Intelligent diagnosis method for crimping quality of strain clamp based on infrared thermal imaging

Abstract

The invention provides a strain clamp crimping quality intelligent diagnosis method based on infrared thermal imaging, and the method comprises the steps: carrying out the noise reduction and enhancement processing of a thermal image data set through employing a wavelet transform and morphological filtering algorithm, highlighting the temperature distribution characteristics of a crimping part, segmenting a thermal image through employing a region growing algorithm, extracting a region of interest, and carrying out the segmentation of a region of interest. Calculating characteristic parameters such as a temperature mean value, a variance and a gradient of the region, and constructing a crimping quality characteristic vector; in order to solve the problem of crimp quality defect sample scarcity, a generative adversarial network algorithm is adopted, through adversarial learning of a generator and a discriminator, defect samples are synthesized, a data set is expanded, a transfer learning technology is utilized, a convolutional neural network model is pre-trained on a crimp quality normal sample, general feature representation is extracted, and fine tuning is performed on defect data. And the model generalization ability is improved.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to an intelligent diagnosis method for the crimping quality of a tension clamp based on infrared thermal imaging. Background Art

[0002] The tension clamp in the overhead transmission line is an important hardware connection device, and its crimping quality directly affects the safe operation of the line. At present, the quality detection of tension clamp crimping mainly relies on manual infrared temperature measurement, which has problems such as low detection efficiency and strong subjectivity. In order to realize the intelligent diagnosis of the crimping quality of tension clamps, the following technical difficulties need to be solved: standardized acquisition and preprocessing of infrared thermal imaging data; crimping quality feature extraction and quantitative analysis; machine learning model training and optimization under small sample data; interference elimination of complex environmental factors and improvement of model generalization ability. The acquisition of infrared thermal imaging data needs to consider the influence of factors such as shooting angle, distance, and ambient temperature. Minor deviations in the acquisition process may cause data distortion. At the same time, there is a lot of redundant and noisy information in the original thermal image, and preprocessing operations such as image enhancement, segmentation, and feature extraction are required to improve data quality. In terms of crimping quality feature extraction and quantitative analysis, it is necessary to study the intrinsic relationship between temperature distribution and crimping quality, and extract key characteristic parameters that can reflect crimping quality, such as temperature gradient and temperature dispersion. Since it is difficult to obtain samples of crimping quality defects, the problem of data imbalance is prominent. It is necessary to use technologies such as data enhancement and transfer learning to improve the training effect of the model under small sample data. In addition, the environment in which overhead lines are located is complex and changeable. Factors such as wind and bird droppings may interfere with infrared imaging. It is necessary to study the environmental adaptability of the model and improve the robustness and generalization performance of the diagnostic model. Summary of the invention

[0003] The present invention provides an intelligent diagnosis method for the crimping quality of a tension clamp based on infrared thermal imaging, which mainly includes:

[0004] According to the crimping process specification of tension clamps, the infrared thermal imaging data acquisition standard is formulated, which stipulates that the shooting angle is 45°±5°, the distance is 0.5m-1m, and the ambient temperature range is 10℃-30℃. The infrared thermal imager is used to obtain the thermal image data of the crimping part of the tension clamp, and the image registration and standardization are performed to obtain a thermal image data set with a uniform size of 1024×768 pixels and a JPEG format;

[0005] Wavelet transform and morphological filtering algorithm are used to perform noise reduction and enhancement on the thermal image data set to highlight the temperature distribution characteristics of the crimping part. The thermal image is segmented using the region growing algorithm to extract the region of interest, and the characteristic parameters such as the temperature mean, variance, and gradient of the region are calculated to construct the crimping quality feature vector.

[0006] According to the crimping process parameters (such as crimping force and crimping time) and mechanical simulation analysis results (such as stress distribution and deformation), a mapping model between temperature distribution characteristics and crimping quality is established, and the crimping quality evaluation index is quantified. The correlation analysis and principal component analysis methods are used to screen the key temperature characteristics that have a significant impact on the crimping quality, reduce the feature dimension, and obtain the optimized crimping quality feature set.

[0007] In order to solve the problem of scarcity of samples with crimping quality defects, a generative adversarial network algorithm is used to synthesize defective samples and expand the data set through adversarial learning between the generator and the discriminator. The convolutional neural network model is pre-trained on samples with normal crimping quality using transfer learning technology to extract common feature representations, and fine-tune the defective data to improve the generalization ability of the model.

[0008] Data augmentation technology is used to simulate thermal images in various environments through operations such as image rotation (±15°), translation (±10 pixels), scaling (0.9-1.1 times), and noise addition (Gaussian noise, σ=0.01), to enhance the model's environmental adaptability, and to add L2 regularization terms in model training to control model complexity and improve model robustness and anti-interference ability.

[0009] In the construction of the crimping quality diagnosis model, machine learning algorithms such as support vector machine, random forest and convolutional neural network are used. Through grid search and cross-validation, model hyperparameters are optimized to improve the diagnosis accuracy. The spatial attention mechanism and gradient weighted class activation mapping (Grad-CAM) interpretability module are introduced to enhance the model interpretability and output the diagnosis results such as the location, type and severity of crimping defects.

[0010] Based on the above method, an online real-time diagnosis system for the crimping quality of tension wire clamps was developed. The infrared thermal imager collects thermal images of the crimping parts in real time, transmits them to the edge computing device for preprocessing and feature extraction, and uploads them to the cloud server for quality diagnosis and evaluation. Through model verification and optimization, including using independent test sets to evaluate model performance and regularly updating model parameters, intelligent monitoring of the entire crimping quality process is achieved, crimping defects are discovered and warned in time, and safe operation of transmission lines is guaranteed.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses an intelligent diagnosis method for the crimping quality of a tension wire clamp based on infrared thermal imaging. The method collects thermal images of the crimping part by an infrared thermal imager, performs standardization processing and noise reduction enhancement, and extracts temperature distribution features. In view of the scarcity of samples with crimping quality defects, a generative adversarial network is used to synthesize defect samples, and transfer learning and data augmentation techniques are used to improve the generalization ability of the model. In the construction of the diagnostic model, a spatial attention mechanism and a gradient-weighted class activation mapping are introduced to enhance interpretability. The present invention realizes intelligent monitoring of the entire crimping quality process, and can timely detect and warn crimping defects. By establishing a mapping model between temperature distribution features and crimping quality, the evaluation index is quantified, and the diagnostic accuracy is improved. Online real-time diagnosis is achieved by combining edge computing with cloud servers. The present invention solves the problems of low efficiency and poor accuracy of traditional manual detection, significantly improves the automation level and reliability of crimping quality detection of tension wire clamps, and is of great significance to ensuring the safe operation of transmission lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flow chart of an intelligent method for diagnosing the crimping quality of a tension clamp based on infrared thermal imaging.

[0014] Figure 2 It is a schematic diagram of an intelligent diagnosis method for crimping quality of tension clamps based on infrared thermal imaging of the present invention.

[0015] Figure 3 It is another schematic diagram of the intelligent diagnosis method of the crimping quality of tension clamps based on infrared thermal imaging of the present invention. DETAILED DESCRIPTION

[0016] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] like Figure 1-3 In this embodiment, a method for intelligent diagnosis of crimping quality of a tension clamp based on infrared thermal imaging may specifically include:

[0018] Step S101, according to the tension clamp crimping process specification, formulate the infrared thermal imaging data acquisition standard, specifying the shooting angle of 45°±5°, the distance of 0.5m-1m, and the ambient temperature range of 10℃-30℃. Use an infrared thermal imager to obtain the thermal image data of the tension clamp crimping part, perform image registration and standardization processing, and obtain a thermal image data set with a uniform size of 1024×768 pixels and a JPEG format.

[0019] A thermal image data set of the crimping part of the tension clamp is obtained, wherein the thermal image data set includes multiple thermal image data, each thermal image data has corresponding shooting angle, shooting distance and ambient temperature information; an image registration algorithm is used to geometrically correct the thermal image data set to obtain a geometrically corrected thermal image data set; an image standardization algorithm is used to unify the size and format of the geometrically corrected thermal image data set to obtain a standardized thermal image data set, wherein the standardized thermal image data is in 1024×768 pixel and JPEG format; temperature compensation is performed on the standardized thermal image data set according to ambient temperature information to eliminate the influence of ambient temperature changes on the thermal image data to obtain a temperature compensated thermal image data set; an image segmentation algorithm is used to divide the temperature compensated thermal image data set into regions to extract the crimping part of the tension clamp. Thermal image area, obtain the data set of the area of ​​interest; obtain the multi-scenario adaptability knowledge base of the drone electrical testing equipment, determine the voltage level and line type information of the current electrical testing scene according to the knowledge base, associate the voltage level and line type information with the data set of the area of ​​interest, and construct a thermal image data set with scene labels; use a feature extraction algorithm to analyze the thermal image data set with scene labels, extract the temperature distribution uniformity characteristics and hot spot area characteristics that reflect the crimping quality of the tension clamp, and obtain a thermal image feature set; input the thermal image feature set into a pre-constructed crimping quality assessment model based on deep learning, and judge whether the current crimping quality of the tension clamp is qualified by comparing it with the historical crimping cases in the crimping quality assessment model. If qualified, generate a crimping quality qualified report; if unqualified, generate a crimping quality unqualified report.

[0020] For example, obtaining thermal image data of the crimping part of the tension clamp is an important means to evaluate the safety of the transmission line. First, it is necessary to collect thermal images at multiple angles and distances, and record information such as ambient temperature. For example, an infrared thermal imager can be used to shoot the tension clamp at different distances such as 10 meters, 20 meters, and 30 meters, from different angles such as 0°, 45°, and 90°, while recording the ambient temperature at that time as 25°C. Image registration is the process of aligning thermal images taken at different angles to the same coordinate system. A feature point-based registration algorithm, such as the SI FT algorithm, can be used. The algorithm can extract key points in the thermal image and match them, thereby achieving accurate alignment of thermal images at different angles. This step can eliminate the geometric deformation caused by different shooting angles and lay the foundation for subsequent analysis. Image standardization is to unify thermal images to the same size and format. For example, a bilinear interpolation algorithm can be used to adjust all thermal images to a resolution of 1024×768 pixels and save them in JPEG format. The purpose of this is to eliminate the differences in image size and format under different devices and different shooting conditions, so as to facilitate subsequent processing. Temperature compensation is to eliminate the influence of ambient temperature on thermal images. The thermal image can be corrected using a linear compensation model based on the recorded ambient temperature. For example, if the standard ambient temperature is 20°C and the ambient temperature during actual shooting is 25°C, the temperature value in the thermal image needs to be reduced by 5°C as a whole. This can obtain more accurate temperature distribution information. Image segmentation is to extract the area of ​​the tension clamp crimping part from the entire thermal image. A threshold-based segmentation algorithm, such as the Otsu algorithm, can be used. The algorithm can automatically calculate the optimal threshold and divide the thermal image into high temperature and low temperature areas, thereby accurately extracting the thermal image area of ​​the crimping part. The multi-scenario adaptive knowledge base of the drone electrical inspection equipment contains feature information of different voltage levels and line types. For example, for a 220kV overhead transmission line, the knowledge base will contain information such as the normal operating temperature range and the maximum allowable temperature at this voltage level. Associating this information with the extracted thermal image area can provide an important reference for subsequent analysis. Feature extraction is to extract key features reflecting the crimping quality from thermal images. The temperature distribution uniformity feature can be represented by calculating the temperature standard deviation of the thermal image area. The smaller the standard deviation, the more uniform the temperature distribution. The hot spot area feature can be represented by counting the number of pixels exceeding a certain temperature threshold. These features can intuitively reflect the temperature anomaly of the crimping part. The crimping quality assessment model based on deep learning is obtained through training with a large number of historical cases. A convolutional neural network (CNN) can be used to build a model, with the input being the extracted thermal image features and the output being the crimping quality assessment result. By comparing with historical cases, the model can accurately determine whether the current crimping quality is qualified. For example, in a certain test, the extracted temperature distribution uniformity feature value was 2.5°C (temperature standard deviation), and the hot spot area feature was 50 square centimeters.After inputting these features into the evaluation model, the model outputs a crimping quality score of 85 points (out of 100 points). According to the preset qualification standard (80 points), the system determines that the crimping quality is qualified and generates a qualified report containing detailed test data and evaluation results. This crimping quality evaluation method based on thermal image analysis combines a variety of advanced technologies such as image processing and machine learning. It can quickly and accurately evaluate the crimping quality of tension clamps and effectively improve the safety and reliability of transmission lines. By continuously accumulating data and optimizing models, the evaluation accuracy of this method will continue to improve, providing strong support for the intelligent operation and maintenance of power systems.

[0021] Step S102: Use wavelet transform and morphological filtering algorithm to perform noise reduction and enhancement on the thermal image data set to highlight the temperature distribution characteristics of the crimping part. Use the region growing algorithm to segment the thermal image, extract the region of interest, calculate the temperature mean, variance, gradient and other characteristic parameters of the region, and construct the crimping quality feature vector.

[0022] The thermal image data collected by the UAV electrical testing equipment is obtained, the thermal image data is preprocessed, and the thermal image data is subjected to noise reduction and enhancement processing by using wavelet transform and morphological filtering algorithms to highlight the temperature distribution characteristics of the crimping parts; based on the preprocessed thermal image data, the thermal image is segmented by using a regional growing algorithm to extract the region of interest and obtain the regional image of the crimping part; the temperature mean, variance, and gradient feature parameters are calculated from the regional image of the crimping part to construct a crimping quality feature vector; the crimping quality feature vector is input into a pre-trained crimping quality assessment model, and the crimping quality assessment result is obtained through model reasoning; if the crimping quality assessment result is lower than a preset threshold, the crimping quality is judged to be unqualified, a crimping fault alarm is triggered, and the fault information is sent to a remote control center; if If the crimping quality assessment result is higher than the preset threshold, the crimping quality is judged to be qualified, the crimping quality assessment result is stored in the knowledge base, and the crimping quality data is updated; based on the historical crimping quality data in the knowledge base, an incremental learning algorithm is used to update the parameters of the crimping quality assessment model in real time to improve the generalization ability and adaptability of the model; the environmental perception data of the drone electrical testing equipment is obtained, and the environmental perception data includes temperature, humidity, and wind speed parameters. The environmental perception data is correlated with the crimping quality assessment result to determine the influence of environmental factors on the crimping quality; based on the influence rules obtained from the correlation analysis, the working parameters of the drone electrical testing equipment are dynamically adjusted, the temperature measurement range and sensitivity of the thermal imager are adjusted, the crimping quality assessment model is optimized, and the accuracy and reliability of crimping quality detection are improved.

[0023] Exemplarily, after obtaining the thermal image data collected by the UAV electrical inspection equipment, preprocessing is first performed. This step includes using wavelet transform and morphological filtering algorithms to reduce noise and enhance the thermal image. Wavelet transform can effectively separate the signal and noise in the image, while morphological filtering can retain edge information and smooth the area at the same time. For example, for a thermal image with a resolution of 640×480, the db4 wavelet can be first applied for 3-layer decomposition, and then the high-frequency subband can be soft-thresholded, and finally reconstructed. Then, a circular structure element with a radius of 3 pixels is used for opening and closing operations to remove small noise points and smooth the edges. After preprocessing, the thermal image is segmented using a regional growing algorithm to extract the region of interest. The regional growing algorithm starts from a pre-selected seed point and gradually merges adjacent similar pixels into the region. In the crimping part detection, the highest temperature point can be selected as the seed point, and the growth threshold is set to 1.2 times the average temperature. In this way, the thermal abnormal area of ​​the crimping part can be effectively extracted. The characteristic parameters are calculated from the extracted regional image to construct the crimping quality feature vector. The mean temperature reflects the overall thermal condition, the variance indicates the degree of discreteness of the temperature distribution, and the gradient feature describes the severity of the temperature change. For example, for a normal crimping part, the mean temperature may be around 40°C, the variance does not exceed 5°C, and the maximum gradient value is less than 10°C / pixel. The feature vector is input into the pre-trained crimping quality assessment model for inference. The model can be a classifier built based on algorithms such as support vector machine (SVM) or random forest. The model outputs a score between 0 and 1, indicating the quality of the crimping. The threshold is set to 0.8. If it is lower than this value, it is judged as unqualified and a fault alarm is triggered; if it is higher, it is judged as qualified and the result is stored in the knowledge base. The historical data accumulated in the knowledge base is used for incremental learning of the model. An online stochastic gradient descent algorithm can be used to update the model parameters every time a new sample is received, so that the model can adapt to changes in data distribution. This method is particularly suitable for processing continuously generated crimping quality data streams. The correlation analysis of environmental perception data and crimping quality assessment results helps to understand the impact of external factors. For example, it may be found that when the ambient temperature exceeds 35°C, the accuracy of crimping quality assessment decreases by 5%. Based on this finding, the upper limit of the temperature measurement range of the thermal imager can be appropriately increased in hot weather, or the judgment threshold of the model can be adjusted to maintain the reliability of detection. This dynamic adjustment mechanism can significantly improve the adaptability and accuracy of the UAV electrical inspection system in complex environments.

[0024] Step S103, based on the crimping process parameters (such as crimping force, crimping time) and mechanical simulation analysis results (such as stress distribution, deformation), a mapping model between temperature distribution characteristics and crimping quality is established to quantify crimping quality evaluation indicators. Correlation analysis and principal component analysis methods are used to screen key temperature characteristics that have a significant impact on crimping quality, reduce feature dimensions, and obtain an optimized crimping quality feature set.

[0025] The crimping process parameters are obtained, and the crimping process parameters include crimping force and crimping time; the force and time data in the crimping process are collected in real time by sensors, and the collected data are transmitted to a data processing module; a finite element analysis method is used to perform mechanical simulation analysis on the crimping process according to the crimping process parameters, and stress distribution and deformation data of the crimping area are obtained; the temperature distribution data in the crimping process is collected in real time by infrared thermal imaging technology, and the temperature field distribution characteristics of the crimping area are obtained; the crimping process parameters, the stress distribution and deformation data obtained by the mechanical simulation analysis, and the temperature field distribution characteristics are integrated to construct a multi-source heterogeneous data set; a data mining algorithm is used to extract and select features of the multi-source heterogeneous data set, and the correlation analysis and principal component analysis methods are used to screen out the key factors that have a significant impact on the crimping quality. The key features include temperature distribution features and mechanical features, forming an optimized crimping quality feature set; based on the optimized crimping quality feature set, a mapping model between temperature distribution features and crimping quality is established by using a machine learning algorithm to achieve quantitative evaluation of crimping quality; the crimping quality evaluation model is deployed in an embedded system to achieve real-time monitoring and quality judgment of the crimping process; if the evaluation model determines that the crimping quality is unqualified, an early warning signal is triggered; an incremental learning algorithm is used to update and optimize the crimping quality evaluation model online according to the newly collected crimping process parameters and crimping quality data, so as to improve the generalization ability and robustness of the model; the crimping quality evaluation results are correlated with the crimping process parameters and the mechanical simulation analysis results, and a crimping quality report is generated through data visualization technology.

[0026] For example, crimping quality assessment is a key link in the maintenance of power equipment, involving the integration of multiple technical fields. First, obtain crimping process parameters such as crimping force and time, which directly affect the crimping effect. For example, for the crimping of copper wires, a crimping force of 2000N and a crimping time of 5 seconds may be required. Real-time collection of these data can ensure the accurate execution of process parameters. Finite element analysis method is used to simulate the mechanical behavior during crimping. By establishing a geometric model of the crimping component, setting material properties and boundary conditions, the stress distribution and deformation of the crimping area can be calculated. For example, the analysis results may show that the maximum stress in the crimping area is 300MPa and the maximum deformation is 0.2mm. These data help to evaluate the strength and stability of the crimping. Infrared thermal imaging technology can capture temperature changes during crimping. During the crimping process, the temperature of the crimping area will increase due to material deformation and friction. By analyzing the temperature field distribution, potential hot spots or abnormal temperature areas can be identified. For example, normal crimping may cause the temperature of the crimping area to increase by 10-15℃, while in abnormal cases, the local temperature may rise by more than 30℃. The fusion of multi-source heterogeneous data is the key to achieving a comprehensive evaluation. Combining process parameters, mechanical analysis results, and temperature distribution characteristics can build a more comprehensive crimp quality evaluation model. Through data mining and feature selection, the most influential features can be identified. For example, it may be found that the crimping force is positively correlated with the final crimp strength, while the maximum temperature in the crimping area is negatively correlated with the crimp quality. Machine learning algorithms can build crimp quality evaluation models based on these features. For example, a support vector machine (SVM) or random forest algorithm can be used, with inputs including crimping force, time, maximum stress, maximum deformation, and maximum temperature, and outputs a quantitative score of crimp quality. Model training may be based on a large amount of historical data, including crimp samples with known good and bad quality. Deploying the evaluation model into an embedded system can achieve real-time monitoring. For example, using an ARM-based microprocessor with a lightweight machine learning framework, the quality score can be calculated in real time during the crimping process. If the score is lower than a preset threshold (such as 80 points), the system will trigger an early warning signal to prompt the operator to check or rework. Incremental learning algorithms allow the model to be continuously optimized with new data. After each crimping is completed, new data samples can be used to update the model, gradually improving the accuracy and adaptability of the model. This is particularly effective for crimping tasks under different materials and environmental conditions. Finally, the crimping quality report is generated through data visualization technology, which can intuitively display the various indicators of the crimping process and the final quality assessment results. For example, a thermal map is used to display the temperature distribution in the crimping area, a line graph is used to show the change of crimping force over time, and a dashboard is used to display the final quality score. This visual report helps technicians quickly understand the crimping status and provides a basis for subsequent process optimization.

[0027] Step S104, in order to solve the problem of scarcity of samples with crimping quality defects, a generative adversarial network algorithm is used to synthesize defective samples and expand the data set through adversarial learning between the generator and the discriminator. The convolutional neural network model is pre-trained on samples with normal crimping quality by using transfer learning technology, and the general feature representation is extracted. Fine-tuning is performed on the defective data to improve the generalization ability of the model.

[0028] Acquire samples with normal crimping quality, adopt convolutional neural network model for pre-training, and extract the common feature representation of the sample data through network structure and loss function; acquire samples with defective crimping quality, adopt generative adversarial network algorithm, construct generator and discriminator network, take the defective samples as input, and train the generator to synthesize realistic defective samples through adversarial learning; acquire newly collected crimping quality data, adopt incremental learning algorithm, and regularly update the parameters of the defect detection model to adapt it to changes in the production environment and improve the robustness and real-time performance of the model.

[0029] For example, obtaining samples with normal crimping quality is the basis for building a defect detection model. Taking automotive wiring harness crimping as an example, a large number of qualified crimping terminal images can be collected, including samples under different models, materials and crimping parameters. Convolutional neural networks are used for pre-training, and classic network structures such as ResNet or VGGNet can be selected to extract the shape, texture and other features of the crimped terminals through multi-layer convolution and pooling. The loss function can use cross entropy loss to optimize network parameters to minimize prediction errors. The purpose of this step is to allow the network to learn the general representation of samples with normal crimping quality, laying the foundation for subsequent defect detection. Obtaining samples with crimping quality defects is the key to improving the detection capability of the model. Since there are few defective samples in actual production, a generative adversarial network (GAN) can be used to synthesize realistic defective samples. Taking insufficient crimping as an example, the generator network can receive normal crimping images as input and generate simulated insufficient crimping images through convolution and deconvolution operations. The discriminator network is responsible for distinguishing between real defective samples and generated samples. Through adversarial learning, the generator gradually improves the quality of the synthesized image, and finally produces samples that are highly similar to real defects. This method can significantly increase the number and diversity of defect samples and improve the generalization ability of the model. Acquiring newly collected crimping quality data and performing incremental learning can ensure that the model adapts to changes in the production environment. For example, when new crimping equipment or materials are introduced, new defect types may appear. By regularly collecting new crimping data and updating model parameters using incremental learning algorithms, the model can gradually adapt to these changes. In specific implementation, technologies such as Elastic Weight Consolidation can be used to retain original knowledge while learning new knowledge to avoid catastrophic forgetting. This method can improve the robustness of the model and enable it to maintain stable detection performance under different production conditions. The whole process forms a closed loop: starting from normal sample pre-training, expanding defective samples through GAN, and then continuously optimizing through incremental learning. This method can not only effectively deal with the problem of sample imbalance, but also adapt to the dynamic changes of the production environment, and ultimately achieve high-precision and high-reliability online detection of crimping quality.

[0030] Step S105, using data augmentation technology, through image rotation (±15°), translation (±10 pixels), scaling (0.9-1.1 times), noise addition (Gaussian noise, σ=0.01) and other operations, simulate thermal images in various environments to enhance the model's environmental adaptability. Add L2 regularization terms in model training to control model complexity and improve model robustness and anti-interference ability.

[0031] The thermal image data collected by the UAV electrical testing equipment is obtained, and the thermal image data is preprocessed to obtain the preprocessed thermal image data, wherein the preprocessing includes image denoising, image enhancement and image standardization; the preprocessed thermal image data is augmented by using data augmentation technology to obtain an augmented thermal image data set, wherein the data augmentation technology includes image rotation, image translation, image scaling and Gaussian noise addition; the augmented thermal image data set is divided into a training set, a validation set and a test set; a thermal image feature extraction model is constructed, and a convolutional neural network is used to extract multi-scale and multi-level features of the thermal images in the training set, and a comprehensive feature representation of each thermal image in the training set is obtained by feature fusion; in On the basis of the thermal image feature extraction model, a fault diagnosis model is constructed, and the support vector machine algorithm is used to classify the comprehensive feature representation to determine whether the equipment has a fault; if the fault diagnosis model determines that the equipment has a fault, the fault location model is triggered, and the decision tree algorithm is used to perform fine-grained classification of the fault type to determine the specific fault location and cause; according to the fault location result, the corresponding fault handling solution is obtained from the pre-constructed knowledge base to generate fault maintenance suggestions; using an online learning strategy, according to the newly collected thermal image data of the drone, the thermal image feature extraction model, the fault diagnosis model and the fault location model are regularly incrementally trained and optimized, and the model parameters are updated in real time.

[0032] For example, thermal image data preprocessing is a key step to improve the accuracy of fault diagnosis. Image denoising can be done by using methods such as median filtering or wavelet transform to effectively remove thermal noise and salt and pepper noise. Image enhancement can improve contrast through histogram equalization to make fault features more prominent. Image standardization can normalize pixel values ​​to the [0,1] interval, which is beneficial for subsequent model training. Data augmentation technology can effectively expand training samples and improve the generalization ability of the model. For example, the thermal image is rotated by ±15°, translated within a range of 10% up, down, left, and right, zoomed in and out by 20%, and Gaussian noise with a mean of 0 and a standard deviation of 0.01 is added. These operations can simulate various changes in the actual acquisition process and enhance the robustness of the model. The thermal image feature extraction model adopts a convolutional neural network structure, such as ResNet50. The shallow convolution layer extracts low-level features such as edges and textures, and the deep convolution layer extracts semantic-level high-level features. Multi-scale feature fusion is achieved through the feature pyramid network (FPN) to obtain a more comprehensive feature representation. The fault diagnosis model uses the support vector machine (SVM) algorithm, which can effectively process high-dimensional feature space. The radial basis kernel function is used, the penalty parameter C is set to 10, the kernel parameter γ is set to 0.1, and the hyperparameters are optimized by grid search. SVM can find the optimal classification hyperplane and achieve accurate classification of faults and normal states. The fault location model uses a decision tree algorithm, such as C4.5. The decision tree selects the optimal feature for splitting through the information gain ratio to form a series of if-then rules. For example, if the temperature in a certain area exceeds 80℃ and the temperature gradient is greater than 10℃ / cm, it is determined to be an insulation fault. The decision tree has the advantage of strong interpretability, which makes it easier for engineers to understand the cause of the fault. The knowledge base stores solutions for various faults. For example, for insulation faults, it is recommended to check whether the insulation layer is aged or damaged, and give specific steps for replacing the insulation material or local repair. The system automatically matches the corresponding solution based on the fault location results to provide guidance for maintenance personnel. The online learning strategy uses incremental learning methods, such as the Learn++ algorithm. When the newly collected thermal image data accumulates to a certain amount, the system automatically triggers the model update. The new data is combined with the original data set to retrain the model and update the parameters. This approach not only retains the original knowledge, but also adapts to the dynamic changes of the device status and environment, and continuously improves the accuracy of diagnosis.

[0033] Step S106, in the construction of the crimping quality diagnosis model, machine learning algorithms such as support vector machine, random forest and convolutional neural network are used to optimize model hyperparameters and improve diagnostic accuracy through grid search and cross-validation. The spatial attention mechanism and gradient weighted class activation mapping (Grad-CAM) interpretability module are introduced to enhance the model interpretability and output diagnostic results such as the location, type and severity of crimping defects.

[0034] The real-time status data of the UAV is obtained, and the status data is preprocessed by a low-power embedded processor to remove abnormal values ​​and obtain normalized status data; the normalized status data is modeled by a support vector machine algorithm to train an equipment fault diagnosis model; if the fault diagnosis model determines that a fault occurs, a fault alarm is triggered, and the alarm information is sent to a remote control center through the UAV communication module.

[0035] Exemplarily, the real-time status data of the drone is obtained, such as flight altitude, speed, attitude angle, battery voltage, motor speed, control signal strength, etc. These data reflect the operating status of the drone during the mission and are an important basis for fault diagnosis. These raw status data are preprocessed by a low-power embedded processor, such as an ARM architecture processor. The main purpose of preprocessing is to remove outliers, such as erroneous data caused by instantaneous sensor failure or electromagnetic interference. The removal of outliers can be based on statistical methods, for example, setting a reasonable numerical range, and data beyond this range is considered to be outliers, or calculating the mean and standard deviation of the data, and treating data that differs from the mean by more than three times the standard deviation as outliers and removing them. After preprocessing, normalized status data is obtained. For example, the original motor speed data may vary widely. Through normalization processing, it is mapped to between 0 and 1, which can eliminate the influence of dimensions between different data and improve the efficiency and accuracy of subsequent model training. The normalized status data is modeled using a support vector machine algorithm. Support vector machine is a powerful machine learning algorithm that is particularly suitable for processing high-dimensional data and complex nonlinear relationships. For example, the flight altitude, speed, attitude angle and other status data of the drone can be used as input features, and whether the equipment is normal can be used as the output label. The support vector machine model can be trained through a large amount of historical data. During the training process, the support vector machine algorithm will find an optimal hyperplane to separate samples of different categories and maximize the classification interval. This hyperplane is the obtained equipment fault diagnosis model. For example, if the status data of the drone is input into the model at a certain moment, and the result of the model output shows that the data point is on the "fault" side of the hyperplane, then it is determined that the equipment may have a fault. If the fault diagnosis model determines that a fault has occurred, a fault alarm is triggered. For example, the model outputs a signal indicating a fault, triggering the indicator light on the drone to flash, and sends the alarm information to the remote control center through the drone communication module, such as a 4G or 5G module. The alarm information may include the time when the fault occurred, the location of the drone, the type of fault (such as motor failure, battery failure, control system failure, etc.). In this way, the operator of the remote control center can understand the status of the drone in a timely manner and take corresponding measures, such as ordering the drone to return or make an emergency landing to avoid more serious accidents. In this way, the safety of drone inspections can be effectively improved. Because drones may encounter various emergencies when performing tasks, such as motor failure, low battery, control system failure, etc. By monitoring the status data of drones in real time and analyzing them using trained fault diagnosis models, potential faults can be discovered in time and measures can be taken before the fault occurs to avoid accidents. This can greatly improve the safety and reliability of drone inspections.

[0036] Step S107, based on the above method, develop an online real-time diagnosis system for the crimping quality of tension wire clamps. The infrared thermal imager collects thermal images of the crimping parts in real time, transmits them to the edge computing device for preprocessing and feature extraction, and uploads them to the cloud server for quality diagnosis and evaluation. Through model verification and optimization, including using independent test sets to evaluate model performance and regularly updating model parameters, the whole process of intelligent monitoring of crimping quality can be achieved, crimping defects can be discovered and warned in time, and the safe operation of transmission lines can be guaranteed.

[0037] Acquire a real-time infrared thermal image of the crimping part of the tension wire clamp collected by an infrared thermal imager; preprocess the infrared thermal image through an edge computing node to obtain a clear infrared thermal image; perform feature extraction based on the clear infrared thermal image, and if there is uneven temperature distribution in the infrared thermal image, extract temperature feature parameters to obtain a feature data set; use a feature selection algorithm to filter out a feature subset from the feature data set, and if the correlation between the feature in the feature subset and the crimping quality is lower than a preset threshold, remove the feature to obtain an optimized feature set; transmit the optimized feature set to a cloud-based distributed big data processing platform, and if the cloud server receives the data, perform quality diagnosis and evaluation to obtain a preliminary diagnosis result; based on the preliminary diagnosis result, use a distributed computing The framework performs parallel calculations on the historical crimping quality data stored in the data warehouse, and extracts feature information from the historical data through a data mining algorithm to obtain a historical feature database; the machine learning model is trained through the historical feature database, and if the crimping quality grade predicted by the model does not match the actual grade, the model parameters are adjusted to obtain an optimized crimping quality diagnosis model; an incremental learning algorithm is used to update the parameters of the crimping quality diagnosis model in real time according to the newly collected thermal image data and the corresponding actual crimping quality labels to obtain an updated model; the real-time thermal image feature data uploaded by the edge computing node is inferred through the updated model, and if the inference result shows that the crimping quality is unqualified, an alarm message and maintenance suggestions are generated to determine the final crimping quality diagnosis result.

[0038] Exemplarily, a real-time infrared thermal image of the crimping part of the tension clamp is obtained. For example, the infrared radiation information of the clamp area is captured at a rate of 30 frames per second by a FLIRVueProR infrared thermal imager mounted on a drone, and an original thermal image data stream is generated. Since the drone inspection process may be affected by factors such as airflow and jitter, the thermal image may have problems such as noise and blur, so preprocessing is required. The edge computing node performs preprocessing. For example, a Raspberry Pi 4B is used as an edge computing node, and its quad-core ARMCortex-A72 processor and OpenCV library are used to perform median filtering on the original thermal image to remove salt and pepper noise, and then Gaussian filtering is performed to smooth the image. Finally, histogram equalization is performed to enhance the image contrast to obtain a clear infrared thermal image. This preprocessing method can effectively improve the image quality and lay the foundation for subsequent feature extraction. Feature extraction is performed based on clear infrared thermal images. If there is uneven temperature distribution in the thermal image, such as a significant temperature difference between the wire clamp body and the wire connection, then temperature feature parameters such as maximum temperature, minimum temperature, average temperature, temperature difference, temperature variance, etc. are extracted. The shape characteristics of the temperature distribution, such as temperature gradient, hot spot area, hot spot shape, etc., can also be extracted. These feature parameters together constitute a feature data set. A feature selection algorithm is used to screen feature subsets, such as a feature selection algorithm based on mutual information, to calculate the correlation between each feature and the crimping quality. Assuming that after calculation, it is found that the correlation coefficient between the feature "temperature variance" and the crimping quality is only 0.2, which is lower than the preset threshold of 0.5, then it is considered that the feature has a small contribution to the diagnosis of crimping quality, and it is eliminated, and finally the optimized feature set is obtained. In this way, the feature dimension can be reduced, and the model training efficiency and diagnostic accuracy can be improved. The optimized feature set is transmitted to the cloud, for example, the feature data is uploaded to the cloud server via a 4G / 5G network. After receiving the data, the cloud server first performs a preliminary quality diagnosis and evaluation. For example, the feature data is input into a pre-trained threshold-based diagnostic model. According to rules such as whether the maximum temperature exceeds 80 degrees Celsius and whether the temperature difference is greater than 10 degrees Celsius, it is preliminarily judged whether there is any abnormality in the crimping quality, and a preliminary diagnostic result is obtained. Based on the preliminary diagnostic results, a distributed computing framework is used for parallel computing. For example, if the preliminary diagnostic results show that a certain wire clamp has a risk of poor crimping, the deep diagnostic process is triggered. The Spark distributed computing framework is used to parallelly calculate the historical crimping quality data of this model of wire clamp stored in the data warehouse, and the historical data is divided into several clusters through the K-means clustering algorithm. Each cluster represents a typical crimping state. Then, the random forest algorithm is used to extract key features from each cluster, such as the skewness and kurtosis of the temperature distribution, to build a historical feature database.The machine learning model is trained through the historical feature database. For example, the support vector machine (SVM) algorithm is used to train an SVM classification model using the historical feature data and the corresponding crimping quality labels (qualified, slight defects, and severe defects) as training samples. If the crimping quality grade predicted by the model does not match the actual grade, for example, a sample that is actually "slightly defective" is predicted as "qualified", the kernel function parameters and penalty coefficients of the SVM model are adjusted according to the prediction error to optimize the model performance, and finally the optimized crimping quality diagnosis model is obtained. The model parameters are updated using the incremental learning algorithm. For example, the online support vector machine (On li neSVM) algorithm is used. Whenever new thermal image data and the corresponding actual crimping quality labels are collected, they are used as new training samples, and the support vectors and weight coefficients of the SVM model are updated in real time to obtain the updated model. In this way, the model can continuously adapt to the new data distribution and improve the accuracy and robustness of the diagnosis. The inference calculation is performed through the updated model. For example, the real-time thermal image feature data uploaded by the edge computing node is input into the updated SVM model, and the model outputs the predicted grade of the crimping quality according to the learned feature patterns and decision boundaries. If the inference result shows that the crimping quality is unqualified, for example, the predicted level is "serious defect", an alarm message is generated to remind the operation and maintenance personnel that the wire clamp has a serious crimping problem, and maintenance suggestions are given, such as recommending the replacement of the wire clamp, and finally the crimping quality diagnosis result is determined. In this way, real-time, accurate, and automated diagnosis of the crimping quality of the tension wire clamp can be achieved, potential safety hazards can be discovered and handled in a timely manner, and the safe and stable operation of the power system can be guaranteed.

[0039] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent diagnosis method for the crimping quality of tension clamps based on infrared thermal imaging, characterized in that: The method comprises: According to the crimping process specification of tension clamps, the infrared thermal imaging data acquisition standard is formulated, which stipulates that the shooting angle is 45°±5°, the distance is 0.5m-1m, and the ambient temperature range is 10℃-30℃. The infrared thermal imager is used to obtain the thermal image data of the crimping part of the tension clamp, and the image registration and standardization are performed to obtain a thermal image data set with a uniform size of 1024×768 pixels and a JPEG format; Wavelet transform and morphological filtering algorithm are used to perform noise reduction and enhancement on the thermal image data set to highlight the temperature distribution characteristics of the crimping part. The thermal image is segmented using the region growing algorithm to extract the region of interest, and the characteristic parameters such as the temperature mean, variance, and gradient of the region are calculated to construct the crimping quality feature vector. According to the crimping process parameters and mechanical simulation analysis results, a mapping model between temperature distribution characteristics and crimping quality is established, and the crimping quality evaluation index is quantified. The correlation analysis and principal component analysis methods are used to screen the key temperature characteristics that have a significant impact on the crimping quality, reduce the feature dimension, and obtain the optimized crimping quality feature set. In order to solve the problem of scarcity of samples with crimping quality defects, a generative adversarial network algorithm is used to synthesize defective samples and expand the data set through adversarial learning between the generator and the discriminator. The convolutional neural network model is pre-trained on samples with normal crimping quality using transfer learning technology to extract common feature representations, and fine-tune the defective data to improve the generalization ability of the model. Data augmentation technology is used to simulate thermal images in various environments through operations such as image rotation, translation, scaling, and noise addition, to enhance the model's environmental adaptability. L2 regularization terms are added to model training to control model complexity and improve model robustness and anti-interference capabilities. In the construction of the crimping quality diagnosis model, machine learning algorithms such as support vector machine, random forest and convolutional neural network are used. Through grid search and cross-validation, model hyperparameters are optimized to improve the diagnosis accuracy. The spatial attention mechanism and gradient weighted class activation mapping interpretability module are introduced to enhance the model interpretability and output the diagnosis results such as the location, type and severity of the crimping defects. Based on the above method, an online real-time diagnosis system for the crimping quality of tension wire clamps was developed. The infrared thermal imager collects thermal images of the crimping parts in real time, transmits them to the edge computing device for preprocessing and feature extraction, and uploads them to the cloud server for quality diagnosis and evaluation. Through model verification and optimization, including using independent test sets to evaluate model performance and regularly updating model parameters, intelligent monitoring of the entire crimping quality process is achieved, crimping defects are discovered and warned in time, and safe operation of transmission lines is guaranteed.

2. The method according to claim 1, characterized in that include: Acquire a thermal image data set of the crimping position of the tension clamp, wherein the thermal image data set includes a plurality of thermal image data, each thermal image data having corresponding shooting angle, shooting distance and ambient temperature information; Performing geometric correction on the thermal image data set by using an image registration algorithm to obtain a geometrically corrected thermal image data set; Using an image standardization algorithm to unify the size and format of the thermal image data set after geometric correction to obtain a standardized thermal image data set, wherein the standardized thermal image data is 1024×768 pixels and JPEG format; Performing temperature compensation on the standardized thermal image data set according to the ambient temperature information to eliminate the influence of the ambient temperature change on the thermal image data, thereby obtaining a temperature-compensated thermal image data set; Using an image segmentation algorithm to divide the temperature-compensated thermal image data set into regions, extracting the thermal image region of the tension clamp crimping part, and obtaining a data set of an area of ​​interest; Acquire a multi-scenario adaptability knowledge base of the UAV electrical testing equipment, determine the voltage level and line type information of the current electrical testing scene according to the knowledge base, associate the voltage level and line type information with the data set of the region of interest, and construct a thermal image data set with scene labels; The thermal image data set with scene labels is analyzed by using a feature extraction algorithm to extract temperature distribution uniformity features and hot spot area features reflecting the crimping quality of the tension clamp, thereby obtaining a thermal image feature set; The thermal image feature set is input into a pre-built crimping quality assessment model based on deep learning, and compared with the historical crimping cases in the crimping quality assessment model to determine whether the current crimping quality of the tension clamp is qualified. If qualified, a crimping quality qualified report is generated; if unqualified, an unqualified crimping quality report is generated.

3. The method according to claim 1, characterized in that include: Acquire thermal image data collected by the UAV electrical testing equipment, pre-process the thermal image data, use wavelet transform and morphological filtering algorithm to perform noise reduction and enhancement processing on the thermal image data, and highlight the temperature distribution characteristics of the crimping part; According to the preprocessed thermal image data, the thermal image is segmented by using a region growing algorithm to extract the region of interest and obtain a regional image of the crimping part; Calculate the temperature mean, variance and gradient characteristic parameters from the regional image of the crimping part to construct a crimping quality characteristic vector; Inputting the crimping quality feature vector into a pre-trained crimping quality assessment model, and obtaining a crimping quality assessment result through model reasoning; If the crimping quality evaluation result is lower than a preset threshold, the crimping quality is judged to be unqualified, a crimping fault alarm is triggered, and the fault information is sent to a remote control center; If the crimping quality evaluation result is higher than a preset threshold, the crimping quality is judged to be qualified, the crimping quality evaluation result is stored in a knowledge base, and the crimping quality data is updated; According to the historical crimping quality data in the knowledge base, an incremental learning algorithm is used to update the parameters of the crimping quality assessment model in real time to improve the generalization ability and adaptability of the model; Obtain environmental perception data of the UAV electrical testing device, the environmental perception data including temperature, humidity and wind speed parameters, and perform correlation analysis on the environmental perception data and the crimping quality assessment result to determine the influence of environmental factors on the crimping quality; According to the influence rules obtained from the correlation analysis, the working parameters of the UAV electrical testing equipment are dynamically adjusted, the temperature measurement range and sensitivity of the thermal imager are adjusted, the crimping quality evaluation model is optimized, and the accuracy and reliability of crimping quality detection are improved.

4. The method according to claim 1, characterized in that: include: Acquiring crimping process parameters, wherein the crimping process parameters include crimping force and crimping time; The sensor collects the force and time data in the crimping process in real time, and transmits the collected data to the data processing module; Using the finite element analysis method, according to the crimping process parameters, a mechanical simulation analysis is performed on the crimping process to obtain stress distribution and deformation data of the crimping area; Through infrared thermal imaging technology, the temperature distribution data during the crimping process is collected in real time to obtain the temperature field distribution characteristics of the crimping area; The crimping process parameters, the stress distribution and deformation data obtained by the mechanical simulation analysis, and the temperature field distribution characteristics are integrated to construct a multi-source heterogeneous data set; Using a data mining algorithm, feature extraction and selection are performed on the multi-source heterogeneous data set, and through correlation analysis and principal component analysis methods, key features that have a significant impact on crimping quality are screened out, wherein the key features include temperature distribution features and mechanical features, to form an optimized crimping quality feature set; According to the optimized crimping quality feature set, a mapping model between temperature distribution features and crimping quality is established by using a machine learning algorithm to achieve quantitative evaluation of crimping quality; Deploy the crimping quality assessment model into an embedded system to achieve real-time monitoring and quality judgment of the crimping process; If the evaluation model determines that the crimping quality is unqualified, a warning signal is triggered; Adopting an incremental learning algorithm, the crimping quality assessment model is updated and optimized online according to the newly collected crimping process parameters and crimping quality data, so as to improve the generalization ability and robustness of the model; The crimping quality assessment result is correlated with the crimping process parameters and the mechanical simulation analysis result, and a crimping quality report is generated through data visualization technology.

5. The method according to claim 1, characterized in that include: Obtain samples with normal crimping quality, use a convolutional neural network model for pre-training, and extract the general feature representation of the sample data through the network structure and loss function; Obtain crimping quality defect samples, use a generative adversarial network algorithm to construct a generator and a discriminator network, take the defect samples as input, and train the generator to synthesize realistic defect samples through adversarial learning; Newly collected crimping quality data is obtained, and an incremental learning algorithm is used to regularly update the parameters of the defect detection model to adapt it to changes in the production environment and improve the robustness and real-time performance of the model.

6. The method according to claim 1, characterized in that include: Acquire thermal image data collected by the UAV electrical testing equipment, and preprocess the thermal image data to obtain preprocessed thermal image data, wherein the preprocessing includes image denoising, image enhancement, and image standardization; The preprocessed thermal image data is augmented by using a data augmentation technique to obtain an augmented thermal image data set, wherein the data augmentation technique includes image rotation, image translation, image scaling, and Gaussian noise addition; Dividing the augmented thermal image dataset into a training set, a validation set, and a test set; Constructing a thermal image feature extraction model, using a convolutional neural network to extract multi-scale and multi-level features of the thermal images in the training set, and obtaining a comprehensive feature representation of each thermal image in the training set through feature fusion; Based on the thermal image feature extraction model, a fault diagnosis model is constructed, and a support vector machine algorithm is used to classify the comprehensive feature representation to determine whether there is a fault in the equipment; If the fault diagnosis model determines that the equipment is faulty, the fault location model is triggered, and a decision tree algorithm is used to classify the fault type in fine granularity to determine the specific fault location and cause; According to the fault location result, a corresponding fault handling solution is obtained from a pre-built knowledge base to generate a fault repair suggestion; An online learning strategy is adopted to regularly perform incremental training and optimization on the thermal image feature extraction model, the fault diagnosis model and the fault location model according to the thermal image data newly collected by the UAV, and the model parameters are updated in real time.

7. The method according to claim 1, characterized in that include: Acquire the real-time status data of the drone, pre-process the status data through a low-power embedded processor, remove abnormal values, and obtain normalized status data; The normalized state data is modeled using a support vector machine algorithm to train an equipment fault diagnosis model; If the fault diagnosis model determines that a fault occurs, a fault alarm is triggered and the alarm information is sent to the remote control center through the UAV communication module.

8. The method according to claim 1, characterized in that: include: Obtaining real-time infrared thermal images of the crimping parts of the tension clamp collected by the infrared thermal imager; Preprocessing the infrared thermal image through an edge computing node to obtain a clear infrared thermal image; Performing feature extraction based on the clear infrared thermal image, if there is uneven temperature distribution in the infrared thermal image, extracting temperature feature parameters to obtain a feature data set; A feature selection algorithm is used to select a feature subset from the feature data set, and if the correlation between the feature in the feature subset and the crimping quality is lower than a preset threshold, the feature is eliminated to obtain an optimized feature set; The optimized feature set is transmitted to the cloud distributed big data processing platform. If the cloud server receives the data, quality diagnosis and evaluation are performed to obtain preliminary diagnosis results; According to the preliminary diagnosis results, a distributed computing framework is used to perform parallel computing on the historical crimping quality data stored in the data warehouse, and feature information in the historical data is extracted through a data mining algorithm to obtain a historical feature database; The machine learning model is trained by using the historical feature database, and if the crimping quality grade predicted by the model does not match the actual grade, the model parameters are adjusted to obtain an optimized crimping quality diagnosis model; Adopting an incremental learning algorithm, according to the newly acquired thermal image data and the corresponding actual crimping quality labels, the parameters of the crimping quality diagnosis model are updated in real time to obtain an updated model; The updated model is used to perform inference calculations on the real-time thermal image feature data uploaded by the edge computing node. If the inference results show that the crimping quality is unqualified, alarm information and maintenance suggestions are generated to determine the final crimping quality diagnosis result.

Images