阅读说明：本技术 一种铸件产品内部缺陷智能检测装置 (Casting product internal defect intelligent detection device ) 是由 刘骁佳 吴远峰 周鹏飞 刘世钰 洪海波 赵*** 王桃章 于 2019-10-11 设计创作，主要内容包括：一种铸件产品内部缺陷智能检测装置,X射线管发出的X射线透照铸件产品在成像板上得到包含铸件产品内部质量信息的数字图像,通过光路仿真与控制单元实现最佳成像,生成图像自动上传至云平台服务器后,通过采用数字图像处理技术、深度学习神经网络算法对图像进行智能预处理、判读缺陷存在与否、缺陷定位、缺陷类型识别、缺陷评级,实现图像表征质量的检测。通过该装置,在检测过程中实现了对复杂结构铸件检测过程的精确控制,从而最佳图像；在图像评价过程通过缺陷智能识别,代替人工评片过程,有效地缩减人工检测时长,在保证缺陷识别准确率的前提下,避免人为误差,提高铸件产品质量检测工作效率。(An X-ray transmitted casting product sent by an X-ray tube obtains a digital image containing casting product internal quality information on an imaging plate, optimal imaging is achieved through a light path simulation and control unit, the generated image is automatically uploaded to a cloud platform server, intelligent preprocessing, defect existence judging, defect positioning, defect type identification and defect grading are carried out on the image through the adoption of a digital image processing technology and a deep learning neural network algorithm, and detection of image representation quality is achieved. By the device, the precise control of the detection process of the casting with the complex structure is realized in the detection process, so that an optimal image is obtained; through defect intelligent identification in the image evaluation process, replace artifical and appraise the piece process, it is long effectively to reduce artifical measuring, under the prerequisite of guaranteeing defect identification rate of accuracy, avoids human error, improves foundry goods product quality testing work efficiency.)

1. The utility model provides a foundry goods product internal defect intellectual detection system device which characterized in that includes: the system comprises an X-ray tube, an imaging plate, a light path simulation and control unit and a cloud platform server;

the optical path simulation and control unit can select the optimal transillumination optical path according to the appearance or the three-dimensional model of the casting product, and can adjust the positions of the X-ray tube and the imaging plate according to the optimal transillumination optical path so that the casting product is positioned at the optimal position between the X-ray tube and the imaging plate; the light path simulation and control unit controls the X-ray tube to emit X-rays, and the X-rays are imaged on the imaging plate after transmitting the casting product to obtain an X-ray image, namely an original gray image of the casting product; pushing the original gray level image of the casting product to a cloud platform server;

the cloud platform server is used for preprocessing the original gray level image of the casting product to ensure that the contrast of the original gray level image meets the requirement, and the preprocessed image is obtained; inputting the preprocessed image into a convolutional neural network model CNN, carrying out defect screening, and judging whether the preprocessed image contains defects;

inputting the image judged to contain the defect into a convolutional neural network model RCNN with an interested region, positioning the defect in the defective image, and labeling a minimum block diagram of the position of the defect to realize defect positioning;

and extracting the image in the minimum block diagram after the defect positioning, inputting a deep learning neural network model DNN, and classifying the defect to obtain the type of the defect.

2. The intelligent detection device for internal defects of casting products according to claim 1, characterized in that: the imaging plate is to be located at the focus of the X-ray tube radiation.

3. The intelligent detection device for internal defects of casting products according to claim 1, characterized in that: the method comprises the following steps of preprocessing an original gray level image of a casting product, and specifically comprises the following steps:

for the original gray-scale image with severe thickness gradient change, the preprocessing method includes but is not limited to: image enhancement, image nonlinear change, filtering and noise reduction and image sharpening;

for the rough-surface original gray-scale image, the preprocessing method includes but is not limited to: enhancing images, filtering and denoising, and smoothing images;

for the original gray-scale image of the edge of the casting, the preprocessing modes include but are not limited to: image enhancement, image nonlinear change, filtering and denoising, and wavelet transformation.

4. The intelligent detection device for internal defects of casting products according to claim 1, characterized in that: the contrast requirement of the original gray image is specifically as follows: the gray difference value of four adjacent pixel points of one pixel point meets the requirement of the minimum distinguishing limit of the detection threshold.

5. The intelligent detection device for internal defects of casting products according to claim 1, characterized in that: the defects comprise air holes, slag inclusion, shrinkage porosity, cracks and segregation.

6. The intelligent detection device for internal defects of casting products according to claim 1, characterized in that: the convolutional neural network model CNN is trained in advance, and the specific training mode is as follows:

(1) establishing a training image set with a label, wherein the training image is an image obtained by preprocessing an original gray image of a casting product, and the image is marked in advance to form the label if the image is flawless; labels are classified as defective and non-defective labels;

(2) and inputting the images in the training image set into an untrained convolutional neural network model CNN one by one, and training the convolutional neural network model CNN according to a preset structure, an objective function and a model parameter optimization method of the convolutional neural network model CNN to obtain the trained convolutional neural network model CNN, namely, the parameters in the convolutional neural network model CNN are optimized.

7. The intelligent detection device for internal defects of casting products according to claim 1, characterized in that: the trained convolutional neural network model CNN can output labels, whether the casting products have defects or not is judged according to the labels, and defect screening is achieved.

8. The intelligent detection device for internal defects of casting products according to claim 1, characterized in that: the minimum block diagram of the position of the defect refers to: a group of parallel lines tangent to the longest diameter of the defect and a group of parallel lines vertical to the parallel lines form a rectangular frame.

9. The intelligent detection device for internal defects of casting products according to claim 1 is characterized in that: the convolutional neural network model RCNN with the region of interest needs to be trained in advance in the following way:

(1) establishing a training image set with a defect minimum block diagram;

(2) the images in the training image set with the minimum defect block diagram are input into an untrained convolutional neural network model RCNN with a region of interest one by one, the convolutional neural network model RCNN with the region of interest is trained according to a preset structure, an objective function and a model parameter optimization method of the convolutional neural network model RCNN with the region of interest, and the trained convolutional neural network model RCNN with the region of interest is obtained, namely parameters in the trained convolutional neural network model RCNN with the region of interest are optimized.

10. The intelligent detection device for internal defects of casting products according to claim 1, characterized in that: the trained convolutional neural network model RCNN with the region of interest can output a minimum block diagram of the position of the defect, and the defect location is realized according to the minimum block diagram of the position of the defect.

Technical Field

The invention relates to an intelligent detection device for internal defects of a casting product, and belongs to the technical field of intelligent detection.

Background

At present, in the industrial field, the detection of internal defects of casting products is an important part of quality detection in the production process, the ray imaging technology is generally adopted to photograph internal information of the casting products, and then workers with abundant experience interpret the internal defects of the casting products according to the photographs. However, the defect evaluation of the digital radiographic image needs qualified skilled workers to complete, a great deal of time and money are consumed for training the skilled workers, and a great deal of evaluation work easily causes fatigue of the skilled workers, so that human errors such as misjudgment and missed judgment are easily caused. Meanwhile, the increase of the casting yield accompanying the development of the productivity puts higher demands on the evaluation efficiency of the sheet. With the development of artificial intelligence technology, deep learning neural network algorithms based on digital images take more and more important tasks in the field of nondestructive testing.

Disclosure of Invention

The technical problem solved by the invention is as follows: the intelligent detecting device for the internal defects of the casting products is applied to quality detection tasks of the casting products, analyzes digital radiation images by using a computer deep learning method according to the internal defect characteristics of the casting products, intelligently judges and reads results, effectively shortens the time for manual detection, avoids human errors on the premise of ensuring the defect identification accuracy, and improves the quality detection work efficiency of the casting products.

The technical scheme of the invention is as follows: an intelligent detection device for internal defects of a casting product comprises: the system comprises an X-ray tube, an imaging plate, a light path simulation and control unit and a cloud platform server;

the optical path simulation and control unit can select the optimal transillumination optical path according to the appearance or the three-dimensional model of the casting product, and can adjust the positions of the X-ray tube and the imaging plate according to the optimal transillumination optical path so that the casting product is positioned at the optimal position between the X-ray tube and the imaging plate; the light path simulation and control unit controls the X-ray tube to emit X-rays, and the X-rays are imaged on the imaging plate after transmitting the casting product to obtain an X-ray image, namely an original gray image of the casting product; pushing the original gray level image of the casting product to a cloud platform server;

the cloud platform server is used for preprocessing the original gray level image of the casting product to ensure that the contrast of the original gray level image meets the requirement, and the preprocessed image is obtained; inputting the preprocessed image into a convolutional neural network model CNN, carrying out defect screening, and judging whether the preprocessed image contains defects.

And inputting the image judged to contain the defect into a convolutional neural network model RCNN with an interested region, positioning the defect in the defective image, and labeling the minimum block diagram of the position of the defect to realize defect positioning.

And extracting the image in the minimum block diagram after the defect positioning, inputting a deep learning neural network model DNN, and classifying the defect to obtain the type of the defect.

Preferably, the imaging plate is located at the focus of the X-ray tube radiation.

Preferably, the original gray level image of the casting product is preprocessed, specifically as follows:

for the original gray-scale image with severe thickness gradient change, the preprocessing method includes but is not limited to: image enhancement, image nonlinear change, filtering and noise reduction and image sharpening.

For the rough-surface original gray-scale image, the preprocessing method includes but is not limited to: image enhancement, filtering and noise reduction and image smoothing.

For the original gray-scale image of the edge of the casting, the preprocessing modes include but are not limited to: image enhancement, image nonlinear change, filtering and denoising, and wavelet transformation.

Preferably, the contrast requirement of the original grayscale image is specifically: the gray difference value of four adjacent pixel points of one pixel point meets the requirement of the minimum distinguishing limit of the detection threshold.

Preferably, the defects include pores, slag inclusions, shrinkage porosity, cracks and segregation.

Preferably, the convolutional neural network model CNN is trained in advance, and the specific training mode is as follows:

(1) establishing a training image set with a label, wherein the training image is an image obtained by preprocessing an original gray image of a casting product, and the image is marked in advance to form the label if the image is flawless; labels are classified as defective and non-defective labels;

(2) the images in the training image set are input to the untrained convolutional neural network model CNN one by one, the convolutional neural network model CNN is trained according to the preset structure, objective function and model parameter optimization method of the convolutional neural network model CNN, and as shown in fig. 3, the trained convolutional neural network model CNN is obtained, namely parameters in the convolutional neural network model CNN are optimized.

Preferably, the trained convolutional neural network model CNN can output labels, and whether the casting product has defects is judged according to the labels, so that defect screening is realized.

Preferably, the minimum block diagram of the position of the defect is as follows: a group of parallel lines tangent to the longest diameter of the defect and a group of parallel lines vertical to the longest diameter of the defect (the position tangent to the outermost side of the edge of the defect) form a rectangular frame.

Preferably, the convolutional neural network model with the region of interest RCNN needs to be trained in advance in the following way:

(1) establishing a training image set with a defect minimum block diagram;

(2) the images in the training image set with the minimum defect block diagram are input into an untrained convolutional neural network model RCNN with a region of interest one by one, the convolutional neural network model RCNN with the region of interest is trained according to a preset structure, an objective function and a model parameter optimization method of the convolutional neural network model RCNN with the region of interest, and as shown in FIG. 4, the trained convolutional neural network model RCNN with the region of interest is obtained, namely parameters in the trained convolutional neural network model RCNN with the region of interest are optimized.

Preferably, the trained convolutional neural network model RCNN with the region of interest can output a minimum block diagram of the position of the defect, and the defect location is realized according to the minimum block diagram of the position of the defect.

Preferably, the deep learning neural network model DNN needs to be trained in advance, and the training mode is as follows:

(1) establishing a training image set with a defect type label and a defect minimum block diagram; each defect type label of the training image with the minimum defect block diagram corresponds to one defect type;

(2) and (3) inputting the images in the training image set with the defect type label and the defect minimum block diagram into an untrained deep learning neural network model DNN one by one, and training the deep learning neural network model DNN according to a preset structure, an objective function and a model parameter optimization method of the deep learning neural network model DNN to obtain the trained deep learning neural network model DNN, namely, parameters in the deep learning neural network model DNN are optimized.

Preferably, the deep learning neural network model DNN can output the label with the defect type, and the defect classification is realized according to the label with the defect type.

Preferably, after the defects are classified, the classified defects may be further subjected to defect rating, whether the defect rating meets different-level patterns of the same type of defects input by previous DNN network training is judged, and a detection result is output.

Compared with the prior art, the invention has the advantages that:

(1) the invention realizes the confirmation of the optimal detection light path by adopting the light path simulation and control module, replaces the light path arrangement made by detection personnel by experience, obtains the optimal image and improves the image precision and sensitivity.

(2) According to the invention, the cloud platform server is adopted to carry out intelligent analysis on the digital radiographic image, and the cloud service is adopted, so that gradual quality information and internal defects of different products and different times in the detection process can be effectively collected and summarized, data support is improved for optimizing the design and production process, and the quality of the product is finally improved

(3) The invention analyzes the digital radiographic image by a computer deep learning method, intelligently interprets the result, partially or completely replaces the manual film evaluation process, effectively shortens the manual detection time, avoids human errors on the premise of ensuring the defect identification accuracy, and improves the quality detection work efficiency of casting products.

(4) According to the method, the defect screening of the digital radiographic image is carried out through the convolutional neural network model CNN, the method can effectively identify macroscopic information and microscopic information in the image, and can carry out model upgrading optimization along with data volume accumulation, so that support is provided for the iteration of a defect screening algorithm.

(5) According to the method, the defect of the digital radiographic image is positioned through the convolutional neural network model RCNN with the region of interest, the block diagram with the minimum defect can be effectively identified, and support is provided for defect rating of subsequent defect size measurement.

Drawings

FIG. 1 is a schematic block diagram of the intelligent defect detection method of digital radiographic images of a casting product according to the present invention;

FIG. 2 is a schematic view of an intelligent detection device for internal defects of a casting product according to the present invention;

FIG. 3 is a schematic diagram of a defect screening algorithm for digital radiographic images of cast products according to the present invention;

FIG. 4 is a schematic diagram of the defect localization algorithm of the digital radiographic image of the casting product of the present invention.

Detailed Description

The invention is described in further detail below with reference to the figures and specific embodiments.

The invention relates to an intelligent detection device for internal defects of casting products, wherein X-rays emitted by an X-ray tube transilluminate the casting products to obtain digital images containing internal quality information of the casting products on an imaging plate, optimal imaging is realized through a light path simulation and control unit, generated images are automatically uploaded to a cloud platform server, and then intelligent preprocessing, defect existence judgment, defect positioning, defect type identification and defect grading are carried out on the images through the adoption of a digital image processing technology and a deep learning neural network algorithm, so that the detection of image representation quality is realized. By the device, the precise control of the detection process of the casting with the complex structure is realized in the detection process, so that an optimal image is obtained; through defect intelligent identification in the image evaluation process, replace artifical and appraise the piece process, it is long effectively to reduce artifical measuring, under the prerequisite of guaranteeing defect identification rate of accuracy, avoids human error, improves foundry goods product quality testing work efficiency. Real-time uploading and processing of detection data are realized through the cloud platform server, and conditions are created for production whole-process digitization.

The invention is mainly applied to the detection of complex structure castings, such as aerospace complex space topological structure castings, complex revolving body structure castings, complex casings and engine castings, effectively improves the detection quality and detection consistency, and improves the accuracy and efficiency of detection image evaluation, thereby finally improving the quality of the castings, and meanwhile, the detection data realizes the storage and pushing of a cloud platform, and provides a foundation for the digital and intelligent application of the production flow.

Fig. 1 is a schematic block diagram of an intelligent defect detection method for a casting product digital radiographic image, in fig. 1, 1 is an original image, 2 is image preprocessing, 3 is defect screening, 4 is defect positioning, 5 is defect type identification, 6 is defect 1, 7 is defect 2, 8 is defect 3, 9 is qualified, 10 is rework maintenance, 11 is yield acceptance, 12 is scrapped, fig. 2 is a schematic diagram of an intelligent internal defect detection device for a casting product, in fig. 2, ① is an optical path simulation and control unit, ② is an X-ray tube, ③ is an imaging plate, ④ is a cloud platform, fig. 3 is a schematic diagram of a defect screening algorithm for a casting product digital radiographic image, in fig. 3, an a-rolling layer, a-down sampling layer, C-rolling layer, D-down sampling layer, e-full connection layer, f-output layer (full connection + Softmax activation), fig. 4 is a schematic diagram of a digital radiographic image positioning algorithm for a casting product digital radiographic image, in fig. 4, A-extraction of candidate blocks, B-N extraction feature extraction, C-Pong-Po classification, and C-Po classification.

The invention discloses an intelligent detection device for internal defects of a casting product, which comprises: the system comprises an X-ray tube, an imaging plate, a light path simulation and control unit and a cloud platform server; as shown in fig. 2;

as shown in fig. 1, the optical path simulation and control unit can optimally select an optimal transillumination optical path through the built-in simulation module according to the morphological characteristics or the three-dimensional model of the casting product, that is, X-rays are emitted from the ray tube and are perpendicular to the surface of the casting product or a tangent plane thereof to transmit the casting product to the imaging plate, or when the special structure of the casting product cannot be injected from the perpendicular surface, the shortest path of the X-rays to transmit the casting product is ensured; controlling the X-ray tube and the imaging plate according to the optimal transillumination pipeline, and adjusting the positions of the X-ray tube and the imaging plate to enable the casting product to be located at the optimal position between the X-ray tube and the imaging plate, namely the imaging plate is located on a focal plane where the focal point of the X-ray emitted by the X-ray tube is located; the light path simulation and control unit controls the X-ray tube to emit X-rays, the X-rays are imaged on the imaging plate after transmitting the casting product, and the light path simulation and control unit controls the imaging plate to obtain an X-ray image, namely an original gray image of the casting product; the light path simulation and control unit is used for pushing the original gray level image of the casting product to the cloud platform server;

the cloud platform server is used for preprocessing the original gray level image of the casting product to enable the contrast of the original gray level image to meet the requirement and obtain a preprocessed image; inputting the preprocessed image into a convolutional neural network model CNN, carrying out defect screening, and judging whether the preprocessed image contains defects. The cloud platform server can pack and integrate the intelligent defect detection method, preferably can be deployed on the Internet, and outputs cloud service; different products in the process of summarizing and detecting can be effectively collected to form casting digital ray image data sets of different products, and a foundation is provided for research and development of the defect intelligent detection method.

And inputting the image judged to contain the defect into a convolutional neural network model RCNN with an interested region, positioning the defect in the defective image, and labeling the minimum block diagram of the position of the defect to realize defect positioning.

And extracting the image in the minimum block diagram after the defect positioning, inputting a deep learning neural network model DNN, and classifying the defect to obtain the type of the defect.

The imaging plate is located at the focus of the X-ray tube radiation.

The method comprises the following steps of preprocessing an original gray level image of a casting product, and specifically comprises the following steps:

for the original gray-scale image with severe thickness gradient change, the preprocessing method includes but is not limited to: image enhancement, image nonlinear variation, filtering and noise reduction, image sharpening and the like.

For the rough-surface original gray-scale image, the preprocessing method includes but is not limited to: image enhancement, filtering and denoising, image smoothing and the like.

For the original gray-scale image of the edge of the casting, the preprocessing modes include but are not limited to: image enhancement, image non-linear change, filtering noise reduction, wavelet transformation, and the like.

When image preprocessing is actually performed, different preprocessing method combinations and parameters are tried, and the preprocessing method combination and the parameters with good effects are selected preferably.

The contrast requirement of the original grayscale image is preferably: the gray difference value of four adjacent pixel points of one pixel point meets the requirement of the minimum distinguishing limit of the detection threshold.

The defects preferably include pores, slag inclusions, shrinkage porosity, cracks, and segregation.

The convolutional neural network model CNN is trained in advance, preferably in the following manner:

①, establishing a training image set with labels, wherein the training images are images of casting products after the original gray level images are preprocessed, and the images are marked in advance to form the labels, wherein the labels are divided into defective labels and non-defective labels;

②, inputting the images in the training image set to the CNN, setting the structure, target function, model parameter optimization method of CNN according to the requirement, training CNN, not only setting the structure, target function, model parameter optimization method of CNN based on the VGG network model structure, but also setting the optimization scheme as the optimization scheme, using random function to initialize the model parameters based on VGG network model structure, defining the error between the model output and input as the target function, using error back propagation method to update the model parameters, using gradient descent method to find the parameters to update and descend the target function minimal value, using the above process to train CNN, as shown in FIG. 3, obtaining characteristic diagram after image input, obtaining characteristic diagram after convolution, repeating convolution and descent sampling process, inputting the finally obtained characteristic diagram into the full connection layer, outputting the trained convolution neural network model, obtaining the good convolution neural network model after training, namely obtaining the characteristic diagram after image input, repeating convolution and descent sampling process, inputting characteristic diagram, selecting the characteristic diagram finally obtained characteristic diagram, using AlCNN as the combined training algorithm, selecting the algorithm of the algorithm, selecting the algorithm, the.

The trained convolutional neural network model CNN can output labels, whether the casting products have defects or not is judged according to the labels, and defect screening is achieved.

The minimum block diagram of the position of the defect is preferably as follows: a group of parallel lines tangent to the longest diameter of the defect and a group of parallel lines vertical to the longest diameter of the defect (the position tangent to the outermost side of the edge of the defect) form a rectangular frame.

The convolutional neural network model with the region of interest, RCNN, needs to be trained in advance, and the preferred training mode is as follows:

① creating a training image set with a minimum defect block diagram, wherein the training images in the training image set have all possible defects of the casting product, each training image at least comprises one defect, and all possible defects comprise air holes, slag inclusion, shrinkage porosity, cracks and segregation.

②, inputting the images in the training image set with the minimum defect block diagram into the untrained convolutional neural network model RCNN with the region of interest one by one, training the convolutional neural network model RCNN with the region of interest according to the preset structure, objective function and model parameter optimization method of the convolutional neural network model RCNN with the region of interest, without being limited to the specific structure, objective function and model parameter optimization method, setting according to the requirements, preferably, initializing the model parameters by using a random function on the basis of the structure of the YoLO network model, defining the error between the output and input of the model as the objective function, updating the model parameters by using an error back propagation method, searching parameters by using a gradient descent method, updating and putting down to obtain the minimum value of the objective function, as shown in FIG. 4, extracting candidate frames on the input image, mapping the candidate frames to a characteristic diagram by using a CNN extraction method, adjusting the ROI to a fixed size by using ROI, classifying and obtaining the ROI after obtaining the characteristics, classifying and carrying out the classification, the training of the regression, namely, training the training of the region of the convolutional neural network model RCNN with the minimum defect block diagram, selecting the algorithm, and the algorithm can also select the algorithm of the algorithm, and the algorithm can be used for the algorithm.

The trained convolutional neural network model RCNN with the region of interest preferably can output a minimum block diagram of the position of the defect, and the defect location is realized according to the minimum block diagram of the position of the defect.

The deep learning neural network model DNN needs to be trained in advance, and the preferred training mode is as follows:

① creating a training image set with a defect type label and a minimum defect block diagram, wherein the defect type label of each training image with the minimum defect block diagram corresponds to a defect type;

②, inputting the images in the training image set with the defect type label and the defect minimum block diagram into the untrained deep learning neural network model DNN one by one, training the deep learning neural network model DNN according to the preset structure, objective function and model parameter optimization method of the deep learning neural network model DNN, setting the optimal scheme according to the requirement by using a random function to initialize the model parameters on the basis of the AlexNet network model structure, defining the error between the output and input of the model as an objective function, updating the model parameters by using an error back propagation method, searching parameters by using a gradient descent method to update and put down to obtain the minimum value of the objective function, training the deep learning neural network model DNN by using the process to obtain the trained deep learning neural network model DNN, namely optimizing the parameters in the deep learning neural network model DNN, performing various combination training on the DNN structure, the objective function and the model parameters, preferably selecting a gradient descent method of a dynamic quantity combination of a DNN, and the like, and performing a batch optimization on the DNN by using a gradient descent method of a gradient descent method, and the optimal selection method of a gradient descent method of a virtual learning neural network convolution model, a virtual network model, a virtual descent method of a virtual learning network convolution model, a virtual descent method of a virtual learning network, a virtual descent method of a virtual descent, a virtual descent method of a virtual descent method, a virtual descent method of a virtual descent, a virtual learning neural network, a virtual descent method of a virtual descent method.

The deep learning neural network model DNN can output labels with the defect types, and defect classification is achieved according to the labels with the defect types.

After the defects are classified, the classified defects can be graded, whether the defect grades meet different levels of patterns of the same type of defects input in advance or not is judged, and a detection result is output;

the invention carries out pretreatment, and further scheme for realizing defect contrast improvement is as follows: and (4) introducing an optimization algorithm, and searching a preprocessing combination with maximized defect contrast.

The invention sets the contrast requirement of the original gray level image, and further scheme for realizing the contrast improvement of the image is as follows: and adjusting the X-ray irradiation intensity according to the size and the material of the casting to be detected.

The convolutional neural network model CNN is trained, and the further scheme for realizing defect screening improvement is as follows: and selecting the structure of the residual error network ResNet as the structure of the volume and neural network model CNN for training.

The convolutional neural network model RCNN with the region of interest is trained, and the further scheme for realizing the defect positioning improvement is as follows: the structure of the Faster R-CNN model is selected as the structure of the volume with the region of interest and the neural network model RCNN for training.

The deep learning neural network model DNN is trained, and the further scheme for realizing defect classification improvement is as follows: optimizing learning rate, regularization term portion in an objective function of a deep learning neural network model DNN

After the defects are classified, the classified defects can be graded, whether the defect grades meet the inspection standard or not is judged, and the detection result is output, wherein the preferable scheme is as follows:

and after the defects are classified, a DNN network is applied to carry out defect rating on the classified defects, specifically, images to be rated are input into the trained network, labels representing defect grades are output, and the judgment basis is to evaluate whether the images meet different-grade patterns of the same type of defects input by DNN network trainers in advance.

The invention can realize the further proposal of improving the indexes of the device as follows: the further scheme for realizing the reduction of the missing judgment rate of the intelligent film evaluation comprises the following steps of writing the missing judgment rate into a target function and optimizing

In the detection process of the casting product, the invention overcomes the problems that the internal information of the casting product is photographed by a ray imaging technology which is usually adopted, and then the internal defects of the casting product are interpreted by experienced workers according to the photograph. And the problems that during quality detection, workers with high labor cost and rich experience are scarce, the period of technical workers with detection qualification is long, and manual detection is often accompanied by artificial errors such as low detection efficiency, erroneous judgment and missing judgment of detection results and the like along with the increase of detection working strength, and the product quality detection work is seriously influenced are solved.

By taking the casting cabin body as a further preferred scheme, the intelligent detection device and method for the defects of the casting products provided by the invention have the key processes as follows: (1) the X-ray source is arranged in the cylindrical cabin, the X-ray is emitted from the X-ray source and is emitted into the cabin perpendicular to the surface section of the casting cabin to form an optimal transillumination light path, the optimal transillumination light path is received by an imaging plate which is arranged outside the cabin and is positioned on a ray focal plane to form a digital ray original gray image of the casting cabin, and the digital ray original gray image is pushed to a cloud platform server. (2) The cloud platform server preprocesses the original gray level image of the digital ray of the casting cabin body, so that the contrast of the original gray level image meets the requirement, and the preprocessed image is obtained; (3) inputting the preprocessed image into a convolutional neural network model CNN, carrying out defect screening, and judging whether the preprocessed image contains defects. The convolutional neural network model CNN needs to be trained in advance, model parameters are initialized by using a random function on the basis of the VGG network model structure to obtain an untrained convolutional neural network model CNN, errors between model output and model input are defined as a target function, the model parameters are updated by adopting an error back propagation method, parameters are searched by using a gradient descent method, and the minimum value of the target function is obtained by putting down the parameters to update, so that the training of the convolutional neural network model CNN is completed. The trained convolutional neural network model CNN has defect screening capability, and can screen input images with defects. (4) And inputting the image judged to contain the defect into a convolutional neural network model RCNN with an interested region, positioning the defect in the defective image, and labeling the minimum block diagram of the position of the defect to realize defect positioning. The convolutional neural network model RCNN with the region of interest needs to be trained in advance, model parameters are initialized by using a random function on the basis of a YOLO network model structure to obtain an untrained convolutional neural network model RCNN with the region of interest, errors between model output and model input are defined as a target function, the model parameters are updated by adopting an error back propagation method, parameters are searched by using a gradient descent method, the parameters are updated and put down to obtain the minimum value of the target function, and the training of the convolutional neural network model RCNN with the region of interest is completed. The trained convolutional neural network model RCNN with the region of interest has defect positioning capacity, and can mark the minimum block diagram of the position of a defect to realize defect positioning. (5) And extracting the image in the minimum block diagram after the defect positioning, inputting a deep learning neural network model DNN, and classifying the defect to obtain the type of the defect. The deep learning neural network model DNN needs to be trained in advance, model parameters are initialized by using a random function on the basis of an AlexNet network model structure to obtain an untrained deep learning neural network model DNN, errors between model output and model input are defined as an objective function, model parameters are updated by adopting an error back propagation method, parameters are searched by using a gradient descent method, the parameters are updated and put down to obtain the minimum value of the objective function, and the deep learning neural network model DNN is trained. The trained deep learning neural network model DNN has defect classification capability, and can classify specific types (pores, slag inclusion, shrinkage porosity, cracks and segregation) of the defect images in the input minimum block diagram.

The invention carries out the film evaluation efficiency test, randomly extracts 100 images from the casting digital radiographic image, respectively sends the images to an experienced film evaluating person to evaluate the film and the intelligent defect detection device and method provided by the invention to evaluate the film, records the total time of manual film evaluation and machine film evaluation, and calculates the film evaluation time of each image. The test results showed that the manual sheet flattening time was 30.2 seconds per sheet and the machine sheet flattening time was 14.6 seconds per sheet. Compared with the traditional manual film evaluation mode, the intelligent defect detection device and method provided by the invention improve the efficiency by about 100%.

By contrast, the detection mode of a computer combined with a deep learning neural network artificial intelligence algorithm replaces the traditional artificial detection mode, the detection knowledge of experienced inspection skilled personnel is converted into a machine language which can be quantized by the computer, the images are analyzed by adopting an artificial intelligence new technology through digital images collected by X-ray detection equipment, the defects in the images are positioned, identified and graded, and the automatic and intelligent interpretation and output results are obtained. And under the condition that the detection data quantity is accumulated continuously, the detection accuracy is improved.