阅读说明：本技术 一种基于红外图像的电解槽极板故障识别方法、系统、终端及可读存储介质 (Electrolytic cell polar plate fault identification method, system, terminal and readable storage medium based on infrared image ) 是由 朱红求 彭天宇 阳春华 李勇刚 周灿 戴宇思 黄科科 阳波 彭磊 于 2021-08-23 设计创作，主要内容包括：本发明公开了一种基于红外图像的电解槽极板故障识别方法、系统、终端及可读存储介质,其包括以下步骤：获取电解槽极板的红外图像数据；基于红外图像数据构建电解槽极板故障识别模型,其中,电解槽极板故障识别模型构建过程为：基于红外图像数据依次进行深度特征提取、检测框生成、筛选检测框得到候选检测框、基于候选检测框对应的特征图进行分类网络训练和回归网络训练得到电解槽极板故障识别模型,从而利用电解槽极板故障识别模型得到待识别的电解槽极板的故障识别结果。本发明所述方法利用深度学习的优势提高了故障识别的准确率以及速度,将红外图像成功应用于电解槽的极板检测,并保证了检测精度。(The invention discloses an electrolytic cell polar plate fault identification method, a system, a terminal and a readable storage medium based on infrared images, which comprises the following steps: acquiring infrared image data of an electrolytic cell polar plate; an electrolytic cell polar plate fault identification model is constructed based on infrared image data, wherein the construction process of the electrolytic cell polar plate fault identification model is as follows: and sequentially carrying out depth feature extraction, detection frame generation and detection frame screening on the basis of infrared image data to obtain candidate detection frames, and carrying out classification network training and regression network training on the basis of feature maps corresponding to the candidate detection frames to obtain an electrolytic cell polar plate fault identification model, so that a fault identification result of the electrolytic cell polar plate to be identified is obtained by using the electrolytic cell polar plate fault identification model. The method improves the accuracy and speed of fault identification by using the advantages of deep learning, successfully applies the infrared image to the detection of the polar plate of the electrolytic cell and ensures the detection precision.)

1. An electrolytic cell polar plate fault identification method based on infrared images is characterized in that: the method comprises the following steps:

step 1: acquiring infrared image data of an electrolytic cell polar plate;

step 2: performing network training by taking the infrared image data in the step 1 as sample data to obtain an electrolytic cell polar plate fault identification model; the network of electrolytic cell plate fault identification models comprises at least: the method comprises the following steps of extracting a characteristic network, generating a region network, classifying a network and a regression network, wherein the training process of the electrolytic cell polar plate fault identification model comprises the following steps:

s2-1: inputting infrared image data into the feature extraction network to obtain a feature map of the sample;

s2-2: generating a network by inputting the feature map into the area to obtain a detection frame, and screening the detection frame to obtain a candidate detection frame;

s2-3: inputting the feature map corresponding to each candidate detection frame into a classification network and a regression network for training to obtain a fault identification model of the electrolytic cell polar plate;

the classification network is used for classifying the detection frame area, and the identification category at least comprises a fault category; the regression network is used for adjusting the detection frame boundary;

and step 3: and inputting the infrared image data of the electrolytic cell polar plate to be identified into the electrolytic cell polar plate fault identification model to obtain a fault identification result.

2. The method of claim 1, wherein: in step S2-2, an improved non-maximum suppression algorithm is used to screen the detection frames to obtain candidate detection frames, wherein the formula of the improved non-maximum suppression algorithm is as follows:

in the formula, M is the detection frame with the highest confidence corresponding to the calibration target, bi is the current detection frame, iou (M, bi) is the intersection ratio of the current detection frame and the detection frame with the highest confidence, s_iFor the original confidence level of the current detection box,for the confidence after reassignment, sigma is a confidence threshold adjustment coefficient;

wherein the confidence after reassignmentThe detection frame of 0 is eliminated, iou (M, bi) is less than or equal to a₁The detection frame area of (2) is regarded as a background, and the calibration target is a classification target marked in the sample.

3. The method of claim 2, wherein: a is₁The value of (A) is 0.35-0.45; a is₂The value of (A) is 0.8-0.9.

4. The method of claim 1, wherein: the identification types of the detection frames in the electrolytic cell polar plate fault identification model comprise: electrolyte outlet, electrolyte inlet, light short circuit, normal short circuit, and severe short circuit.

5. The method of claim 1, wherein: and 3, the number of the detection frames correspondingly arranged in each feature point in the area generation network is 4, the size area is {64 × 64,128 × 128}, and the aspect ratio is a combination of {1:4,1:2 }.

6. The method of claim 1, wherein: after acquiring the infrared image data of the electrode plate of the electrolytic cell in the step 1, the method further comprises the following steps: the method comprises the following steps of (1) carrying out standardization processing on original infrared image data, specifically:

step 1.1, denoising an image by using a self-adaptive mean filtering algorithm;

step 1.2, using a standard zinc electrolytic tank as a template, and performing single-tank segmentation on the obtained image;

step 1.3, correcting barrel distortion of the image, wherein the coordinate conversion relation between the undistorted image and the original distorted image is as follows:

in the formula, m and n are transverse coordinates and longitudinal coordinates in the undistorted image; for transverse sitting in distorted imagesA vertical coordinate is marked; the horizontal center coordinate and the vertical center coordinate in the distorted image; k is a radical of₁ k₂The distortion coefficient of the image can be obtained by manually calibrating the distorted image;

and 1.4, carrying out classified target calibration on the fault electrode plate in the infrared image of the zinc electrolytic cell after the standardized treatment.

7. The method of claim 1, wherein: the electrolytic bath is a zinc electrolytic bath.

8. A system based on the method of any one of claims 1-7, characterized by: the method comprises the following steps:

the image data acquisition module is used for acquiring infrared image data of the electrolytic cell polar plate;

the characteristic diagram generating module is used for inputting infrared image data into the characteristic extraction network to obtain a characteristic diagram of the sample;

the detection frame generation module is used for inputting the characteristic diagram into the area generation network to obtain a detection frame;

the detection frame screening module is used for screening the detection frames to obtain candidate detection frames;

the model training module is used for inputting the feature map corresponding to each candidate detection box into the classification network and the regression network for training;

and the detection module is used for inputting the infrared image data of the electrolytic cell polar plate to be identified into the electrolytic cell polar plate fault identification model to obtain a fault identification result.

9. A terminal, characterized by: the method comprises the following steps:

one or more processors;

a memory storing one or more programs;

the processor calls the program to implement:

the process steps of any one of claims 1 to 7.

10. A readable storage medium, characterized by: a computer program is stored, which is invoked by a processor to implement:

the process steps of any one of claims 1 to 7.

Technical Field

The invention belongs to the technical field of electrolytic bath fault detection, and particularly relates to an electrolytic bath polar plate fault identification method, system, terminal and readable storage medium based on infrared images.

Background

With the continuous development of zinc smelting process, zinc hydrometallurgy has become the mainstream zinc smelting method at present, and the process flow mainly comprises the working procedures of neutral leaching, purification, electrolysis and the like. The efficiency of electrolytic production directly affects the productivity and quality of finished zinc.

In the zinc electrolysis process, the cathode plate and the anode plate are alternately arranged and connected in parallel in a circuit, and are electrified through contact with a busbar, the pole plates are likely to be short-circuited due to the factors of too small pole plate distance, adhesion of anode mud, uneven current density and the like, the current efficiency can be seriously influenced by the short circuit, a large amount of electric energy is consumed, the temperature of the pole plates is also greatly increased, the overvoltage of hydrogen ions in electrolyte is reduced, the precipitation efficiency of elemental zinc is influenced, and the phenomenon of 'plate burning' can be caused by the serious short circuit. Therefore, the timely discovery of short-circuit faults has positive significance on the zinc electrolysis process.

At present, two methods are mainly used for detecting the short circuit of the polar plate, one method is a reed pipe hand-held slot checking device, the current condition is reflected through the magnetic field intensity, although the method has the advantages of simple principle, convenient operation and low cost, the sensitivity, the precision and the real-time property are difficult to meet the actual requirement, and the labor is very consumed. The other is to use a sensor to carry out contact measurement on the voltage and the current of the electrolytic cell, although the real-time performance of detection is improved, and the state of the polar plate can be directly reflected, under the high-acid and high-humidity environment, the sensor is easy to corrode, and the reliability and the maintainability of equipment are low.

Based on the drawbacks of the two detection methods, the infrared thermal imaging technology is concerned. The infrared thermal imaging technology is utilized to acquire the bath surface temperature of the whole electrolytic bath in an all-weather, remote and non-contact manner, and fault identification, classification and positioning are carried out through the temperature of the electrode plate. Although the state of the pole plate can be intuitively reflected through the infrared image, the problem of quickly, accurately and completely detecting the fault target in the image still exists. The traditional infrared image target detection methods mainly comprise a dynamic threshold segmentation method, a frame difference method and the like, and the methods have large positioning deviation degrees on a fault polar plate and are very easy to cause missed detection and false detection. The reason is that the environment of the electrolytic plant is severe and complex, if the covering cloth is incomplete, the normal polar plate can be exposed, and although no fault exists, the infrared image can show a high-temperature state; the acid mist can also influence the infrared imaging, and the emissivity of the groove surface is changed, so that the gray level of the infrared image is abnormal; when slight faults or tiny faults in the busbar area occur, the fault target is small, and the characteristic information is weakly expressed on the image. When a feature extraction rule is designed, the traditional target detection algorithm hardly considers the complex features, the pertinence is not strong, missing detection and false detection are easy to occur when the above conditions occur, and the identification effect is poor. According to the traditional infrared image target detection, interested areas with various scales are selected on an original image in a sliding window mode, feature extraction is carried out on each interested area, and then classification judgment is carried out by using a classifier, so that the algorithm has extremely large calculation amount and is not suitable for industrial fields.

Disclosure of Invention

The invention provides an electrolytic cell polar plate fault identification method, system, terminal and readable storage medium based on infrared images, aiming at the problem that the traditional infrared image target detection cannot be effectively adapted to the complex severe environment of an electrolytic plant and thus cannot be effectively applied to polar plate fault detection of the electrolytic plant. The method utilizes the advantages of the neural network to extract the depth features of the image to obtain the feature map, the feature map contains richer semantics compared with the original infrared image, and the polar plate features can be accurately extracted under the complex environment of an electrolysis workshop, so that a polar plate fault detection means based on the infrared image is realized, and the polar plate fault detection means are enriched.

On one hand, the invention provides an electrolytic bath pole plate fault identification method based on an infrared image, which comprises the following steps:

step 1: acquiring infrared image data of an electrolytic cell polar plate;

step 2: performing network training by taking the infrared image data in the step 1 as sample data to obtain an electrolytic cell polar plate fault identification model; the network of electrolytic cell plate fault identification models comprises at least: the method comprises the following steps of extracting a characteristic network, generating a region network, classifying a network and a regression network, wherein the training process of the electrolytic cell polar plate fault identification model comprises the following steps:

s2-1: inputting infrared image data into the feature extraction network to obtain a feature map of the sample;

s2-2: generating a network by inputting the feature map into the area to obtain a detection frame, and screening the detection frame to obtain a candidate detection frame;

s2-3: inputting the feature map corresponding to each candidate detection frame into a classification network and a regression network for training to obtain a fault identification model of the electrolytic cell polar plate;

the classification network is used for classifying the detection frame area, and the identification category at least comprises a fault category; the regression network is used for adjusting the detection frame boundary;

and step 3: and inputting the infrared image data of the electrolytic cell polar plate to be identified into the electrolytic cell polar plate fault identification model to obtain a fault identification result.

The invention extracts the depth characteristics of the image, the characteristic image contains richer semantics than the original image, and the characteristics of the complex scene of the electrolytic cell can be accurately extracted, thereby effectively relieving the influence of the severe environment of the electrolytic workshop and improving the fault detection precision.

The classification network in the electrolytic cell polar plate fault identification model is used for detecting the frame region classification, and the training process can refer to the training means of the existing classification network. The detection frame region classification can be set according to actual requirements, for example, the polar plate is marked in an infrared image sample, faults and non-faults are classified and marked, and other positions are taken as backgrounds; or the fault polar plate and the electrolyte inlet and outlet are marked in the infrared image sample, the polar plate is regarded as a foreground, fault types such as a slight short circuit, a common short circuit, a serious short circuit and the like are classified aiming at the polar plate, and other positions are regarded as backgrounds; in a word, the target calibration is carried out in the infrared image sample according to the classification requirement, so that in the training process, the training is carried out according to the category scores of the detection frames in the calibration target range and the actual category of the calibration target, and the trained classification network can output the category of the detection frames.

And the regression network in the electrolytic cell polar plate fault identification model is used for representing the boundary of the detection frame to obtain the detection frame with accurate position. The training process can refer to the training means of the existing regression network. For example, the training process may be understood as adjusting the boundary positions of the detection frames with reference to the calibration targets in the infrared image samples.

It should be understood that the fault recognition model for the trained electrolytic cell polar plate obtains the fault recognition result as the accurate position and the classification of the detection frame on the infrared image, so that the position of the fault polar plate can be determined.

Optionally, in step S2-2, the detection frame is screened by using an improved non-maximum suppression algorithm to obtain a candidate detection frame, where the formula of the improved non-maximum suppression algorithm is as follows:

in the formulaM is the detection frame with the highest confidence coefficient corresponding to the calibration target, bi is the current detection frame, iou (M, bi) is the intersection ratio of the current detection frame and the detection frame with the highest confidence coefficient, s_iFor the original confidence level of the current detection box,for the confidence after reassignment, sigma is a confidence threshold adjustment coefficient;

wherein the confidence after reassignmentThe detection frame of 0 is eliminated, iou (M, bi) is less than or equal to a₁The detection frame area of (2) is regarded as a background, and the calibration target is a classification target marked in the sample.

Alternatively, a₁Is 0.35-0.45, such as 0.4; a is₂Is 0.8-0.9, such as 0.85.

When a plurality of continuous polar plates are in fault in the electrolytic cell, a plurality of target detection frames are stacked, and if the frames are directly screened in a mode of reserving the maximum scoring frame and rejecting the frames with the intersection ratio exceeding the threshold value, the detection frames adjacent to the fault are possibly rejected, so that the detection is missed. The improved non-maximum value suppression algorithm provided by the invention only rejects the maximum cross-over ratio (the threshold value is a)₂) And then, reducing the confidence coefficient of the remaining cross-over and higher detection frames in a Gaussian weighting mode, and not rejecting the detection frames. The larger the intersection ratio, the greater the probability that the detection box is a redundant detection box, giving it greater penalty weight. Although the confidence is reduced, the detection box is still retained, thereby reducing the probability of missed detection. Are relatively small (threshold a)₁) The confidence of the detection frame is not adjusted.

Optionally, the identification category of the detection frame in the electrolytic cell pole plate fault identification model includes: electrolyte Outlet (Outlet), electrolyte Inlet (Inlet), light short (Slightshort), general short (general short), and severe short (severeshot).

The method includes the steps of setting a slight short (Slightshort), a general short (Generalshort) and a severe short (Severeshort) based on conventional knowledge in the field and application requirements, wherein the method is not particularly limited in this respect, and the classification standard of the model is based on the classification standard in the sample target calibration.

Optionally, the number of the detection frames correspondingly arranged in each feature point in the area generation network in step 3 is 4, the size area is {64 × 64,128 × 128}, and the aspect ratio is a combination of {1:4,1:2 }.

The invention sets the number of the anchor frames to be 4 from the actual characteristics of the electrolytic cell polar plate, reduces the number of the detection frames, improves the initial size, is more suitable for the size of the electrolytic cell polar plate, and greatly reduces the regression calculation amount of the boundary frame in the subsequent training process.

Optionally, after acquiring the infrared image data of the electrolytic cell plate in step 1, the method further includes: the method comprises the following steps of (1) carrying out standardization processing on original infrared image data, specifically:

step 1.1, denoising an image by using a self-adaptive mean filtering algorithm;

step 1.2, using a standard zinc electrolytic tank as a template, and performing single-tank segmentation on the obtained image;

step 1.3, correcting barrel distortion of the image, wherein the coordinate conversion relation between the undistorted image and the original distorted image is as follows:

in the formula, m and n are transverse coordinates and longitudinal coordinates in the undistorted image;the horizontal coordinate and the vertical coordinate in the distorted image;the horizontal center coordinate and the vertical center coordinate in the distorted image; k is a radical of₁k₂As a distortion coefficient of the imageThe distortion coefficient can be obtained by manually calibrating the distorted image. Such as: k is a radical of₁＝1.1×10^-4，k₂＝4.4×10^-6。

And 1.4, carrying out classified target calibration on the fault electrode plate in the infrared image of the zinc electrolytic cell after the standardized treatment.

Optionally, the electrolytic cell is a zinc electrolytic cell.

In a second aspect, the invention provides a system based on the method for identifying the fault of the electrode plate of the electrolytic cell, which comprises:

the image data acquisition module is used for acquiring infrared image data of the electrolytic cell polar plate;

the characteristic diagram generating module is used for inputting infrared image data into the characteristic extraction network to obtain a characteristic diagram of the sample;

the detection frame generation module is used for inputting the characteristic diagram into the area generation network to obtain a detection frame;

the detection frame screening module is used for screening the detection frames to obtain candidate detection frames;

the model training module is used for inputting the feature map corresponding to each candidate detection box into the classification network and the regression network for training;

and the detection module is used for inputting the infrared image data of the electrolytic cell polar plate to be identified into the electrolytic cell polar plate fault identification model to obtain a fault identification result.

In a third aspect, the present invention provides a terminal, comprising:

one or more processors;

a memory storing one or more programs;

the processor calls the program to implement:

the electrolytic cell pole plate fault identification method based on the improved Mask-RCNN comprises the following steps.

In a fourth aspect, the present invention provides a readable storage medium storing a computer program for invocation by a processor to implement:

the electrolytic cell pole plate fault identification method based on the improved Mask-RCNN comprises the following steps.

Advantageous effects

1. According to the electrolytic cell polar plate fault identification method based on the infrared image, the depth characteristics of the image are extracted, the characteristic image contains richer semantics than the original image, more dimensional information can be obtained, and the characteristics of the complex scene of the electrolytic cell can be accurately extracted, so that the influence of the severe environment of an electrolytic plant is effectively relieved, and the fault detection precision is improved. Therefore, the method of the present invention is more effective than the conventional algorithm that performs recognition based on features such as a single pixel or a contour texture.

2. In a further preferred scheme of the invention, an improved non-maximum suppression algorithm is introduced to screen the detection frames, so that the number of the detection frames is reduced, and particularly, the improved non-maximum suppression algorithm provided by the invention only eliminates the maximum cross-over-cross ratio (the threshold value is a)₂) The confidence of the detection frames which are left to be crossed and are higher is reduced, and the probability of missed detection is reduced.

Drawings

FIG. 1 is a schematic flow chart of the construction of an electrolytic bath electrode plate fault identification model in the zinc electrolytic bath electrode plate fault identification method based on infrared images, which is disclosed by the invention;

FIG. 2 is a Mask-RCNN structure diagram (a network structure diagram of an electrolytic cell plate fault recognition model) of the zinc electrolytic cell plate fault recognition method based on infrared images;

FIG. 3 is a diagram of the RPN layer structure of the method for identifying the fault of the zinc electrolytic cell plate based on the infrared image, wherein (a) is the network structure and (b) is a schematic diagram of the RPN generation detection frame corresponding to the initial position of the detection frame;

FIG. 4 is an infrared image of a zinc electrolytic cell plate of the infrared image-based zinc electrolytic cell plate fault identification method of the present invention, wherein (a) is a black-and-white image of an original infrared image, which is seriously distorted, and (b) is a black-and-white image of a corrected infrared image, which is improved in distortion;

fig. 5 is a black-and-white image of the zinc electrolytic cell plate fault detection result image of the infrared image-based zinc electrolytic cell plate fault identification method according to the present invention, which classifies and locates the fault and includes the detection frame, the corresponding category and the score thereof.

Detailed Description

The invention provides an electrolytic cell polar plate fault identification method based on infrared images, which utilizes a neural network to construct an electrolytic cell polar plate fault identification model for electrolytic cell polar plate fault detection. For a better understanding of the present invention, the contents of the present invention are further described in detail below with reference to examples. The specific embodiments described herein are merely illustrative of the invention and are not intended to be limiting, i.e., the described embodiments are only some, but not all, of the embodiments of the invention.

Example 1:

1. acquisition and normalization processing of data sets

In this embodiment, in a zinc electrolysis plant, the electrolytic cell includes a plurality of electrolytic cells, and the array arrangement is basically presented in single electrolytic cell, and the negative and positive plates are placed alternately, and the both ends of electrolytic cell are electrolyte entry and export.

The infrared camera is arranged at a position 3-5 meters above the electrolytic cell according to the field angle of the infrared camera, so that the whole cell surface can be covered. The operator uses an infrared camera to shoot the electrolytic cell to obtain original infrared image data.

After field data are collected, firstly, a noise point of an original image is removed by a self-adaptive mean value filtering algorithm. In addition, since the collected picture includes the whole electrolytic cell, in order to conveniently position the fault board, a standard cell is used as a template, and a standardized difference sum of squares matching method is used for extracting a single-cell image. The barrel distortion of the image is caused by the inconsistent refractive indexes of the lenses of the wide-angle lens of the camera, so the image needs to be corrected according to the distortion rate, and a standard training sample is obtained.

Through a series of operations, 2000 pieces of experimental data are obtained in total, data labels are marked by professional operators to ensure the accuracy of training data, 1600 pieces of the experimental data are divided into training sets, and the remaining 400 pieces of the experimental data are divided into verification sets without test sets.

In the embodiment, target calibration is performed on a fault polar plate and an electrolyte inlet and outlet in an infrared image sample, the polar plate is taken as a foreground, fault types such as a slight short circuit, a common short circuit and a serious short circuit are classified for the polar plate, and other positions are taken as a background. Therefore, the detection frame categories in this embodiment include: electrolyte Outlet (Outlet), electrolyte Inlet (Inlet), light short (Slightshort), general short (general short), and severe short (severeshot).

In other possible embodiments, the target calibration may be adjusted according to the requirement of fault classification, for example, the target calibration is performed on the electrode plate in the infrared image sample, and the fault and the non-fault are classified and marked, and other positions are regarded as the background; or dividing the fault types by different standards. The invention is not particularly limited in this regard, but it should be understood that the sample markings necessarily include target calibration of faulty plates, thereby enabling the trained network to be used for plate fault detection in an electrolytic cell.

2. Design detection network (network structure of electrolytic bath polar plate fault identification model)

Firstly, depth features of an image are extracted by using a ResNet101 residual network, and then 4 detection frames with different scales are generated on a feature layer for each feature point by using a region generation network (RPN).

Because the size of the polar plate to be detected is fixed, only the shooting angle of the camera slightly affects the size of the polar plate in the image, a detection frame with the length-width ratio of 1:4 is mainly selected to detect a fault target according to the size characteristics of the polar plate, but in a slight short-circuit state, a heating point only appears near a contact point of the polar plate and a busbar, and therefore the detection frame with the length-width ratio of 1:2 is also necessary. Compared with the traditional generation strategy, the arrangement of the embodiment can reduce the detection frames by about 50%, reduce the regression calculation amount of partial boundary frames and improve the identification speed of the fault polar plate.

It should be understood that many detection frames are obtained through the RPN, and a calibration target range corresponds to a series of detection frames, so that it is necessary to perform screening, in this embodiment, an improved non-maximum suppression algorithm is preferably used for screening, and in other possible embodiments, screening may be performed in other manners.

In this embodiment, the improved non-maximum suppression algorithm eliminates or penalizes the detection frame with high intersection-to-parallel ratio. The following were used:

in the formula, M is the detection frame with the highest confidence corresponding to the calibration target, bi is the current detection frame, iou (M, bi) is the intersection ratio of the current detection frame and the detection frame with the highest confidence, s_iFor the original confidence level of the current detection box,for the confidence after reassignment, sigma is a confidence threshold adjustment coefficient; wherein the confidence after reassignmentThe detection frame of 0 is eliminated, iou (M, bi) is less than or equal to a₁The detection frame area of (2) is regarded as a background, and the calibration target is a classification target marked in the sample.

In the screening process, aiming at a series of detection frames corresponding to each calibration target, determining the detection frame with the highest confidence coefficient, wherein the confidence coefficient is the foreground score (probability of foreground) of the target in each frame given by the RPN network; therefore, the confidence of the detection frames is adjusted according to the formula, partial detection frames are eliminated, and the scores of the partial detection frames are adjusted. And separates the foreground and background (i.e., positive and negative samples) based on the foreground-background score. After screening, the detection frame corresponding to each calibration target is effectively screened.

Then, the remaining detection frames are input into the roiign layer, the feature map size corresponding to each detection frame is scaled to be consistent (in this embodiment, 7 × 7), and then the feature map size is input into the classification network and the regression network respectively to perform target classification and bounding box adjustment (network training). The ROIAlign layer functions to scale the feature map size corresponding to the detection box to a consistent level, and the implementation process of the ROIAlign layer refers to the prior art, which is not limited by the present invention. However, it is preferable to eliminate the operation of quantizing the feature image values and boundaries, calculate the image values by using a bilinear interpolation method, and then perform the maximum pooling operation.

The classification network is a full connection layer connected behind the roilign layer, the number of nodes is set to 6 in the embodiment, and the Softmax classifier is connected, because 5 categories are output, namely an electrolyte outlet, an electrolyte inlet, a slight short circuit, a common short circuit, a severe short circuit and a background. The regression network is connected with the classification network in parallel at a ROIAlign layer rear full connection layer, the number of nodes is set to be 24 in the embodiment, and the regression boundary box Regressor is connected with the regression network, and because 5 categories and 1 background have 4 boundary box regression parameters, the detection frames are classified, the boundaries of the detection frames are adjusted, and the accurate positions of the detection frames are obtained.

3. Model training and assessment

And setting hyper-parameters and iteration times of the model, wherein if the basic learning rate is set to be 0.001, the momentum factor is set to be 0.9, the attenuation value is set to be 0.0005, the optimization method is a random gradient descent algorithm (SGD), and the iteration times are 10000.

The training set is loaded into a constructed Mask-RCNN model (the electrolytic bath electrode plate fault identification model of the embodiment) and training is started. In this embodiment, the number of images in the training set and the verification set preferably accounts for 80% and 20% of the total data set, respectively.

And after iteration is completed, evaluating the performance of the model, and if the precision ratio and the recall ratio can both reach 90% and the multi-class average accuracy (mAP) is not lower than 85%, indicating that the model has better fault identification capability, otherwise, continuously repeating the steps and continuing training.

In summary, the method for identifying the fault of the electrode plate of the electrolytic cell based on the infrared image in the embodiment includes the following steps:

step 1: and acquiring infrared image data of the electrolytic cell polar plate.

For example, collecting infrared image data of the zinc electrolytic cell plate, standardizing the infrared image data, and then dividing the infrared image data into a training set and a verification set.

Step 2: performing network training by taking the infrared image data in the step 1 as sample data to obtain an electrolytic cell polar plate fault identification model;

and step 3: and inputting the infrared image data of the electrolytic cell polar plate to be identified into the electrolytic cell polar plate fault identification model to obtain a fault identification result.

Preprocessing infrared image data of the electrolytic cell polar plate to be identified, extracting depth features, generating detection frames, screening out candidate detection frames, obtaining the category and the accurate detection frame from a classification network and a regression network respectively, and finally obtaining the fault area and the type of the polar plate.

The training process of the electrolytic cell polar plate fault identification model is as follows:

s2-1: inputting infrared image data into the feature extraction network (such as a residual error network) to obtain a feature map of the sample;

s2-2: generating a network (RPN) by using the characteristic diagram input area to obtain a detection frame, and screening the detection frame by using an improved non-maximum suppression algorithm (NMS) to obtain a candidate detection frame;

s2-3: and inputting the characteristic diagram corresponding to each candidate detection frame into a classification network and a regression network for training to obtain the electrolytic cell polar plate fault identification model.

In this embodiment, the optimization method is a random gradient descent algorithm (SGD), the basic learning rate is 0.001, the attenuation value is 0.0005, the momentum factor is 0.9, and the number of iterations is 10000. And (3) performing performance evaluation on the trained model, if the precision ratio and the recall ratio can reach 80%, and the multi-class average accuracy (mAP) is not lower than 75%, indicating that the model has better fault identification capability, and if the requirement is not met, repeating the step 5 and continuing to perform iterative optimization.

It should be understood that if the precision or the number of detection frames does not meet the preset requirement, the iterative operation may be repeated. The iterative process may be based on the regression network to obtain the accurate detection box, and then the procedure returns to step S2-2 again for screening and training. When the electrolytic cell pole plate fault identification model is used for identification in the step 3, if the number of the detection frames is too high, the detection frames can be screened by adopting an improved non-maximum suppression algorithm to obtain a final detection frame, which is not specifically limited by the invention.

Example 2:

the embodiment provides a system based on an electrolytic cell pole plate fault identification method, which comprises the following steps: the device comprises an image data acquisition module, an electrolytic cell polar plate fault identification model construction module and a detection module, wherein the electrolytic cell polar plate fault identification model construction module comprises a characteristic diagram generation module, a detection frame screening module and a model training module.

The image data acquisition module is used for acquiring infrared image data of the electrode plate of the electrolytic cell; the electrolytic cell polar plate fault identification model building module is used for carrying out network training by taking infrared image data as sample data to obtain an electrolytic cell polar plate fault identification model.

The characteristic diagram generating module is used for inputting infrared image data into the characteristic extraction network to obtain a characteristic diagram of the sample; the detection frame generation module is used for inputting the characteristic diagram into the area generation network to obtain a detection frame; the detection frame screening module is used for screening the detection frames to obtain candidate detection frames;

the model training module is used for inputting the feature map corresponding to each candidate detection box into the classification network and the regression network for training;

and the detection module is used for inputting the infrared image data of the electrolytic cell polar plate to be identified into the electrolytic cell polar plate fault identification model to obtain a fault identification result.

In other feasible modes, the characteristic diagram generation module, the detection frame screening module and the model training module work as independent modules and are not integrated into an electrolytic bath pole plate fault identification model construction module. That is, the division of the functional module unit is only a division of a logic function, and there may be another division manner in actual implementation.

For the specific implementation process of each unit module, refer to the corresponding process of the foregoing method. It should be understood that, the specific implementation process of the above unit module refers to the method content, and the present invention is not described herein in detail, and the division of the above functional module unit is only a division of a logic function, and there may be another division manner in the actual implementation, for example, multiple units or components may be combined or may be integrated into another system, or some features may be omitted, or may not be executed. Meanwhile, the integrated unit can be realized in a hardware form, and can also be realized in a software functional unit form.

Example 3:

the present embodiment provides a terminal, which includes a processor and a memory, where the memory stores one or more programs, and the processor calls the programs to implement:

step 1: acquiring infrared image data of an electrolytic cell polar plate;

step 2: performing network training by taking the infrared image data in the step 1 as sample data to obtain an electrolytic cell polar plate fault identification model; the network of electrolytic cell plate fault identification models comprises at least: the method comprises the following steps of extracting a characteristic network, generating a region network, classifying a network and a regression network, wherein the training process of the electrolytic cell polar plate fault identification model comprises the following steps:

s2-1: inputting infrared image data into the feature extraction network to obtain a feature map of the sample;

s2-2: generating a network by inputting the feature map into the area to obtain a detection frame, and screening the detection frame to obtain a candidate detection frame;

s2-3: inputting the feature map corresponding to each candidate detection frame into a classification network and a regression network for training to obtain a fault identification model of the electrolytic cell polar plate;

the classification network is used for classifying the detection frame area, and the identification category at least comprises a fault category; the regression network is used for adjusting the detection frame boundary;

and step 3: and inputting the infrared image data of the electrolytic cell polar plate to be identified into the electrolytic cell polar plate fault identification model to obtain a fault identification result.

The terminal further includes: and the communication interface is used for communicating with external equipment and carrying out data interactive transmission.

The memory may include high speed RAM memory, and may also include a non-volatile defibrillator, such as at least one disk memory.

If the memory, the processor and the communication interface are implemented independently, the memory, the processor and the communication interface may be connected to each other through a bus and perform communication with each other. The bus may be an industry standard architecture bus, a peripheral device interconnect bus, an extended industry standard architecture bus, or the like. The bus may be divided into an address bus, a data bus, a control bus, etc.

Optionally, in a specific implementation, if the memory, the processor, and the communication interface are integrated on a chip, the memory, the processor, that is, the communication interface may complete communication with each other through the internal interface.

The specific implementation process of each step refers to the explanation of the foregoing method.

It should be understood that in the embodiments of the present invention, the Processor may be a Central Processing Unit (CPU), and the Processor may also be other general purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other Programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The memory may include both read-only memory and random access memory, and provides instructions and data to the processor. The portion of memory may also include non-volatile random access memory. For example, the memory may also store device type information.

Example 4:

the present embodiments provide a readable storage medium storing a computer program for invocation by a processor to implement:

step 1: acquiring infrared image data of an electrolytic cell polar plate;

step 2: performing network training by taking the infrared image data in the step 1 as sample data to obtain an electrolytic cell polar plate fault identification model; the network of electrolytic cell plate fault identification models comprises at least: the method comprises the following steps of extracting a characteristic network, generating a region network, classifying a network and a regression network, wherein the training process of the electrolytic cell polar plate fault identification model comprises the following steps:

s2-1: inputting infrared image data into the feature extraction network to obtain a feature map of the sample;

s2-2: generating a network by inputting the feature map into the area to obtain a detection frame, and screening the detection frame to obtain a candidate detection frame;

s2-3: inputting the feature map corresponding to each candidate detection frame into a classification network and a regression network for training to obtain a fault identification model of the electrolytic cell polar plate;

the classification network is used for classifying the detection frame area, and the identification category at least comprises a fault category; the regression network is used for adjusting the detection frame boundary;

and step 3: and inputting the infrared image data of the electrolytic cell polar plate to be identified into the electrolytic cell polar plate fault identification model to obtain a fault identification result.

The specific implementation process of each step refers to the explanation of the foregoing method.

The readable storage medium is a computer readable storage medium, which may be an internal storage unit of the controller according to any of the foregoing embodiments, for example, a hard disk or a memory of the controller. The readable storage medium may also be an external storage device of the controller, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), and the like provided on the controller. Further, the readable storage medium may also include both an internal storage unit of the controller and an external storage device. The readable storage medium is used for storing the computer program and other programs and data required by the controller. The readable storage medium may also be used to temporarily store data that has been output or is to be output.

Based on such understanding, the technical solution of the present invention essentially or partially contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product stored in a storage medium and including instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned readable storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read-only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disk.

It should be emphasized that the examples described herein are illustrative and not restrictive, and thus the invention is not to be limited to the examples described herein, but rather to other embodiments that may be devised by those skilled in the art based on the teachings herein, and that various modifications, alterations, and substitutions are possible without departing from the spirit and scope of the present invention.