation applications. A machine process may be a process that includes one or more machines that may operate in series, parallel, or a combination thereof. The sensor to be predicted may be any type of sensor, e.g., for optics, light, imaging, photons, acoustics, sound, vibration, automotive, chemical, electrical current, electrical potential, magnetic, radio, environmental, weather, humidity, flow, fluid velocity, ionizing radiation, sub-atomic particles, navigation, position, angle, displacement, distance, velocity, acceleration, pressure, force, density, level, heat, temperature, proximity, presence, velocity, etc.

The machine learning software model may be any type of machine learning software model that is trainable and executable in a computer-based environment.

The optimization process disclosed herein may include any type of process that is performed iteratively by comparing various solutions until an optimal or satisfactory solution is found. The trained machine learning software model facilitates optimization because it feeds back instantaneous sensor predictions to the optimizer at each iteration, thereby reducing the total time required to achieve a particular degree of optimization goal.

The 3D printer disclosed herein is just one example of a 3D printing machine. Many variations are possible. The sensors discussed herein in particular relation to 3D printers are photodiode sensors. However, other sensors are possible.

The present application discloses powder bed manufacturing applications as an example of applications that can generate high resolution machine tool paths. However, other tools may also be capable of generating machine tool paths.

Physics-based features of significant interest to this application include energy flux features and efficient conduction features. In various embodiments, other physics-based features may also be calculated and used to train the machine learning software model.

Furthermore, while this specification contains many specifics of particular embodiments, these should not be construed as limitations on the scope of any invention or of what may be claimed, but rather as descriptions of features specific to particular embodiments of particular inventions. Certain features that are described in this specification in the context of a single embodiment can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment of the specification can also be implemented in multiple embodiments separately or in any suitable subcombination thereof. Moreover, although descriptions of features above may be described as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

Similarly, while operations disclosed herein occur in a particular order, it should not be understood that such operations need to be performed in the particular order shown or in that order, or that all illustrated operations need to be performed, to achieve desirable results.

Other embodiments are within the scope of the following claims.