Color-coated sheet surface defect identification method and system based on machine learning

Abstract

The invention discloses a color-coated sheet surface defect identification method and system based on machine learning, and relates to the technical field of defect identification. According to the color-coated sheet surface defect identification method and system based on machine learning, the color-coated sheet surface image and electrochemical data are collected, the related features are extracted, the fusion weight is determined, the feature vector is fused, and then the CNN-LSTM hybrid prediction model for reinforcement learning training is used for identification, so that fusion analysis of the image and electrochemical features is realized; by means of machine learning and reinforcement learning, the accuracy of defect recognition and the model generalization ability are improved, the color-coated sheet surface defects can be recognized more comprehensively and accurately, an effective basis is provided for production quality control, and the problems that a traditional recognition method is difficult to comprehensively utilize multi-source data, low in recognition accuracy and incapable of dynamically adapting to complex defect scenes can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of defect recognition, and in particular to a method and system for recognizing surface defects of color-coated plates based on machine learning. Background Art

[0002] With the development of society, the demand for color-coated steel sheets is increasing, and so are the quality requirements. With the development of consumer industries such as automobiles and home appliances, the ever-increasing surface quality requirements for color-coated steel sheets have led to increasingly prominent issues with surface quality. According to statistics, the vast majority of quality disputes and complaints between color-coated steel sheet manufacturers and users in recent years have been related to surface quality. Therefore, both color-coated steel sheet manufacturers and users attach great importance to surface quality testing.

[0003] Surface defect detection of strip steel using machine vision and artificial intelligence methods has become a research focus and development trend in the field of industrial quality inspection. However, existing defect recognition methods for color-coated steel sheets have difficulty in comprehensively utilizing multi-source data, suffer from low recognition accuracy, and are unable to dynamically adapt to complex defect scenarios. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method and system for identifying surface defects of color-coated steel plates based on machine learning, which solves the problems that the existing color-coated steel plate defect identification methods are difficult to comprehensively utilize multi-source data, have low recognition accuracy, and cannot dynamically adapt to complex defect scenarios.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for identifying surface defects of color-coated plates based on machine learning, comprising the following steps: performing surface image acquisition and electrochemical detection on the color-coated plates to obtain surface image data and electrochemical feature data of the color-coated plates; extracting surface defect feature data and surface defect degree feature data from the surface image data of the color-coated plates; extracting electrochemical defect feature data and electrochemical defect degree feature data from the electrochemical feature data; determining fusion weight data of the surface defect feature data and the electrochemical defect feature data based on the surface defect degree feature data and the electrochemical defect degree feature data; fusing the surface defect feature data and the electrochemical defect feature data based on the fusion weight data to determine a fusion feature vector; and inputting the fusion feature vector into a CNN-LSTM hybrid prediction model trained by reinforcement learning to obtain a surface defect recognition result of the color-coated plates.

[0006] Furthermore, the surface defect degree characteristic data includes the defect impact area ratio index , defect shape complexity index and defect depth influence coefficient ;Electrochemical defect degree characteristic data including corrosion rate change index and electrochemical parameter abnormality index ; Determining the fusion weight data of the surface defect characteristic data and the electrochemical defect characteristic data includes the following steps: calculating the surface defect degree evaluation coefficient based on the surface defect degree evaluation model ; Calculate the electrochemical defect degree evaluation coefficient based on the electrochemical defect degree evaluation model ; Get the defect degree matching data stored in the database - fusion weight mapping set, and convert the surface defect degree evaluation coefficient and electrochemical defect degree evaluation coefficient Perform cosine similarity comparison with each defect degree matching data in the defect degree matching data-fusion weight mapping set to determine the cosine similarity value; determine the defect degree matching data corresponding to the largest cosine similarity value, and obtain the corresponding fusion weight data.

[0007] Furthermore, the surface defect degree evaluation coefficient is calculated based on the surface defect degree evaluation model. , including the following steps: obtaining the surface defect degree benchmark parameter data, including the defect impact area ratio index benchmark value , defect shape complexity index benchmark value and defect depth influence coefficient benchmark value ; Input the surface defect degree characteristic data and surface defect degree benchmark parameter data into the surface defect degree evaluation model to obtain the surface defect degree evaluation coefficient , the surface defect degree evaluation model is as follows: ; Calculation of electrochemical defect degree evaluation coefficient based on electrochemical defect degree evaluation model , including the following steps: obtaining electrochemical defect degree benchmark parameter data, including the corrosion rate change rate index benchmark value and the baseline value of the abnormality index of electrochemical parameters ; Input the electrochemical defect degree characteristic data and the electrochemical defect degree benchmark parameter data into the electrochemical defect degree evaluation model to obtain the electrochemical defect degree evaluation coefficient , the electrochemical defect degree evaluation model is as follows: .

[0008] Furthermore, the method for obtaining the defect degree matching data-fusion weight mapping set is as follows: obtain the historical surface defect feature data and historical electrochemical defect feature data of the historical color-coated plate; process the historical surface defect feature data and historical electrochemical defect feature data based on the surface defect degree assessment model and the electrochemical defect degree assessment model respectively to obtain the historical surface defect degree assessment coefficient and the historical electrochemical defect degree assessment coefficient; record the historical surface defect degree assessment coefficient and the historical electrochemical defect degree assessment coefficient as defect degree matching data; randomly assign several groups of fusion weights to each defect degree matching data; normalize the defect degree matching data and fuse it with the fusion weight to obtain the fused historical defect feature vector; determine the defect degree matching data-fusion weight mapping set based on the fused historical defect feature vector.

[0009] Furthermore, the defect degree matching data-fusion weight mapping set is determined based on the fused historical defect feature vectors, including the following steps: the fused historical defect feature vectors corresponding to the same fusion weight are respectively input into the constructed initial CNN prediction model and the initial LSTM prediction model for pre-training; after the pre-training is completed, the performance data of the pre-trained initial CNN prediction model and the initial LSTM prediction model corresponding to the same fusion weight are counted, including the CNN prediction model performance data and the LSTM prediction model performance data; the performance coefficient is obtained based on the analysis of the CNN prediction model performance data and the LSTM prediction model performance data; the fusion weight corresponding to the largest performance coefficient is recorded as the optimal fusion weight; the defect degree matching data and the optimal fusion weight are matched one by one to obtain the defect degree matching data-fusion weight mapping set.

[0010] Furthermore, the CNN prediction model performance data includes CNN prediction accuracy , CNN prediction accuracy and CNN recall , LSTM prediction model performance data includes LSTM prediction accuracy , LSTM prediction accuracy and LSTM recall ; Get the performance coefficient , including the following steps: ; in, and are all weight factors.

[0011] Furthermore, the CNN-LSTM hybrid prediction model is trained through reinforcement learning, including the following steps: building an initial CNN-LSTM hybrid prediction model, including a CNN layer, an LSTM layer and a fully connected layer, wherein the CNN layer is used to process the surface image data of the color-coated plate to obtain surface defect feature data, the LSTM layer is used to process the electrochemical feature data to obtain electrochemical defect feature data, and the fully connected layer is used to perform weighted fusion and splicing of the surface defect feature data and the electrochemical defect feature data to obtain a comprehensive vector, and obtain the surface defect recognition result of the color-coated plate through multiple neuron processing; the initial CNN-LSTM hybrid prediction model is trained and optimized based on the reinforcement learning algorithm to obtain a trained CNN-LSTM hybrid prediction model.

[0012] Furthermore, the initial CNN-LSTM hybrid prediction model is trained and optimized based on the reinforcement learning algorithm, including the following steps: defining the parameters of the current CNN layer, the parameters of the LSTM layer and the comprehensive vector as states; obtaining an action set stored in a database, wherein each action in the action set corresponds to a set of parameter adjustment values, including CNN layer parameter adjustment values ​​and LSTM layer parameter characteristic values; determining the reward value corresponding to each action stored in the database, adjusting the parameters of the current CNN layer and the LSTM layer based on the action corresponding to the maximum reward value, and updating the state; obtaining a prediction performance evaluation coefficient after each parameter adjustment, and stopping the adjustment if the prediction performance evaluation coefficient is greater than the set performance evaluation threshold; otherwise, making an adjustment and updating the reward value corresponding to the action at the same time.

[0013] Furthermore, obtaining the prediction performance evaluation coefficient includes the following steps: obtaining the prediction performance data of the initial CNN-LSTM hybrid prediction model after parameter adjustment, including the prediction accuracy , the ratio of the predicted time to the set benchmark time and load occupancy , perform weighted calculation on the prediction performance data to obtain the prediction performance evaluation coefficient : ;in, , and are weight factors; The process of updating the reward value corresponding to the action when the predicted performance evaluation coefficient is not greater than the set performance evaluation threshold is as follows: subtract the predicted performance evaluation coefficient from the performance evaluation threshold to obtain the difference; compare the difference with the corresponding matching difference stored in the database to determine the update value corresponding to the matching difference; add the reward value and the update value to obtain the updated reward value.

[0014] The machine learning-based color-coated plate surface defect recognition system comprises a data acquisition module for surface image acquisition and electrochemical detection of the color-coated plate to obtain color-coated plate surface image data and electrochemical characteristic data; a feature extraction module for extracting surface defect feature data and surface defect degree feature data from the color-coated plate surface image data, and extracting electrochemical defect feature data and electrochemical defect degree feature data from the electrochemical characteristic data; a fusion weight acquisition module for determining fusion weight data of the surface defect feature data and the electrochemical defect feature data based on the surface defect degree feature data and the electrochemical defect degree feature data; a vector fusion module for fusing the surface defect feature data and the electrochemical defect feature data based on the fusion weight data to determine a fusion feature vector; and a defect recognition module for inputting the fusion feature vector into a CNN-LSTM hybrid prediction model trained through reinforcement learning to obtain a color-coated plate surface defect recognition result.

[0015] The present application has the following advantages: The machine learning-based color-coated plate surface defect recognition method and system collect color-coated plate surface images and electrochemical data, extract relevant features, determine fusion weights and fuse feature vectors, and then use a CNN-LSTM hybrid prediction model trained through reinforcement learning to recognize, realize fusion analysis of images and electrochemical features, improve the accuracy of defect recognition and the generalization ability of the model with the aid of machine learning and reinforcement learning, and more comprehensively and accurately recognize color-coated plate surface defects, providing an effective basis for production quality control and solving the problems of traditional recognition methods, such as difficulty in comprehensively utilizing multi-source data, low recognition accuracy, and inability to dynamically adapt to complex defect scenarios.

[0016] Of course, implementing any product of the present application does not necessarily require achieving all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The present application is a machine learning-based color-coated plate surface defect recognition method flowchart.

[0018] Figure 2 The present application is a machine learning-based color-coated plate surface defect recognition system flowchart. DETAILED DESCRIPTION

[0019] Please refer to Figure 1 The present application provides a machine learning-based color-coated plate surface defect recognition method, comprising the following steps: surface image acquisition and electrochemical detection of the color-coated plate to obtain color-coated plate surface image data and electrochemical characteristic data; extracting surface defect feature data and surface defect degree feature data from the color-coated plate surface image data; and extracting electrochemical defect feature data and electrochemical defect degree feature data from the electrochemical characteristic data.

[0020] Surface defect feature data include normalized geometric features (defect perimeter, area, circularity, aspect ratio), texture features (contrast, entropy, angular second moment, correlation calculated from gray-level co-occurrence matrix) and color features (mean and variance of hue, saturation, and brightness); electrochemical defect feature data include normalized open circuit potential, polarization potential, coating resistance, double-layer capacitance, charge transfer resistance, and corrosion current density.

[0021] Surface defect degree characteristic data including defect impact area ratio index , defect shape complexity index and defect depth influence coefficient ;Electrochemical defect degree characteristic data including corrosion rate change index and electrochemical parameter abnormality index .

[0022] The defect shape complexity index is the normalized fractal dimension. The defect depth influence coefficient is obtained by first acquiring defect depth information using nondestructive testing technology, then determining the defect depth level based on the corresponding defect depths stored in the database. The electrochemical parameter abnormality index is the number of abnormal parameters.

[0023] By capturing surface images and conducting electrochemical testing on color-coated steel sheets and extracting relevant features, we can comprehensively characterize surface defects in color-coated steel sheets from multiple dimensions. The normalized geometric, texture, and color features of the surface defect feature data accurately describe the visual morphology of the defect; the normalized parameters of the electrochemical defect feature data reflect the impact of the defect on the electrochemical performance of the coating. The surface defect severity feature data and electrochemical defect severity feature data further quantify the severity of the defect, providing a comprehensive and quantitative basis for subsequent fusion weight determination and defect identification. This makes defect identification more accurate and comprehensive, and provides multi-dimensional scientific support for color-coated steel sheet quality assessment.

[0024] Determining fusion weight data of the surface defect characteristic data and the electrochemical defect characteristic data based on the surface defect degree characteristic data and the electrochemical defect degree characteristic data; Calculate the surface defect degree assessment coefficient based on the surface defect degree assessment model ; Obtain surface defect degree benchmark parameter data, including the benchmark value of the defect impact area ratio , defect shape complexity index benchmark value and defect depth influence coefficient benchmark value By setting the benchmark values ​​of defect impact area ratio, shape complexity, and depth impact coefficient, a unified quantitative standard is provided for defect degree assessment, ensuring that different defect data can be compared at the same scale to avoid assessment deviation due to dimensional differences. The surface defect degree characteristic data and surface defect degree benchmark parameter data are input into the surface defect degree assessment model to obtain the surface defect degree assessment coefficient. , the surface defect degree evaluation model is as follows: .

[0025] Quantify the degree of deviation of defects from the benchmark in dimensions such as area, shape, and depth, so that the abstract severity of defects can be converted into specific values, which facilitates the determination of subsequent fusion weights.

[0026] Calculation of electrochemical defect degree assessment coefficient based on electrochemical defect degree assessment model ; Obtain electrochemical defect degree benchmark parameter data, including corrosion rate change rate index benchmark value and the baseline value of the abnormality index of electrochemical parameters ; Clarify the baseline values ​​of the corrosion rate change rate and the abnormality of electrochemical parameters, establish a unified scale for electrochemical defect degree assessment, and make the electrochemical data of different batches and different conditions comparable. Input the electrochemical defect degree characteristic data and the electrochemical defect degree benchmark parameter data into the electrochemical defect degree assessment model to obtain the electrochemical defect degree assessment coefficient , the electrochemical defect degree evaluation model is as follows: .

[0027] It comprehensively reflects the changes in corrosion rate and abnormal electrochemical parameters caused by defects, quantifies the severity of defects at the electrochemical level in numerical form, and provides a quantitative basis in the electrochemical dimension for multimodal feature fusion.

[0028] Obtain the defect degree matching data stored in the database - fusion weight mapping set, and convert the surface defect degree evaluation coefficient and electrochemical defect degree evaluation coefficient Perform cosine similarity comparison with each defect degree matching data in the defect degree matching data-fusion weight mapping set to determine the cosine similarity value; determine the defect degree matching data corresponding to the largest cosine similarity value, and obtain the corresponding fusion weight data. This can find the historical situation closest to the current defect characteristics from the historical data, ensuring the rationality and pertinence of the weight determination.

[0029] The mapping set constructed using historical data provides an empirical reference for the association between the current defect degree assessment and the fusion weight, avoids the blindness of weight determination, and makes full use of the correspondence between the defect characteristics contained in the historical data and the optimal weight.

[0030] The method for obtaining the defect degree matching data-fusion weight mapping set is as follows: obtain the historical surface defect feature data and historical electrochemical defect feature data of the historical color-coated plate; process the historical surface defect feature data and historical electrochemical defect feature data based on the surface defect degree assessment model and the electrochemical defect degree assessment model respectively to obtain the historical surface defect degree assessment coefficient and the historical electrochemical defect degree assessment coefficient; convert the historical feature data into a quantitative assessment coefficient through the established model, unify the measurement standard of the historical defect degree, and make the defect data of different periods and types comparable.

[0031] The historical surface defect degree assessment coefficients and the historical electrochemical defect degree assessment coefficients are recorded as defect degree matching data; the historical assessment coefficients are integrated into matching data to form a standardized defect degree representation, which facilitates feature matching with current defect data. Several sets of fusion weights are randomly assigned to each defect degree matching data; by randomly generating multiple sets of weights, the possibility of different weight combinations is covered, providing a sufficient candidate set for subsequent screening of the optimal weight, avoiding the limitations of weight setting. The defect degree matching data is normalized and fused with the fusion weights to obtain the fused historical defect feature vector; the data is normalized to eliminate dimensional differences, and then fused with the weights to form a feature vector, so that the correlation between historical defect characteristics and weights can be quantitatively analyzed by the model.

[0032] Determine the defect severity matching data-fusion weight mapping set based on the fused historical defect feature vectors. By analyzing the performance of the fused feature vectors, the optimal weight corresponding to each defect severity matching data is selected, and a precise mapping relationship between defect severity and fusion weight is established, providing a reliable weight matching basis for real-time defect identification.

[0033] The defect degree matching data-fusion weight mapping set is determined based on the fused historical defect feature vectors, including the following steps: the fused historical defect feature vectors corresponding to the same fusion weight are respectively input into the constructed initial CNN prediction model and the initial LSTM prediction model for pre-training; by inputting the feature vectors corresponding to the same fusion weight into the initial CNN and LSTM model pre-training, the fusion effect of the multimodal features under the weight can be verified, ensuring the adaptability of the weight to the image and electrochemical characteristics.

[0034] After the pre-training is completed, performance data of the initial CNN prediction model and the initial LSTM prediction model after pre-training corresponding to the same fusion weight are counted, including CNN prediction model performance data and LSTM prediction model performance data; a performance performance coefficient is obtained based on the CNN prediction model performance data and the LSTM prediction model performance data; the fusion weight corresponding to the maximum performance performance coefficient is recorded as the optimal fusion weight; the defect degree matching data is one-to-one corresponding to the optimal fusion weight, and a defect degree matching data-fusion weight mapping set is obtained. The defect degree matching data is one-to-one corresponding to the optimal weight, forming a reusable mapping relationship, so that the subsequent real-time defect recognition can quickly obtain adaptive fusion weight based on historical optimal experience, and the recognition efficiency and accuracy are improved.

[0035] The CNN prediction model performance data includes CNN prediction accuracy , CNN prediction precision and CNN recall rate , and the LSTM prediction model performance data includes LSTM prediction accuracy , LSTM prediction precision and LSTM recall rate . The performance performance coefficient is obtained, including the following steps: ; wherein, and are weight factors.

[0036] The accuracy, precision, recall rate and other indicators of the CNN and LSTM models are collected to quantify the defect recognition ability of the models under different fusion weights from multiple dimensions, providing objective data support for weight screening. The performance data of the CNN and the LSTM are integrated by using the formula, the importance of different indicators is balanced by the weight factor, the multi-dimensional performance is converted into a single coefficient, and the advantages and disadvantages of different weights are compared intuitively.

[0037] The surface defect feature data and the electrochemical defect feature data are fused based on the fusion weight data to determine the fusion feature vector; the fusion feature vector is input into the CNN-LSTM hybrid prediction model trained by reinforcement learning to obtain the color coated board surface defect recognition result.

[0038] The CNN-LSTM hybrid prediction model is trained by reinforcement learning, including the following steps: building an initial CNN-LSTM hybrid prediction model, including a CNN layer, an LSTM layer and a fully connected layer, wherein the CNN layer is used to process the surface image data of the color-coated plate to obtain surface defect feature data, the LSTM layer is used to process the electrochemical feature data to obtain electrochemical defect feature data, and the fully connected layer is used to perform weighted fusion and splicing on the surface defect feature data and the electrochemical defect feature data to obtain a comprehensive vector, and obtain the surface defect recognition result of the color-coated plate through multiple neuron processing; the initial CNN-LSTM hybrid prediction model is trained and optimized based on the reinforcement learning algorithm to obtain a trained CNN-LSTM hybrid prediction model.

[0039] Leveraging the strengths of the CNN-LSTM hybrid model, the CNN processes image features to extract visual information, while the LSTM processes electrochemical sequence features to capture temporal dependencies. After fusion, the fully connected layer processes the neural network to output recognition results, enabling multi-dimensional analysis of defects. Using a reinforcement learning algorithm, model parameters are automatically adjusted based on predictive performance, enabling continuous model optimization during iterations, improving the recognition and generalization capabilities of surface defects on pre-coated steel sheets, and resulting in a more reliable trained model.

[0040] The initial CNN-LSTM hybrid prediction model is trained and optimized based on a reinforcement learning algorithm. The following steps are included: The current CNN layer parameters, LSTM layer parameters, and a comprehensive vector are defined as states, which comprehensively represent the current operating state of the model and provide accurate state input for reinforcement learning, enabling the algorithm to make decisions based on complete state information. A set of actions stored in a database is obtained, with each action corresponding to a set of parameter adjustment values, including CNN layer parameter adjustment values ​​and LSTM layer parameter feature values. The set of actions containing CNN and LSTM layer parameter adjustment values ​​is obtained from the database, providing a variety of adjustment options for model parameter optimization, covering different parameter adjustment possibilities and increasing the probability of finding the optimal parameters.

[0041] Determine the reward value corresponding to each action stored in the database, adjust the parameters of the current CNN layer and LSTM layer based on the action corresponding to the maximum reward value, and update the status; adjust the parameters according to the action corresponding to the maximum reward value, and use a quantitative reward mechanism to guide the direction of parameter optimization, ensuring that each adjustment is made towards the goal of improving model performance, thereby improving the efficiency and pertinence of parameter optimization.

[0042] After each parameter adjustment, the predicted performance evaluation coefficient is obtained. If the predicted performance evaluation coefficient is greater than the set performance evaluation threshold, the adjustment is stopped. Otherwise, the adjustment is continued and the reward value corresponding to the action is updated. After the parameter adjustment, the state is updated, allowing the reinforcement learning algorithm to make subsequent decisions based on the new model state, forming a closed-loop optimization process of state-action-state, driving continuous improvement in model performance.

[0043] Obtaining the prediction performance evaluation coefficient includes the following steps: Obtaining the prediction performance data of the initial CNN-LSTM hybrid prediction model after parameter adjustment, including the prediction accuracy , the ratio of the predicted time to the set benchmark time and load occupancy , perform weighted calculation on the prediction performance data to obtain the prediction performance evaluation coefficient : ;in, , and These are all weight factors. The evaluation coefficient is obtained by weighted calculation of prediction accuracy, ratio of prediction time to benchmark time, and load occupancy rate. The model performance is comprehensively measured by comprehensively considering the model's prediction accuracy, efficiency, and resource occupancy.

[0044] The process of updating the reward value corresponding to the action when the predicted performance evaluation coefficient is not greater than the set performance evaluation threshold is as follows: subtract the predicted performance evaluation coefficient from the performance evaluation threshold to obtain the difference; compare the difference with the corresponding matching difference stored in the database to determine the update value corresponding to the matching difference; add the reward value and the update value to obtain the updated reward value.

[0045] The evaluation coefficient is compared with the threshold to determine whether to stop adjustment. This ensures that optimization stops when model performance reaches the set standard, avoiding over-adjustment. Optimization continues to improve performance if the standard is not met. The action reward value is updated based on the difference between the evaluation coefficient and the threshold, allowing the reward mechanism to dynamically reflect the actual impact of the action on model performance, optimize the reward distribution across the action set, and provide more accurate reward guidance for subsequent parameter adjustments.

[0046] Color-coated sheet surface defect recognition system based on machine learning, such as Figure 2As shown, it includes a data acquisition module, which is used to perform surface image acquisition and electrochemical detection on the color-coated plate to obtain surface image data and electrochemical feature data of the color-coated plate; a feature extraction module, which is used to extract surface defect feature data and surface defect degree feature data from the surface image data of the color-coated plate; extract electrochemical defect feature data and electrochemical defect degree feature data from the electrochemical feature data; a fusion weight acquisition module, which is used to determine the fusion weight data of the surface defect feature data and the electrochemical defect feature data based on the surface defect degree feature data and the electrochemical defect degree feature data; a vector fusion module, which is used to fuse the surface defect feature data and the electrochemical defect feature data based on the fusion weight data to determine the fusion feature vector; and a defect recognition module, which is used to input the fusion feature vector into the CNN-LSTM hybrid prediction model trained by reinforcement learning to obtain the surface defect recognition result of the color-coated plate.

[0047] An electronic device includes: a processor; and a memory, wherein computer program instructions are stored in the memory, and when the computer program instructions are executed by the processor, the processor executes the above-mentioned method for identifying surface defects of color-coated plates based on machine learning.

[0048] A computer-readable storage medium is used to store a program, which, when executed by a processor, implements the method for identifying surface defects of color-coated plates based on machine learning as described above.

[0049] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0050] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0051] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0052] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0053] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0054] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for identifying surface defects of color-coated steel sheets based on machine learning, characterized in that: The following steps are involved: Perform surface image acquisition and electrochemical testing on the color-coated plate to obtain surface image data and electrochemical characteristic data of the color-coated plate; Extracting surface defect feature data and surface defect degree feature data from the surface image data of the color-coated plate; extracting electrochemical defect feature data and electrochemical defect degree feature data from the electrochemical feature data; Determining fusion weight data of the surface defect characteristic data and the electrochemical defect characteristic data based on the surface defect degree characteristic data and the electrochemical defect degree characteristic data; The surface defect feature data and the electrochemical defect feature data are fused based on the fusion weight data to determine a fusion feature vector; The fused feature vector is input into the CNN-LSTM hybrid prediction model trained by reinforcement learning to obtain the surface defect recognition results of the color-coated steel plate.

2. The method for identifying surface defects of color-coated steel sheets based on machine learning according to claim 1, characterized in that: Surface defect degree characteristic data including defect impact area ratio index , defect shape complexity index and defect depth influence coefficient ;Electrochemical defect degree characteristic data including corrosion rate change index and electrochemical parameter abnormality index ; Determining fusion weight data of surface defect feature data and electrochemical defect feature data includes the following steps: Calculate the surface defect degree assessment coefficient based on the surface defect degree assessment model ; Calculation of electrochemical defect degree evaluation coefficient based on electrochemical defect degree evaluation model ; Obtain the defect degree matching data stored in the database - fusion weight mapping set, and convert the surface defect degree evaluation coefficient and electrochemical defect degree evaluation coefficient Perform cosine similarity comparison with each defect degree matching data in the defect degree matching data-fusion weight mapping set one by one to determine the cosine similarity value; Determine the defect degree matching data corresponding to the maximum cosine similarity value and obtain the corresponding fusion weight data.

3. The method for identifying surface defects of color-coated steel sheets based on machine learning according to claim 2, characterized in that: Calculate the surface defect degree assessment coefficient based on the surface defect degree assessment model , including the following steps: Obtain benchmark data on the degree of surface defects, including the benchmark value of the defect impact area ratio , defect shape complexity index benchmark value and defect depth influence coefficient benchmark value ; Input the surface defect degree characteristic data and surface defect degree benchmark parameter data into the surface defect degree evaluation model to obtain the surface defect degree evaluation coefficient , the surface defect degree evaluation model is as follows: ； Calculation of electrochemical defect degree evaluation coefficient based on electrochemical defect degree evaluation model , including the following steps: Obtain electrochemical defect degree benchmark parameter data, including corrosion rate change rate index benchmark value and the baseline value of electrochemical parameter abnormality index ; The electrochemical defect degree characteristic data and the electrochemical defect degree benchmark parameter data are input into the electrochemical defect degree evaluation model to obtain the electrochemical defect degree evaluation coefficient , the electrochemical defect degree evaluation model is as follows: 。 4. The method for identifying surface defects of color-coated steel sheets based on machine learning according to claim 2, characterized in that: The method for obtaining the defect degree matching data-fusion weight mapping set is as follows: Obtain historical surface defect characteristic data and historical electrochemical defect characteristic data of color-coated plates; Based on the surface defect degree assessment model and the electrochemical defect degree assessment model, the historical surface defect characteristic data and the historical electrochemical defect characteristic data are processed respectively to obtain the historical surface defect degree assessment coefficient and the historical electrochemical defect degree assessment coefficient; Recording the historical surface defect degree evaluation coefficient and the historical electrochemical defect degree evaluation coefficient as defect degree matching data; Randomly assign several groups of fusion weights to each defect degree matching data; Normalize the defect degree matching data and fuse it with the fusion weight to obtain the fused historical defect feature vector; The defect degree matching data-fusion weight mapping set is determined based on the fused historical defect feature vector.

5. The method for identifying surface defects of color-coated steel sheets based on machine learning according to claim 4 is characterized in that: Determining a defect degree matching data-fusion weight mapping set based on the fused historical defect feature vector includes the following steps: The fused historical defect feature vectors corresponding to the same fusion weight are respectively input into the built initial CNN prediction model and initial LSTM prediction model for pre-training; After pre-training is completed, the performance data of the pre-trained initial CNN prediction model and the initial LSTM prediction model corresponding to the same fusion weight are counted, including the CNN prediction model performance data and the LSTM prediction model performance data; The performance coefficient is obtained based on the performance data of the CNN prediction model and the LSTM prediction model; The fusion weight corresponding to the maximum performance coefficient is recorded as the optimal fusion weight; The defect degree matching data are matched one-to-one with the optimal fusion weights to obtain a defect degree matching data-fusion weight mapping set.

6. The method for identifying surface defects of color-coated steel sheets based on machine learning according to claim 4, characterized in that: CNN prediction model performance data includes CNN prediction accuracy , CNN prediction accuracy and CNN recall , LSTM prediction model performance data includes LSTM prediction accuracy , LSTM prediction accuracy and LSTM recall ; Get the performance coefficient , including the following steps: ； in, and are all weight factors.

7. The method for identifying surface defects of color-coated steel sheets based on machine learning according to claim 1, characterized in that: Training the CNN-LSTM hybrid prediction model through reinforcement learning includes the following steps: Build an initial CNN-LSTM hybrid prediction model, including CNN layer, LSTM layer and fully connected layer. The CNN layer is used to process the surface image data of the color-coated plate to obtain surface defect feature data. The LSTM layer is used to process the electrochemical feature data to obtain electrochemical defect feature data. The fully connected layer is used to perform weighted fusion and splicing of the surface defect feature data and the electrochemical defect feature data to obtain a comprehensive vector. Through multiple neuron processing, the surface defect recognition result of the color-coated plate is obtained. The initial CNN-LSTM hybrid prediction model is trained and optimized based on the reinforcement learning algorithm to obtain a trained CNN-LSTM hybrid prediction model.

8. The method for identifying surface defects of color-coated steel sheets based on machine learning according to claim 7, characterized in that: The training and optimization of the initial CNN-LSTM hybrid prediction model based on the reinforcement learning algorithm includes the following steps: Define the parameters of the current CNN layer, the parameters of the LSTM layer, and the comprehensive vector as states; Obtain the action set stored in the database. Each action in the action set corresponds to a set of parameter adjustment values, including CNN layer parameter adjustment values ​​and LSTM layer parameter feature values. Determine the reward value corresponding to each action stored in the database, adjust the parameters of the current CNN layer and the LSTM layer based on the action corresponding to the maximum reward value, and update the state; After each parameter adjustment, the predicted performance evaluation coefficient is obtained. If the predicted performance evaluation coefficient is greater than the set performance evaluation threshold, the adjustment is stopped. Otherwise, the adjustment is made and the reward value corresponding to the action is updated at the same time.

9. The method for identifying surface defects of color-coated steel sheets based on machine learning according to claim 8, characterized in that: Obtaining the prediction performance evaluation coefficient includes the following steps: Obtain the prediction performance data of the initial CNN-LSTM hybrid prediction model after parameter adjustment, including the prediction accuracy , the ratio of the predicted time to the set benchmark time and load occupancy , perform weighted calculation on the prediction performance data to obtain the prediction performance evaluation coefficient : ； in, , and are weight factors; The process of updating the reward value corresponding to the action when the predicted performance evaluation coefficient is not greater than the set performance evaluation threshold is as follows: Subtract the predicted performance evaluation coefficient from the performance evaluation threshold to obtain the difference; The difference is compared with the corresponding matching difference stored in the database to determine the updated value corresponding to the matching difference, and the reward value is added to the updated value to obtain the updated reward value.

10. A system for identifying surface defects of color-coated steel sheets based on machine learning, used in the method for identifying surface defects of color-coated steel sheets based on machine learning according to any one of claims 1 to 9, characterized in that: include: The data acquisition module is used to collect surface images and perform electrochemical detection on the color-coated plate to obtain surface image data and electrochemical characteristic data of the color-coated plate; A feature extraction module is used to extract surface defect feature data and surface defect degree feature data from the surface image data of the color-coated plate; and to extract electrochemical defect feature data and electrochemical defect degree feature data from the electrochemical feature data; A fusion weight acquisition module, used to determine fusion weight data of the surface defect feature data and the electrochemical defect feature data based on the surface defect degree feature data and the electrochemical defect degree feature data; A vector fusion module is used to fuse the surface defect feature data and the electrochemical defect feature data based on the fusion weight data to determine the fusion feature vector; The defect recognition module is used to input the fused feature vector into the CNN-LSTM hybrid prediction model trained through reinforcement learning to obtain the surface defect recognition results of the color-coated plate.

Images