A battery protection board health status testing method based on image recognition

Abstract

The present application provides a method for testing the health status of a battery protection board based on image recognition, comprising: arranging air pressure sensors at key positions and weak points of a multi-layer structure according to the analysis results of a simulation model, constructing a multi-layer structure synchronous monitoring unit, monitoring changes in pressure data of each layer, and transmitting the changes to an early warning center; collecting environmental condition information, combining air pressure sensor data and material property data including density, thermal conductivity parameters, and gas permeability parameters, to construct a mathematical model that describes gas transfer and pressure distribution in the multi-layer structure; combining image processing results with real-time pressure data to predict crack expansion trends, analyze pressure change data, extract characteristics of structural damage, identify different types of structural damage patterns, and analyze the health status of each layer of material.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for testing the health status of a battery protection board based on image recognition. Background Art

[0002] There is a key technical difficulty and contradiction in the pressure detection of multi-layer battery protection boards. Since the battery protection board is a composite structure composed of multiple functional layers, including conductive layers, insulating layers and protective layers, the differences in physical properties of different materials will have a complex impact on the transmission and pressure distribution of gas. When the battery leaks, the pressure generated at the leak point will be transmitted to the pressure sensor through each functional layer. Due to the different blocking effects of each layer of material, the pressure will be attenuated, delayed and distorted during the transmission process, which causes the pressure signal received by the pressure sensor to deviate from the actual pressure at the leak point, and even superposition and interference of pressure signals. The interface between layers in the multi-layer structure will also affect the pressure transmission. Factors such as mechanical bonding strength, roughness and defects at the interface may cause loss and reflection of the pressure signal, which further increases the difficulty of accurately detecting the pressure at the leak point. The goal of pressure detection is to accurately locate the leak point and evaluate the degree of leakage. Due to the complexity of the multi-layer structure, the pressure changes between different layers are coupled with each other, making it difficult to directly determine the source of the pressure signal. If the detected pressure change is simply attributed to a certain layer and the influence of other layers is ignored, it may lead to wrong judgment and distorted results. Summary of the invention

[0003] The present invention provides a battery protection board health status testing method based on image recognition, which mainly includes:

[0004] Obtain the physical and chemical property data of each layer of material in the multi-layer structure of the battery protection board, including the density, thermal conductivity and gas permeability of the material. Use simulation models to analyze the diffusion and penetration process of gas in the multi-layer material, identify the pressure change area, and predict the weak points.

[0005] According to the analysis results of the simulation model, air pressure sensors are arranged at key locations and weak points of the multi-layer structure to build a multi-layer structure synchronous monitoring unit to monitor the changes in pressure data of each layer and transmit them to the early warning center;

[0006] Collect environmental conditions information, combine pressure sensor data and material property data including density, thermal conductivity parameters and gas permeability parameters, and construct a mathematical model to describe gas transfer and pressure distribution in multilayer structures;

[0007] Using an intelligent dynamic pressure feedback system, the pressure data of each layer is monitored in real time and the leak point is identified. Using cameras and endoscopes, external and internal images of the multi-layer structure are collected. Using image processing modules, the external and internal images are analyzed to identify cracks or fractures in each layer of material.

[0008] Combine image processing results with real-time pressure data to predict crack expansion trends, analyze pressure change data, extract structural damage characteristics, identify different types of structural damage patterns, and analyze the health status of each layer of material;

[0009] According to the health status of each layer of material in the multi-layer structure, the maintenance and repair areas are identified, and a maintenance report is generated, which includes the health status assessment results of the multi-layer structure, the location and characteristics of cracks and structural damage, the recommended repair plan and the inspection frequency.

[0010] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0011] The present invention provides a battery protection board health status testing method based on image recognition, which, combined with a simulation model, can accurately identify the diffusion path of gas in the structure and potential weak areas, and realize accurate monitoring of pressure changes; the arranged air pressure sensor network can capture pressure fluctuations in real time, detect abnormalities in time, and provide data support for preventive maintenance; the integration of air pressure data and image processing technology can not only quickly identify the leakage point, but also accurately determine the location, length, width and fracture depth of the crack through high-precision image analysis, thereby improving the accuracy and efficiency of damage detection; combining real-time monitoring data with historical data analysis, using advanced mathematical models and machine learning algorithms, can predict the expansion trend of cracks, identify different types of structural damage patterns, and evaluate their impact on the overall structural safety, providing a scientific basis for preventing major accidents; by comprehensively analyzing the health status of each layer of materials, a comprehensive maintenance report is formed, which not only points out the current problems, but also puts forward specific repair suggestions and the frequency of subsequent monitoring, which helps to optimize maintenance strategies, extend the service life of battery protection boards, and reduce unnecessary maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flowchart of a method for testing the health status of a battery protection board based on image recognition.

[0013] Figure 2 It is a schematic diagram of a battery protection board health status testing method based on image recognition according to the present invention.

[0014] Figure 3 It is another schematic diagram of a battery protection board health status testing method based on image recognition according to the present invention. DETAILED DESCRIPTION

[0015] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0016] like Figure 1-3 In this embodiment, a method for testing the health status of a battery protection board based on image recognition may specifically include:

[0017] Step S101, obtaining the physical property data and chemical property data of each layer of material in the multi-layer structure of the battery protection plate, including the density, thermal conductivity and gas permeability of the material; analyzing the diffusion and penetration process of the gas in the multi-layer material through a simulation model, identifying the pressure change area, and predicting the weak point.

[0018] Obtain physical property data of each layer of material in the multilayer structure of the battery protection board, the physical property data including material density and thermal conductivity; input the physical property data into a material physical property database; obtain comprehensive physical property parameters of each layer of material by querying and associating the material physical property database; obtain chemical property data of each layer of material in the multilayer structure of the battery protection board, the chemical property data including gas permeability coefficient of the material; input the chemical property data into a material chemical property database; obtain comprehensive chemical property parameters of each layer of material by querying and associating the material chemical property database; establish a three-dimensional model of the multilayer structure of the battery protection board based on the comprehensive physical property parameters and the comprehensive chemical property parameters; the simulation model includes a finite element analysis method, through which the diffusion and permeation of gas in the three-dimensional model of the multilayer structure under different environmental conditions is simulated. The method comprises the following steps: a process for obtaining the spatiotemporal distribution data of the gas diffusion concentration field and the permeation flux field; performing local pressure gradient calculation and threshold comparison on the gas diffusion concentration field and the permeation flux field data, identifying the pressure area, marking the pressure area as a potential weak point area, and generating a weak point area distribution map; using a machine learning algorithm to train multiple groups of simulation analysis results, and establishing a mapping relationship model between the gas diffusion permeation behavior and the material structure and environmental conditions; predicting the weak point area distribution under different conditions through the mapping relationship model; superimposing the weak point area distribution prediction results with the design model of the battery protection board for analysis, performing material optimization and structural optimization on the areas with weak point risks, and obtaining a strengthened battery protection board design through iterative optimization; generating process parameters according to the strengthened battery protection board design, wherein the process parameters include the type, thickness, stacking order and processing technology of each layer of material, and forming a process specification file for the manufacture of the battery protection board.

[0019] Specifically, obtain the physical property data of each layer of material in the multi-layer structure of the battery protection board, including material density, specific heat capacity, thermal conductivity, thermal expansion coefficient, etc., and input the data into the material physical property database. Obtain the chemical property data of each layer of material, including the material's gas permeability coefficient, solubility coefficient, diffusion coefficient, etc., and input the data into the material chemical property database. Through database query and association analysis, the weighted average and other methods are used to calculate the comprehensive physical property parameters of each layer of material, and the comprehensive chemical property parameters are calculated based on the component concentration of the mixed gas and the permeability coefficient of each component in the material.

[0020] According to the physical and chemical property parameters of multilayer materials, and on the basis of considering the microstructure characteristics of the materials, a three-dimensional gas diffusion and permeation model of the multilayer structure of the battery protection plate is established by using methods such as dual-scale model or porous medium model. The diffusion control equation is discretized by using finite difference method or finite volume method, and solved by implicit or explicit iterative algorithm to obtain the spatiotemporal distribution data of gas diffusion concentration field and permeation flux field under different environmental conditions.

[0021] Using common machine learning algorithms such as support vector machines, random forests or neural networks, with gas diffusion permeation parameters and material structure parameters as input features and weak point area distribution as output labels, multiple groups of simulation result samples are trained to obtain weak point prediction models. According to the prediction results, the local pressure gradient of the gas diffusion concentration field and permeation flux field data is calculated, and by comparing with thresholds such as material strength limit or critical failure pressure, the weak point areas with drastic pressure changes are identified, marked on the pressure cloud map with different colors or symbols, and an intuitive weak point area distribution map is generated.

[0022] Starting from the sensitivity analysis of material performance parameters, the key material parameters that have a greater impact on the distribution of weak points are found, and the materials with the best corresponding parameters are selected from the candidate materials. Using methods such as topology optimization or shape optimization, with the optimization goal of minimizing the weak point area or maximizing the structural strength, the material distribution and structural shape are iteratively optimized to obtain the optimized design of the reinforced battery protection board.

[0023] Detailed process parameters are generated by comprehensively considering the structural characteristics, material properties and processing characteristics of the optimized design scheme. The material stacking order is determined according to the type and thickness of each layer of material, and the processing temperature, pressure and other parameters are selected according to the melting point, softening point, rheological properties, etc. of the material. The process parameter tolerance is determined according to the design dimensional accuracy requirements and the accuracy level of the processing equipment. The process specification document for the manufacture of battery protection boards is formed to guide subsequent product manufacturing and quality control.

[0024] At the same time, in the simulation of gas diffusion and permeation, different material combinations and arrangements are compared and analyzed, and the material combination and arrangement that minimizes the diffusion and permeation path is preferred. Combined with the weak point distribution information in the three-dimensional structural model, the diffusion and permeation path is quantitatively analyzed and predicted, providing an optimization basis for the structural design of the battery protection plate.

[0025] When obtaining the physical and chemical properties data of the multilayer structure material of the battery protection board, the crystal structure and phase composition of the material can be determined by an X-ray diffractometer, the thermal stability and decomposition temperature of the material can be determined by a thermogravimetric analyzer, and the thermal conductivity of the material can be determined by a laser flash method. The test temperature range of the thermal conductivity is 25℃~500℃, the test interval is 25℃, and each temperature point is measured 3 times to take the average value. The oxygen and water vapor permeability coefficients of the material are determined by the Fick diffusion law and the Time Lag method. Three samples of each material are measured and the average value is taken. According to the material ratio, the dual-scale model is used to calculate the equivalent comprehensive physical parameters of the multilayer structure. The temperature distribution is calculated by solving the steady-state heat conduction equation, and the concentration distribution is calculated by solving the gas diffusion equation. In the three-dimensional porous medium diffusion model, the finite volume method is used to numerically discretize the diffusion equation, the staggered grid is used to avoid false diffusion, and the SIMPLE algorithm is used to solve the pressure-velocity coupling, and the residual convergence criterion is 1e-6. Using the simulation result data set, the 10-fold cross validation method was used to train and optimize the random forest model, and the optimal hyperparameters were obtained through grid search, including the number of decision trees of 50, the maximum depth of a single decision tree of 30, the maximum number of features considered when dividing is the square root of the total number of features, and the minimum number of leaf node samples is 2. The trained random forest model was applied to the samples to be predicted, and the probability prediction of the weak point area was performed. The area with a probability greater than 65% was extracted as the high-risk area prone to penetration failure. Comsol finite element analysis software was used, the minimum area of ​​the high-risk area was used as the optimization objective function, the material property parameters were used as the design variables, and the sequential quadratic programming algorithm was used for optimization iteration to obtain the optimal material ratio and structural size of the battery protection plate, which reduced the high-risk penetration failure area by 72%. Combined with the optimization design results, in the injection molding process of the protection plate, the injection temperature of the PEEK material was set to 380℃, the holding pressure was set to 120MPa, the holding time was set to 15s, the cooling time was set to 30s, and the molding size tolerance was set to ±0.02mm, so as to obtain a battery protection plate product with stable performance and quality.

[0026] Step S102, according to the analysis results of the simulation model, air pressure sensors are arranged at key positions and weak points of the multi-layer structure, a multi-layer structure synchronous monitoring unit is constructed, the pressure data changes of each layer are monitored, and the pressure data are transmitted to the early warning center.

[0027] According to the analysis results of the simulation model of the multi-layer structure of the battery protection board, the weak areas and key positions of gas diffusion and penetration are identified; the optimization algorithm is used to determine the best layout plan of the air pressure sensor; the air pressure sensors are arranged in the weak areas and the key positions to build an all-round three-dimensional multi-layer structure synchronous monitoring unit; the air pressure sensor is used to monitor the changes in the pressure data inside each layer of the material in real time; the pressure data is transmitted to the central early warning monitoring center through the wireless sensor network; the data of the pressure sensors of each layer are received through the central early warning monitoring center; the multi-layer pressure data is comprehensively analyzed through the data fusion algorithm to evaluate the penetration failure risk of the battery protection board in real time; if the pressure data exceeds the preset safety threshold range, the early warning mechanism is triggered to send risk prompts and positioning information to the operation and maintenance personnel; based on the historical pressure data and the simulation model, the machine learning algorithm is used to predict the pressure change trend, identify the risk area of ​​penetration failure, and generate an early warning probability heat map;

[0028] Specifically, the pressure sensor collects the changes in pressure data inside each layer of material in real time, and transmits the pressure data to the central early warning monitoring center through a wireless sensor network built with low-power long-distance wireless communication protocols such as Zigbee and LoRa. The sensor node adopts a star topology structure, and transmits the data to the central gateway node through multi-hop routing, and then the gateway node uploads the data to the monitoring center server. The monitoring center receives the data of each layer of pressure sensors, and conducts a comprehensive analysis of the multi-layer pressure data through multi-sensor data fusion algorithms such as Kalman filtering and Bayesian estimation. By weighted fusion of data from multiple sensors, the accuracy and reliability of pressure estimation are improved, and the penetration failure risk of the battery protection board is evaluated in real time. If the pressure data exceeds the preset safety threshold range, the early warning mechanism is automatically triggered. According to the degree of deviation between the pressure data and the safety threshold, multiple warning levels are set, such as yellow warning, orange warning, red warning, etc., corresponding to different warning signal sending methods and disposal measures, and risk warnings and positioning information are sent to operation and maintenance personnel in a timely manner. Based on historical pressure data and simulation models, the monitoring center uses time series prediction models such as ARIMA and LSTM to predict the pressure change trend in the future. The prediction model is trained based on historical pressure data to achieve multi-step prediction of future pressure change trends, identify high-risk areas where penetration failure may occur, and use data visualization tools such as Python matplotlib library or Rggplot2 package to map the predicted pressure change probability into heat maps of different colors to intuitively display the distribution of high-risk areas. In the three-dimensional structural model of the battery protection board, a 1mm×1mm×1mm grid is divided, and the finite element analysis software ANSYS is used to perform transient simulation of the gas diffusion and penetration process to obtain the pressure distribution cloud map under 1000 time steps. The pressure gradient information in the cloud map is extracted by image processing algorithm, and the gradient data is linearly fitted by the least squares method to determine that the area with a pressure gradient greater than 0.5MPa / mm is a weak area for gas diffusion. The genetic algorithm is used to optimize the layout of the air pressure sensor. 20 populations, 50 evolutionary iterations, crossover probability 0.6, mutation probability 0.1, and the fitness function is the mean square error between the pressure value of each layer monitoring point and the simulation value. The layout with the highest fitness is selected as the optimal solution. The air pressure sensor uses a micro pressure sensor based on MEMS silicon piezoresistive technology with a range of 0-5MPa and an accuracy of

[0029] ±0.1%, sampling frequency 10Hz. The wireless sensor network adopts Zigbee communication protocol, transmission frequency 2.4GHz, network topology is tree structure, each layer has 1 routing node and 6 sensor nodes, and data is transmitted between routing nodes and central gateway through Modbus protocol. The monitoring center uses Kalman filter algorithm to fuse multi-sensor pressure data, and the confidence of the fused pressure value is increased by more than 20%. The warning threshold is set to 80% of the safe working pressure, that is, 4MPa. If the pressure value exceeds the warning value, the warning information is sent to the on-duty personnel through SMS and voice call. The ARIMA time series model is used to predict the pressure change trend. The pressure change trend in the next week is predicted based on the historical 60-day pressure data, and the average prediction error rate is less than 5%. The Python matplotlib library is used to draw the warning probability heat map, and the probability of the pressure warning value exceeding 4.5MPa is used as the mapping variable of the heat map to generate a 64×64 pixel color heat map.

[0030] Step S103, collecting environmental condition information, combining air pressure sensor data and material property data including density, thermal conductivity parameters and gas permeability parameters, and constructing a mathematical model describing gas transfer and pressure distribution in the multilayer structure.

[0031] According to the collected environmental condition information, air pressure sensor data and material property data, the material property data includes density, thermal conductivity parameter and gas permeation parameter; a mathematical model describing gas transfer and pressure distribution in a multilayer structure is constructed according to the environmental condition information, the air pressure sensor data and the material property data; by solving the mathematical model, the gas concentration and pressure distribution data in the multilayer structure are obtained; the actually measured air pressure sensor data is obtained, the air pressure sensor data is compared with the gas concentration and pressure distribution data, and the deviation between the air pressure sensor data and the gas concentration and pressure distribution data is calculated; if the deviation exceeds a preset threshold, it is determined that the mathematical model needs to be optimized; if the deviation is within the preset threshold range, it is determined that the mathematical model does not need to be optimized; when it is determined that the mathematical model needs to be optimized, a machine learning algorithm is used to train and adjust the parameters of the mathematical model according to the air pressure sensor data; through multiple iterative optimizations, the deviation between the output of the mathematical model and the air pressure sensor data is continuously reduced until the preset requirements are met; using the optimized mathematical model, the gas concentration and pressure distribution inside the multilayer structure under different environmental conditions are predicted; according to the prediction results, the design parameters and material selection of the multilayer structure are determined.

[0032] Specifically, when constructing a mathematical model, for example, it can be assumed that the multilayer structure consists of three layers of material, the thickness of each layer of material is 2mm, 3mm and 5mm, and the diffusion coefficient of gas between adjacent layers is 2×10-9m2 / s, 5×10-9m2 / s and 8×10-9m2 / s, respectively. When the finite element analysis method is used to discretize the multilayer structure, for example, each layer of material can be divided into 10 discrete units along the thickness direction, and the size of each discrete unit is 2mm, 3mm and 15mm. When the iterative solution algorithm is used to solve the linear algebraic equation system, for example, the iteration error threshold can be set to 1×10-6 and the maximum number of iterations can be set to 1000. The solution results show that the gas concentration value at the center of the first layer of material is 85mol / m3 and the pressure value is 2×105Pa. The gas concentration distribution cloud map generated by visualization processing shows that, for example, in the central area of ​​the second layer of material, the gas concentration value reaches 3mol / m3, which is significantly higher than other areas, indicating that there is gas aggregation in this area. If the pressure value of the edge area of ​​the third layer material is found to reach 5×105Pa in the pressure distribution cloud map, which is much higher than other areas, the material thickness of this area can be increased from 5mm to 2mm to alleviate the pressure concentration problem. After three iterations of optimization, the solution results of the mathematical model show that the gas concentration distribution and pressure distribution of each layer in the multilayer structure are uniform and stable, the maximum deviation of the gas concentration value is less than 1mol / m3, and the maximum deviation of the pressure value is less than 2×105Pa. The final multilayer structure design scheme is determined to meet the actual application requirements. The multilayer structure design scheme is: the first layer material is selected as Class A polymer material with a thickness of 2mm; the second layer material is selected as Class B ceramic material with a thickness of 8mm; the third layer material is selected as Class C metal material with a thickness of 8mm. The third layer material is arranged in the order of A / B / C.

[0033] Step S104, using an intelligent dynamic pressure feedback system to monitor the pressure data of each layer in real time and identify the leakage point, using a camera and an endoscope to collect external and internal images of the multi-layer structure; using an image processing module to analyze the external and internal images and identify cracks or fractures in each layer of material.

[0034] The real-time pressure data of the multi-layer structure is obtained, and the monitoring and analysis are performed through the intelligent dynamic pressure feedback system to identify the position coordinates of the leakage points of each layer structure; according to the position coordinates of the leakage points, the camera and the endoscope are controlled to move to the corresponding position area, and the external image and the internal image of the corresponding position area are collected; the collected external image is enhanced, and the enhanced external image is segmented by an image segmentation algorithm to extract the image block of the suspected crack area to obtain a set of suspected crack image blocks; according to the morphological characteristics of the crack, the morphological characteristics include length, width, and direction, the suspected image block is further analyzed; the edge detection algorithm and the thinning algorithm are used to extract the crack contour in the image block to obtain the crack contour feature, and the crack contour feature is analyzed by connecting and curvature to determine whether the suspected crack image block contains a real crack; the internal image is registered to align the image with the standard internal image to determine the specific position of each component in the internal image; the image block of the internal key component area is extracted by the internal image segmentation, and the fracture recognition algorithm is used to determine whether the internal key component is fractured to obtain the internal fracture detection result.

[0035] Specifically, through the monitoring and analysis of the intelligent dynamic pressure feedback system, the leak point was identified at the coordinates of the third layer (-34, 12, 78). The camera and endoscope were controlled to move to the vicinity of the leak point coordinates (45, 22, -56) to collect external and internal images of the area. The collected 500x600 pixel external image was enhanced to increase the average contrast by 30%, improve the clarity and contrast of the image, and facilitate subsequent crack identification; the enhanced external surface image was obtained, and the image was segmented using the Otsu threshold segmentation algorithm. According to the grayscale of the image and the texture features extracted by the Gabor filter, the image was divided into 8 regions to obtain a set of segmented image regions. By analyzing the shape, size, grayscale and other features of the segmented image region, it was determined whether each region was a suspected crack region. For slender regions with an area of ​​less than 50 square pixels and a grayscale value of less than 100, the corresponding image blocks were extracted to obtain a set of suspected crack image blocks. According to the morphological characteristics of the crack, such as the length is more than 5 times the width, the direction is within 10 degrees horizontally or vertically, etc., the suspected crack image block is further analyzed. The Canny edge detection algorithm is used to extract the contour, and the Zhang-Suen thinning algorithm is used to thin the crack contour to obtain the 1-pixel wide crack contour feature. By analyzing the crack contour features, such as the number of connected domains and the average curvature less than 1, it is judged whether the suspected crack image block contains a real crack. For the image blocks that meet the crack characteristics, they are marked as crack areas, and the external crack detection results are obtained. According to the external crack detection results, the OpenCV library is used to visualize the crack area, such as marking the crack position with a red rectangular box on the original image, marking the crack length and other information. At the same time, the values ​​of the crack position, length, curvature, etc. are stored in JSON format to facilitate subsequent statistical analysis and evaluation, and obtain a complete external crack detection report. At the same time, the endoscope was controlled to collect an internal image of 800×600 pixels. The feature point matching algorithm was used to align the image with the standard image. The center coordinates of the pressure valve in the image were determined to be (380, 270) and the diameter was 80 pixels. The center coordinates of the emergency shut-off valve were determined to be (650, 320) and the diameter was 60 pixels. After comprehensive analysis, it was found that there were 2 and 3 serious cracks in the second and fourth layers respectively, and a key component in the fifth layer was broken. The health status score was 65 points, which was defined as a high-risk state. The early warning information was automatically generated to notify 6 relevant personnel, and the 8 damaged locations were presented in a three-dimensional visualization to clearly show the severity of the damage.

[0036] Analyze the pressure data of each layer in real time to generate a pressure dynamic map describing the pressure changes; analyze the pressure dynamic map to determine the area of ​​abnormal pressure change, including sudden drop or unstable areas, and identify the location of the leak.

[0037] According to the real-time sampling data of the pressure sensors of each layer in the battery protection board, the sampling time interval is a preset threshold, and a pressure dynamic map reflecting the pressure distribution changes of each layer is generated, and the pressure dynamic map contains a two-dimensional pressure distribution cloud map of each layer of material and a pressure time series curve map of each sensor position; by comparing the pressure dynamic map of each layer of material with the standard pressure distribution template, a feature matching algorithm is used to determine the abnormal degree of pressure distribution, and the area with uneven pressure distribution and abnormal pressure gradient is extracted, and the area with uneven pressure distribution and abnormal pressure gradient is marked as a suspected leakage area; the pressure time series data of the suspected leakage area is detected for abnormal points, and the pressure change rate is calculated by the first-order difference method. If the change rate exceeds the preset threshold, it is determined that there is a sudden pressure drop abnormality in the area; use The DBSCAN density clustering algorithm clusters the abnormal points of sudden pressure drops, marks the cluster centers as leakage points, and determines the three-dimensional coordinates of the leakage points in combination with the spatial position of the pressure sensor; performs spectral analysis on the pressure change curve of the area less than a preset threshold value away from the leakage point, and uses fast Fourier transform to calculate the spectral characteristics of the pressure signal. If a spectral peak appears within the preset frequency band, it is judged that the pressure in the area has periodic fluctuations and belongs to an unstable pressure area; in the pressure dynamic map, the leakage point and the affected area of ​​the leakage point are highlighted, and a statistical analysis report of the position, area and shape of the leakage point is generated; the spatial position information of the leakage point is sent to the endoscope image acquisition module, instructing the endoscope to focus on the suspected leakage area and perform imaging analysis on cracks and fractures.

[0038] Specifically, based on the real-time sampling data of the pressure sensors in each layer of the battery protection board, a three-dimensional dynamic map reflecting the changes in the pressure distribution of each layer is generated at a time interval of 100ms. The map contains a two-dimensional pressure distribution cloud map of each layer of material and a pressure time series curve diagram of each sensor position. Finite element simulation or machine learning methods are used to establish a standard pressure distribution model based on the pressure distribution data of the battery under normal working conditions as a reference for judging pressure anomalies. Features such as pressure gradient, pressure mean, and pressure variance that can characterize the shape and change trend of pressure distribution are selected. By calculating the feature similarity between the actual pressure distribution and the standard template, the degree of abnormality of the pressure distribution is measured, and areas with uneven pressure distribution and abnormal pressure gradient are extracted and marked as suspected leakage areas. The pressure time series data of the suspected leakage area is detected for abnormal points. The pressure change rate is calculated by the first-order difference method. If the change rate exceeds the threshold of 3kPa / s, it is judged that there is an abnormal pressure drop in the area. By statistically analyzing a large number of pressure change rate data under normal working conditions, the probability distribution model of the pressure change rate is obtained. Assuming that the pressure change rate obeys the normal distribution and taking 3 times the standard deviation as the threshold, the confidence level is 99.7%, taking into account the detection rate and false alarm rate of abnormal points. If the pressure change rate exceeds the threshold, it is judged that there is an abnormal pressure drop in the area. The DBSCAN density clustering algorithm is used to cluster the abnormal pressure drop points. The clustering effect of different parameter combinations is evaluated by multiple iterations and cross-validation methods, and the parameter combination with the largest silhouette coefficient is selected as the optimal parameter. The cluster radius is estimated according to the spatial scale and distribution density of the leakage point, and the minimum sample number is set according to the proportion of abnormal points and the noise level. The cluster center is marked as the leakage point, and the three-dimensional coordinates of the leakage point are determined in combination with the spatial position of the pressure sensor. According to the frequency characteristics of the pressure change inside the battery protection board, a sampling frequency of 100Hz1kHz is selected to collect the pressure signal to meet the Nyquist sampling theorem. Fast Fourier transform is used to perform spectrum analysis on the pressure signal, and the appropriate time window length is selected, which is generally an integer power of 2, such as 256, 512, 1024, etc. The spectrum resolution and time resolution are weighed according to the stability and change speed of the pressure signal. If a significant spectrum peak appears in the 0.110Hz frequency band, it is judged that the pressure in the area has periodic fluctuations and belongs to the pressure instability area. The leakage point in the three-dimensional space is projected onto the two-dimensional pressure distribution map, and the maximum pooling or average pooling method is used to extract the pressure value of the unit where the leakage point is located. The regional growing algorithm is used to divide the leakage-affected area. The leakage point is used as the seed point. According to the pressure gradient and distance threshold, the regional boundary is iteratively expanded until the pressure gradient is less than the given threshold or the regional area reaches a given ratio. In the pressure dynamic map, the leakage point and its affected area are highlighted, and a statistical analysis report of the position, area, shape and other parameters of the leakage point is generated.The spatial position information of the leak point is sent to the endoscope image acquisition module to guide the endoscope to focus on the suspected leak area, and high-magnification imaging analysis is performed on structural defects such as cracks and fractures. The pressure distribution of the battery protection board at 1C, 2C, and 3C discharge rates is transiently simulated by the finite element simulation software COMSOL Multiphysics, and 100 sets of pressure cloud maps under typical working conditions are obtained. Ten characteristic parameters such as the maximum pressure, average pressure, and pressure gradient of each grid unit are extracted to construct a sample feature set. The K-means clustering algorithm is used to cluster the sample features. The optimal number of clusters is determined to be 4 by the elbow rule, and the centroid of each cluster is used as the standard pressure template. For the real-time collected pressure distribution data, the same characteristic parameters are extracted, and the Euclidean distance between it and the four standard templates is calculated. The cluster corresponding to the template with the smallest distance is the normal category of the pressure distribution. If the actual pressure distribution does not belong to any normal category, it is judged as an abnormal distribution. Statistical analysis of 500 sets of pressure change rate data under normal working conditions shows that the mean of the change rate is 0.6 kPa / s and the standard deviation is 0.2 kPa / s. Then 1.2 kPa / s is used as the threshold of pressure drop, and the corresponding confidence level is 99.7%. DBSCAN clustering is performed on the pressure drop abnormal points. The Eps parameter ranges from 0.11.0, the step length is 0.1, and the MinPts parameter ranges from 310, the step length is 1, and there are 100 sets of parameter combinations. Clustering is performed on each set of parameters, and the contour coefficient is calculated. The optimal parameters are Eps=0.5, MinPts=5, and the contour coefficient is 0.8. The pressure signal sampling frequency is set to 500 Hz, the time window length is 1024, the corresponding frequency resolution is 0.49 Hz, and the time resolution is 2.05 s. If the spectrum of the pressure signal has two obvious peaks at 0.5 Hz and 1 Hz, and the corresponding amplitude is greater than 1 kPa, it is judged that the pressure fluctuation in this area is significant. The two-dimensional pressure projection map of the leak point area is generated by the maximum pooling method, with a pooling window size of 3×3 and a step size of 1. The maximum pressure value in the 3×3 neighborhood is extracted as the pressure display value of the leak point. The growth criteria of the regional growing algorithm are: the pressure gradient is less than 0.5kPa / mm, the distance to the seed point is less than 10mm, the number of growth iterations does not exceed 20 times, and the final area does not exceed the area of ​​the entire protection plate.

[0039] When analyzing external and internal images, the image processing module identifies linear structures to determine cracks, uses morphology to measure the extension distance of the crack in the external or internal image to determine the length of the crack, counts the number of pixels on the crack cross section to determine the crack width, analyzes the edge features and brightness changes of the discontinuous area, and identifies the crack location and fracture depth.

[0040] The internal structure image of the battery protection board collected by the endoscope is converted into a grayscale image, and the grayscale image is binarized by an adaptive threshold segmentation algorithm to extract the contour and edge information of the internal structure image; the Canny algorithm is used to detect the edge of the binarized image, and the linear structure is obtained by setting high and low thresholds; the detected linear structure is subjected to Hough line transformation, and a transformation domain is obtained according to the Hough line transformation. According to the position, number and shape of the peak value in the transformation domain, it is judged whether the linear structure is a crack; if two peaks appear in the transformation domain, If the peaks are parallel to the above and the length is greater than the preset pixel threshold, it is determined that there is a crack; for the area where the crack exists, a local image is extracted, and the geometric parameters of the crack are quantitatively analyzed by using a morphological measurement method; by tracking the contour of the crack, the extension distance of the crack in the image is obtained, and the length of the crack is determined according to the extension distance; by summing the grayscale of the cross-sectional image of the crack, the number of pixels of the cross-section is obtained, and the width of the crack is calculated according to the number of pixels; the grayscale image of the crack area is segmented by local threshold, and the grayscale image of the crack area is segmented by local threshold. A regional growing algorithm is applied to the inner area of ​​the contour, and the growing area is projected in the depth direction to obtain a depth distribution map of the crack; an external image is obtained by analyzing the outer surface image of the battery protection plate taken by a CCD camera; the Roberts algorithm is used to calculate the gradient of the external image to obtain a gradient amplitude image; the gradient amplitude image is subjected to local dynamic threshold segmentation, and the area with a gradient value greater than a preset threshold is extracted as a suspected fracture area; the grayscale co-occurrence matrix is ​​calculated for the grayscale image of the suspected fracture area, and texture features are extracted. If the difference in coefficients in the texture features is greater than the area with a preset threshold, it is determined whether there is a fracture; combined with the image taken from the side, the grayscale change in the vertical direction of the suspected fracture area is analyzed, and the depth of the fracture is calculated according to the position and length of the grayscale mutation; the positions of the crack and the fracture area in the internal image and the external image are marked to obtain a spatial distribution map of the crack and the fracture; according to the length, width, and depth parameters of the crack and the fracture, combined with the spatial distribution position of the crack and the fracture, the severity of the crack and the fracture is determined to form a quantitative analysis report of the crack and fracture defects.

[0041] Specifically, the preset pixel threshold refers to 100 pixels; the spatial resolution of the image is calculated according to the width and height of the image and the corresponding actual physical size, and then converted to the corresponding pixel length as the threshold according to the minimum length requirement of the crack. For images with different resolutions, the threshold will be adaptively adjusted to improve the adaptability of detection. For suspected crack areas, local images are extracted, and the geometric parameters of the cracks are quantitatively analyzed using morphological measurement methods. By tracking the crack contour, the extension distance of the crack in the image is obtained, and the pixel distance is converted into the actual length to determine the severity of the crack. The number of pixels in the cross section is obtained by summing the grayscale of the crack cross-section image. First, a standard crack of known width is made on the sample, and the standard crack is imaged under the same imaging conditions. The number of pixels in the crack cross section is extracted, and a calibration curve between the number of pixels and the actual width is established. In actual crack detection, the crack width is quantitatively estimated according to the calibration curve, and the estimation error is calculated. The regional growth algorithm is applied to the internal area of ​​the crack contour, and the growing area is projected in the depth direction to obtain the depth distribution map of the crack. According to the size of the battery protection board and the minimum size of the defect, the appropriate CCD camera resolution is selected. According to the optical properties of the surface material, the appropriate spectral range is selected. According to the maximum depth of the defect, the appropriate depth of field range is selected. The external image is obtained by adjusting the aperture and object distance. The image gradient is calculated using the Roberts algorithm to obtain a gradient amplitude image. The mean and standard deviation of the grayscale in the local area are calculated by sliding the window. The mean plus or minus a certain multiple of the standard deviation is used as the dynamic threshold of the area. By changing the window size and the multiple factor, the fineness of the segmentation can be adjusted, and the area with a larger gradient value can be extracted as the suspected fracture area. The grayscale co-occurrence matrix of the suspected fracture area is calculated, and the texture features such as contrast, correlation, and entropy are extracted to form a feature vector. The texture feature vectors of the manually labeled normal area and fracture area are used as training samples to train the support vector machine classifier to determine whether the unknown area belongs to the fracture area and give the classification confidence. Combined with the image taken from the side, the grayscale change of the suspected fracture area in the vertical direction is analyzed, and the depth of the fracture is estimated according to the position and length of the grayscale mutation. The positions of cracks and fractures in the internal and external structural images are marked. By establishing a stereo correspondence between the endoscope and the external camera, the image coordinates of the same defect in two perspectives are triangulated to calculate the coordinates of the defect in three-dimensional space. A structured light projection device is attached to the endoscope. The three-dimensional morphology of the defect is reconstructed by projecting the deformation of the stripes on the defect surface, and the defect is accurately located in the depth direction to obtain the spatial distribution map of cracks and fractures. According to the spatial resolution of the image and the inherent error of the algorithm, the maximum error of the crack length and width measurement is estimated. The repetitive accuracy of the measurement is calculated by repeated measurements of multiple samples. The uncertainty of depth estimation is analyzed according to factors such as vibration and defocus during imaging.For the extraction of fractured areas, the missed detection rate and false detection rate were calculated by comparing with the manual annotation results to evaluate the accuracy of the detection algorithm. According to the length, width, and depth parameters of cracks and fractures, combined with their spatial distribution positions, the severity of cracks and fractures was judged, and a quantitative analysis report of crack and fracture defects was formed. The resolution of endoscopic image acquisition was 1600×1200, the field of view was 10mm×8mm, and the corresponding spatial resolution was 6.25μm / pixel. When the image was binarized, the Otsu method was used to automatically determine the optimal global threshold. When edge detection was performed, the high threshold of the Canny algorithm was set to 80, the low threshold was set to 40, and the standard deviation was 1.5. When Hough transform was performed, the θ angle step was 1°, the ρ polar step was 1pixel, and the minimum line segment length threshold was set to 1 / 10 of the image height, that is, 120pixels, corresponding to an actual length of 0.75mm. When calibrating the crack width, standard cracks with widths of 0.05, 0.10, 0.20, and 0.50 mm were made, and the logarithmic relationship between the width w and the number of pixels n was fitted: w = 0.02ln(n+1). When local threshold segmentation was performed, the window size was 50×50 and the multiplication factor was 1.5. The CCD camera used 5 million pixels, a pixel size of 3.45 μm, a physical size of 120 mm × 80 mm, and a visible spectrum band. The depth of field of 50 mm was achieved through an f / 8 aperture and a 200 mm object distance. When calculating the image gradient, the convolution kernel size of the Sobel horizontal and vertical operators was 3×3. When analyzing the texture of the fracture area, the directions of the gray-level co-occurrence matrix were selected as 0°, 45°, 90°, and 135°, the step size was 1, and an 8-bit grayscale was used. The SVM classifier used the RBF kernel function, the number of training samples was 500, the number of cross-validations was 5, and the confidence threshold was set to 0.8. During 3D reconstruction, the baseline distance between the endoscope and the external camera was 10 cm, the structured light used 5 parallel gratings, the stripe width was 1 mm, and the angle between the projector and the camera was 30°. The length measurement was repeated 30 times, the width measurement was repeated 50 times, and the depth measurement was repeated 20 times, and the confidence interval was 95%. When extracting the fracture area, 500 samples were manually annotated, including 250 normal areas and 250 fracture areas. 80% were randomly selected as the training set and 20% as the test set, and the missed detection rate and false detection rate were calculated respectively.

[0042] Step S105, combining the image processing results with the real-time pressure data to predict the crack expansion trend; analyzing the pressure change data, extracting the characteristics of structural damage, identifying different types of structural damage patterns, and analyzing the health status of each layer of material.

[0043] The crack defect parameters output by the image processing module are registered and fused with the real-time pressure data of each layer of material to establish a spatial mapping relationship between the crack defect and the pressure distribution; a data mining algorithm is used to analyze the correlation between the crack defect parameters and the real-time pressure data, and a multivariate linear regression model is constructed to fit the quantitative relationship between the length and depth of the crack and the local pressure value; if the local pressure value exceeds the yield strength threshold of the material, and the depth of the crack is greater than a preset percentage of the material thickness, it is determined that the material is at risk of overload damage; the image processing results of the same crack at multiple consecutive moments are tracked and compared to extract the expansion speed and expansion direction of the crack; combined with the force of the material The system uses the mechanical performance parameters and the fracture mechanics model to predict the expansion trend and critical size of the crack under the current stress state; performs time-frequency domain analysis on the pressure change data and uses wavelet transform to extract the characteristics of the pressure signal, including amplitude, frequency and phase; uses random forest to establish a classification model between the characteristics of the pressure signal and the structural damage mode to identify different types of damage; evaluates the crack image processing results, crack expansion prediction results, and characteristic analysis results of the pressure signal to analyze the health status of the material or structure; when the crack size, expansion speed or characteristics of the pressure signal exceed the preset danger threshold, the system generates a health warning report and takes corresponding control measures according to the danger level.

[0044] Specifically, the preset percentage is 30%; the crack defect parameters such as position, length, width, and depth output by the image processing module are registered and fused with the pressure distribution data of each layer of material at the same time, and the spatial mapping relationship between crack defects and pressure distribution is established. Time series prediction methods such as Holt-Winters exponential smoothing method and ARIMA autoregressive sliding average model are used to model the change trend of parameters such as crack length and depth in the historical time period, and extract dynamic characteristics such as the speed and acceleration of crack extension. Data mining algorithm is used to analyze the correlation between crack defect parameters and pressure distribution data, and a multivariate linear regression model is constructed to fit the quantitative relationship between crack length and depth and local pressure value. According to the plane strain fracture toughness and yield strength of the material, the critical crack size is calculated. When the crack depth exceeds 50% of the critical size and the local pressure exceeds the yield strength threshold of the material, it is determined that the layer of material has an overload damage risk. The image processing results of the same crack at multiple consecutive moments are tracked and compared to extract dynamic characteristic parameters such as the expansion speed and expansion direction of the crack tip. By using finite element simulation and other methods, the quantitative relationship between crack depth and stress concentration is analyzed. Combined with mechanical properties such as fracture toughness and fatigue life of the material, fracture mechanics models such as Paris formula and Forman equation are used to predict the crack expansion rate under the current stress state. According to the theory of structural mechanics and damage mechanics, the influence of different damage modes such as impact, scratching, extrusion, etc. on the frequency domain and time domain characteristics of the pressure signal is analyzed, and key features that can reflect the physical mechanism of damage such as spectrum peaks and signal mutation points are selected, and corresponding damage mode labels are defined. Using supervised learning algorithms such as support vector machines and random forests, a classification model between pressure signal characteristics and structural damage modes is established to achieve intelligent identification of different types of damage. The crack image processing results, crack extension prediction results, and pressure signal feature analysis results are comprehensively evaluated. Combined with other monitoring data such as vibration, acoustic emission, strain, etc., through data fusion methods such as Bayesian inference and Kalman filtering, multi-parameter comprehensive analysis and decision-making are achieved to diagnose and warn the health status of materials or structures. Construct a risk assessment matrix, comprehensively consider the danger level and weight of multiple indicators such as crack size, expansion speed, pressure signal characteristics, and calculate the comprehensive risk level of the material or structure. For different risk levels, formulate corresponding early warning strategies and control measures, and generate a health warning report. When the structure is in a high-risk state, timely take measures such as structural reinforcement, local repair, unloading or shutdown maintenance to ensure structural safety. The least squares method is used to perform multivariate linear regression on the crack defect parameters and pressure distribution data, and the regression equation is obtained: crack depth = 0.5 × local pressure + 0.2 × crack length + 0.1, and the determination coefficient R 2=0.85, indicating that the crack depth has a significant linear correlation with the local pressure and crack length. Based on the Holt-Winters exponential smoothing model, the smoothing parameters α=0.2, β=0.1, γ=0.3 are set, and the crack depth time series is exponentially smoothed three times, and the crack growth rate is 0.02mm / day and the acceleration is 0.001mm / day2. The plane strain fracture toughness KIC=30MPa·m^0.5 and the yield strength σys=400MPa, from which the critical crack size ac=1.77mm is calculated. The finite element simulation results show that when the crack depth exceeds 50% of the plate thickness, the stress concentration factor at the crack tip reaches 2.5, which exceeds the yield strength of the material. In the Paris formula da / dN=C(ΔK)^m, the material constant C=1.5×10^-11, m=3.8, stress ratio R=0.1, stress intensity factor range ΔK=15MPa·m^0.5, and the crack growth rate da / dN=2.5×10^-5mm / cycle is calculated. The pressure time series is decomposed by wavelet packet, and the db4 wavelet base is selected. The decomposition layer is 5, and the energy distribution characteristics of each frequency band are extracted. It is found that the pressure signal energy above 500Hz in the high frequency band accounts for more than 40% during impact damage, while the high frequency energy of scratch damage accounts for less than 20%. SVM is used to train the classification model of pressure signal characteristics and damage patterns, and the RBF kernel function is selected, the penalty factor C=10, the kernel function parameter γ=0.1, and the accuracy of 10-fold cross validation reaches 95%. A 5-level risk assessment matrix was established, with crack depth, expansion rate, and pressure signal mutation amplitude occupying rows, columns, and cells of the 3×3 matrix respectively. Crack depth > 1 mm and expansion rate > 0.01 mm / day were considered high risk, and pressure signal mutation amplitude > 2 times the standard deviation were considered high risk. When the number of high-risk cells exceeded 20%, the comprehensive risk level was 5, and immediate shutdown and maintenance measures were required.

[0045] The machine learning model is used to classify the image and pressure data, and the classification results are fused to identify the preset structural damage patterns, including stress corrosion cracking, fatigue cracking, and material fracture. Based on the identified structural damage patterns, the development trend of the structural damage is analyzed and the severity of the structural damage is determined.

[0046] The crack defect image and pressure signal data are preprocessed. The crack image is denoised by median filtering to extract HOG features and Gabor texture features. The pressure signal is denoised by wavelet transform to extract statistical features and frequency domain features to form a high-dimensional feature vector. A nonlinear dimensionality reduction algorithm is used to reduce the dimensionality of high-dimensional features to extract a low-dimensional feature subset that can reflect the internal structure and category distinction of the sample. A multi-level damage mode label system is constructed, including stress corrosion cracking, fatigue cracking, ductile fracture and brittle fracture, and the crack image and pressure signal data of the training sample are annotated. A deep convolutional neural network is used to construct a crack image classification model. The parameters of the crack image classification model are optimized through transfer learning and fine-tuning methods. The crack image classification model is trained to achieve damage mode classification of the crack image. An ensemble learning algorithm is used to construct a pressure signal feature classification model for the pressure signal features. The feature selection and hyperparameters of the pressure signal feature classification model are optimized through feature importance analysis and cross-validation. Parameter setting is used to realize the damage mode classification of the pressure signal; the image classification results and the pressure signal classification results are comprehensively decided, and the multi-source heterogeneous classification results are fused by the evidence fusion module to form the recognition result of the structural damage mode; based on the confidence probability of the damage mode classification, the severity of stress corrosion cracking and fatigue cracking is quantitatively evaluated; for the stress corrosion cracking, the crack growth rate and crack morphology complexity are analyzed, and the crack growth life is calculated by the Paris model; for the fatigue cracking, the crack length, crack density and crack orientation are analyzed, and the fatigue life is calculated by the cumulative damage module; for the material fracture, the fracture morphology, grain size and inclusion distribution are analyzed, and the fracture toughness is calculated by the fracture mechanics method; in combination with the crack growth rate, damage accumulation degree and fracture toughness degradation, a structural damage development trend prediction model is constructed, and the support vector machine algorithm is used to establish a nonlinear mapping relationship between the damage parameter and the service time, so as to realize the online prediction of the damage evolution trend and the remaining life prediction.

[0047] Specifically, a multi-level damage mode labeling system is constructed, including stress corrosion cracking, fatigue cracking, ductile fracture, brittle fracture and other damage categories and subcategories. The formation mechanism, macro-microscopic characteristics, influencing factors of various damage modes are defined and described in detail, the judgment criteria and classification basis of different damage modes are clarified, and the crack images and pressure signal data of the training samples are annotated. Deep convolutional neural networks such as ResNet are used to construct a crack image classification model. Through transfer learning and fine-tuning methods, parameter optimization and model training are performed on the basis of ImageNet pre-training weights to achieve damage mode classification of crack images. For pressure signal features, integrated learning algorithms such as XGBoost and LightGBM are used to construct a classification model. Through feature importance analysis and cross-validation, the feature selection and hyperparameter setting of the model are optimized to achieve damage mode classification of pressure signals. The results of each classifier are given initial evidence quality as the credibility of the evidence source, and the evidence from different evidence sources is orthogonalized and normalized to form an evidence matrix. The Dempster synthesis rule is used to synthesize the evidence in pairs to form a new evidence quality. The confidence of each damage mode is obtained by calculating the evidence support and evidence conflict, and the one with the highest confidence is selected as the final classification result. Based on the confidence probability of damage mode classification, the severity of key damage modes such as stress corrosion cracking and fatigue cracking is quantitatively evaluated. For stress corrosion cracking, the corrosion of the medium and the electrochemical properties of the material are considered, and the corrosion reaction kinetic equation is established to calculate the corrosion-induced crack tip dissolution rate. At the same time, considering the acceleration effect of the stress field on the dissolution process, the stress synergistic corrosion equation is established to form a quantitative prediction model for stress corrosion cracking. For fatigue cracks, crack density and orientation parameters are introduced on the basis of the Miner module to establish a fatigue life prediction model adapted to it. In the fatigue fracture module, the greater the crack density, the more significant the stress concentration effect between adjacent cracks, and the easier it is for cracks to extend and connect; the closer the crack orientation is to the load direction, the greater the stress intensity factor at the crack tip, and the faster the crack extension rate. Comprehensively considering quantitative indicators such as crack extension rate and damage accumulation degree, the fatigue life prediction is achieved. For material fracture, based on the fracture toughness test, scanning electron microscopy, transmission electron microscopy and other methods are used to observe and analyze the microstructural changes on the crack propagation path, focusing on microscopic damage mechanisms such as plastic deformation, microcrack initiation, and grain destruction. Through quantitative metallographic analysis, electron backscatter diffraction and other technologies, the evolution law of dislocation density, grain orientation, and strain distribution caused by damage is characterized, and the quantitative relationship between microstructural parameters and fracture toughness is constructed to achieve the evaluation of fracture toughness degradation. In terms of damage development trend prediction, on the basis of machine learning models, physical equations and statistical analysis methods of damage mechanics are introduced, such as damage evolution equations, probability distribution models, etc., to form a semi-empirical physical model that is consistent with the experimental data.On this basis, data assimilation, Bayesian inference and other methods are used to achieve real-time correction and update of physical models and monitoring data, thereby improving the applicability and robustness of damage prediction. At the same time, uncertainty quantification methods such as Monte Carlo simulation and confidence interval estimation are used to evaluate the reliability and confidence level of damage prediction results. The crack image is denoised using a 5×5 median filter. When extracting HOG features, the cell size is set to 8×8, the block size is set to 16×16, and the number of bins is set to 9, forming a 324-dimensional feature vector; the angle interval of the Gabor filter is set to 30 degrees, and the scale interval is set to 0.5, forming a 40-dimensional texture feature. The pressure signal is decomposed using a 5-level wavelet, and the db4 wavelet basis is selected to extract 12 statistical features such as the mean, variance, and kurtosis of each frequency band. Eight frequency domain features such as the mean and peak of the amplitude spectrum and phase spectrum are extracted through FFT transformation. The t-SNE algorithm was used to reduce the dimensionality of the 364-dimensional high-dimensional features. The perplexity parameter was set to 30, the learning rate was set to 10, and the number of iterations was set to 1000 to obtain a 2-dimensional damage mode feature map. A label system containing 10 damage modes was constructed, and the average number of words for the macro-micro morphology and mechanism description of each damage mode was no less than 200 words. The ResNet model adopted a 50-layer network structure, was pre-trained on ImageNet for 120 rounds, and then fine-tuned on the crack dataset for 60 rounds. The batchsize was set to 32, the learning rate was set to 0.0001, and the optimizer was Adam. The XGBoost model used 100 decision trees, the maximum tree depth was set to 6, the minimum number of leaf nodes was set to 20, the feature sampling ratio was set to 0.8, and the optimal hyperparameters were obtained through grid search optimization. In the evidence fusion module, the initial evidence quality was taken as the accuracy of the classifier on the validation set. When synthesizing evidence, the Yager rule was used to handle conflicts, and the evidence support threshold was set to 0.7. In the stress corrosion crack growth model, the thermodynamic and kinetic parameters of the material are considered, and the FEM method is used to simulate the chemical potential distribution of the corrosive medium and the stress distribution at the crack tip, the stress field is calculated to improve the dissolution rate, and the coupled corrosion and stress growth model is established. In the fatigue crack module, the three-dimensional morphology data of the crack is collected by CT, the volume density and spatial orientation tensor of the crack are calculated, and the crack growth constitutive equation based on continuous damage mechanics is established. The fracture morphology is observed by stereo microscope, the grain morphology is characterized by SEM+EBSD, the grain orientation distribution is analyzed by OIM software, the strain distribution is determined by HR-DIC method, and the strain concentration factor greater than 2.5 is judged as severe plastic deformation. The damage prediction adopts the Bayes inference method, MCMC sampling is 5000 times, and the upper and lower limits of the 95% confidence interval are taken as the life interval estimation.

[0048] Step S106, identifying the maintenance and repair areas according to the health status of each layer of material in the multi-layer structure, and generating a maintenance report, including the health status assessment results of the multi-layer structure, the location and characteristics of cracks and structural damage, the recommended repair plan and the inspection frequency.

[0049] According to the results of pressure response analysis, leakage point detection, crack and fracture defect identification of the multi-layer structure of the battery protection board, a multi-parameter feature library of material and structural state is established, including key feature parameters, such as pressure distribution, stress strain, leakage rate, crack size, and fracture morphology; the key feature parameters are compared and analyzed with the safety thresholds of material performance and structural strength, and the health factor and safety margin of each layer of material and the overall structure are calculated, wherein the health factor is the ratio of the actual value of the key feature parameter to the safety threshold, and the safety margin is the ratio of the difference between the safety threshold and the actual value to the safety threshold, and the weighted average of the health factors of each layer of material is used as the health index of the multi-layer structure; if the health factor of a layer of material is lower than the preset threshold or the safety margin is lower than the preset threshold, the health state of the layer of material is determined to be sub-healthy, and there is a potential risk of performance degradation; according to the spatial distribution law of the key feature parameters, the K-means clustering algorithm is used to divide the sub-healthy material into regions, and the optimal number of clusters is adaptively determined to obtain the plane positions and spatial boundaries of multiple sub-healthy regions; the clustering results are compared with the spatial positions of the detected cracks and fracture defects. The system matches the positions, extracts the crack and fracture defects falling within each cluster area, collects the characteristic information of the defects, and corrects the health status of each area according to the characteristic information; for weak areas with gas leakage risk, the risk level is divided into mild, moderate and severe according to the leakage rate, and marked with different colors; for areas with the crack and fracture defects, the risk level is divided into mild, moderate and severe according to the type, size, distribution density and expansion trend of the crack and fracture defects; for the gas leakage area, silicone rubber seal and epoxy seal are used to seal the gas leakage area. Resin infusion is used to repair leak points. For moderate and severe leaks, it is recommended to increase the thickness of the sealing material and the number of coating layers. For areas with cracks and fracture defects, resin impregnation, adhesive bonding and metal blending are used for crack repair and fracture bonding. For moderate and severe defects, it is recommended to carry out structural reinforcement or replace laminated materials. The inspection frequency is established according to the risk level of the cracks and fracture defects. A maintenance report is generated based on the health status assessment results of the multi-layer structure, the leakage rate analysis results, the regional analysis results of the cracks and fracture defects, the repair method results and the established inspection frequency results.

[0050] Specifically, based on the results of pressure response analysis, gas leakage detection, crack and fracture defect identification of the multi-layer structure of the battery protection board, a multi-parameter feature library of material and structural state is established, including key characteristic parameters such as pressure distribution, stress strain, gas leakage rate, crack size, fracture morphology, etc. According to the mechanical properties of the material (such as yield strength, fracture toughness, etc.) and the structural design specifications, the upper and lower limits of the safety threshold of each characteristic parameter are determined. Taking the stress index as an example, 80% of the material yield strength can be taken as the upper safety limit and 50% as the lower safety limit; taking the crack size as an example, 50% of the critical crack size corresponding to the fracture toughness can be taken as the upper safety limit. Compare and analyze these characteristic parameters with the safety threshold, calculate the health factor and safety margin of each layer of material and the overall structure, where the health factor = actual value / safety lower limit, the safety margin = (safety upper limit-actual value) / safety upper limit, and the weighted average of the health factor of each layer of material is used as the health index of the multi-layer structure. If the health factor of a layer of material is lower than 0.7 or the safety margin is lower than 0.3, the health state of the material layer is judged to be sub-healthy, and there is a potential risk of performance degradation. The coordinates and values ​​of the discrete points of the health factor or safety margin of each layer of material in three-dimensional space are used as interpolation nodes, and methods such as Kriging interpolation or radial basis function interpolation are used to generate a continuous three-dimensional health status distribution map. On this basis, the K-means clustering algorithm is used to divide the sub-healthy materials into regions. K initial cluster center points are randomly selected, the Euclidean distance from each sample point to the cluster center is calculated, the sample points are divided into the category where the nearest cluster center is located, the geometric center of each category is recalculated, and the cluster center and sample classification are iteratively optimized until the clustering results converge or the maximum number of iterations is reached. In the clustering process, the elbow method or silhouette coefficient method is used to optimize the selection of the number of categories K, and the optimal number of clusters is adaptively determined to obtain the three-dimensional spatial boundaries of multiple sub-healthy areas. The clustering results are matched with the three-dimensional spatial positions of the cracks and fracture defects detected in the previous chapters, and the crack and fracture defects falling inside each cluster area are extracted. The characteristic information such as the type, quantity, and size of the defects are statistically analyzed, and the health status of each area is corrected according to the defect characteristics. For weak areas with gas leakage risk, the risk level is divided into mild (0.10.5mL / min), moderate (0.52.0mL / min), and severe (>2.0mL / min) according to the leakage rate, and marked with different colors in the maintenance report; combined with the material property database and failure mechanism knowledge, a mapping relationship matrix between material performance, defect parameters and risk level is established. For common material defect types, qualitative and quantitative evaluation conditions are preset to obtain risk levels.For areas with crack and fracture defects, the risk level is divided into mild microcracks, crack size <1mm, moderate through cracks, crack size 1-5mm, and severe multi-layer through cracks, fractures, and crack size >5mm according to the type, size, distribution density and expansion trend of the defects, and marked in the maintenance report. Establish a correlation matrix between the material system, defect characteristics, and repair process to form a set of standardized defect repair solution library. For crack and fracture defect areas, according to the matching information of materials and defects, quickly retrieve and generate feasible repair solutions. For mild defects, crack repair is carried out by resin impregnation, adhesive coating and other methods; for moderate defects, mechanical reinforcement or replacement of part of the laminated board is carried out on the basis of resin impregnation and adhesive coating; for severe defects, structural reinforcement and overall replacement are adopted. In the maintenance report, the health status of materials and structures is dynamically updated according to the feedback of actual detection data, and the risk level and detection cycle of the corresponding area are adjusted. For mild risk areas, it is recommended to test once every three months; for moderate risk areas, it is recommended to test once every month; for severe risk areas, it is recommended to test once every week. If no defect expansion or new defects are found after multiple consecutive tests, the detection cycle of the area will be extended accordingly; on the contrary, if there is a trend of defect deterioration or leakage, the detection cycle will be shortened accordingly, and the repair plan will be updated in time. Finally, based on the three-dimensional geometric model of the multi-layer structure, a three-dimensional visualization of the material health status, risk level, and defect distribution is generated, and visual channels such as color, size, and shape are used to intuitively express the spatial distribution of the health status of materials and structures. An interactive interface is designed to allow users to query detailed information such as health status indicators, defect parameters, and repair suggestions of local areas by hovering the mouse or clicking. Functions such as multi-view switching, local zooming, and transparency adjustment are provided to facilitate users to observe the damage of the internal structure from different angles. The three-dimensional visualization results are summarized with information such as health level distribution, main risk factors, repair plans, and detection cycles to form a structured and visualized battery protection board health assessment report, providing an intuitive and reliable reference for subsequent maintenance decisions. Through finite element simulation analysis, the stress cloud map of the battery protection plate under the condition of falling and impact is obtained, and the maximum VonMises stress value is extracted as 220MPa. By querying the material mechanical properties database, the yield strength of the protection plate material is 280MPa, and the upper limit of the safety threshold is 224MPa, and the lower limit is 140MPa. Substituting into the formula, the health factor under this condition is 0.63, and the safety margin is 0.02, which is judged as a sub-healthy state. The health factors of each layer of material are weighted averaged, and the weight coefficient is determined according to the thickness of each layer and the material stiffness. The health index of the overall structure is calculated to be 0.72.The Kriging interpolation algorithm was used, and the three-dimensional coordinates and health factor values ​​of each measuring point were used as samples. By calculating the covariance matrix between the samples, a spatial interpolation model was established, and a continuous health factor distribution cloud map was generated in the three-dimensional geometric model of the protection plate. The health factor distribution cloud map was clustered by K-means, and the clustering quality was evaluated according to the silhouette coefficient. The optimal cluster number K=4 was obtained, and four sub-healthy areas were automatically divided. The spatial boundaries of each area were determined by Euler distance calculation. The three-dimensional point cloud data of cracks obtained by CT scanning were superimposed and analyzed with the clustering results, and the subsets of crack point clouds falling in each sub-healthy area were extracted. The crack size was statistically analyzed, and the maximum crack length in area 1 was 2.5 mm, the average length was 1.2 mm, and the crack density was 0.8 / cm. 2 , query the crack parameter-risk level mapping table, and determine that the risk level of the area is moderate. According to the material type and thickness information of the crack location, the repair process database is searched and matched, and the optimal repair solution is "two-component epoxy resin infusion + carbon fiber sheet winding reinforcement", and detailed process parameters and operation procedures are automatically generated. Combined with the crack growth rate model and the risk update model, the detection cycle of the moderate risk area is dynamically adjusted. According to the predicted result of the crack growth rate of 0.2mm / month, the detection cycle is extended from 1 month to 45 days. Call the three-dimensional visualization engine to render the clustering results of the sub-healthy area, and map the risk level to three colors: red, yellow, and green, to generate an intuitive health status distribution color map. The main information is extracted and summarized into a structured report, presented in the form of a three-dimensional model and a list, and embedded with an interactive function module to realize information query, screening and statistical analysis. The user can select any area with the mouse to dynamically display the risk level, crack parameters, repair plan and other detailed information of the area, and update the defect distribution and health status diagram in the three-dimensional model at the same time, so as to realize human-computer interactive information browsing and auxiliary decision-making.

[0051] The above only lists some preferred embodiments of the present invention, but the present invention is not limited thereto, and many improvements and changes can be made. As long as the improvements and changes are made on the basis of the basic principles of the present invention, they should be regarded as falling within the protection scope of the present invention.

Claims

1. A battery protection board health status testing method based on image recognition, characterized in that: The method comprises: Obtain the physical and chemical property data of each layer of material in the multilayer structure of the battery protection plate, including the density, thermal conductivity and gas permeability of the material, simulate the diffusion and penetration process of gas in the multilayer structure under different environmental conditions by finite element analysis method, obtain multiple groups of simulation analysis results, use machine learning algorithm to train the multiple groups of simulation analysis results, establish a mapping relationship model between gas diffusion and penetration behavior and material structure and environmental conditions, and predict the distribution of weak point areas under different conditions by the mapping relationship model; According to the analysis results of the simulation model, air pressure sensors are arranged at key locations and weak points of the multi-layer structure to build a multi-layer structure synchronous monitoring unit to monitor the changes in pressure data of each layer and transmit them to the early warning center; Collecting environmental condition information, air pressure sensor data and material property data, wherein the material property data includes density, thermal conductivity parameter and gas permeability parameter, and constructing a mathematical model describing gas transfer and pressure distribution in the multilayer structure according to the environmental condition information, the air pressure sensor data and the material property data; Using the intelligent dynamic pressure feedback system, the pressure data of each layer is monitored in real time and the leakage point is identified. The external and internal images of the multi-layer structure are collected using cameras and endoscopes. The external and internal images are analyzed using the image processing module to identify cracks or fractures in each layer of material and output crack defect parameters. The crack defect parameters are registered and fused with the real-time monitored pressure data of each layer, and the spatial mapping relationship between the crack defect and the pressure distribution is established. The data mining algorithm is used to analyze the correlation between the crack defect parameters and the real-time pressure data, and a multivariate linear regression model is constructed to fit the quantitative relationship between the length and depth of the crack and the local pressure value. The image processing results of the same crack at multiple consecutive moments are tracked and compared to extract the expansion speed and expansion direction of the crack; combined with the mechanical property parameters of the material, the fracture mechanics model is used to predict the expansion trend and critical size of the crack under the current stress state, and analyze the health status of each layer of material; According to the health status of each layer of material in the multi-layer structure, the maintenance and repair areas are identified, and a maintenance report is generated, which includes the health status assessment results of the multi-layer structure, the location and characteristics of cracks and structural damage, the recommended repair plan and the inspection frequency.

2. The method according to claim 1, wherein: The diffusion and penetration process of gas in the multilayer structure under different environmental conditions is simulated by a finite element analysis method to obtain multiple sets of simulation analysis results. The multiple sets of simulation analysis results are trained by a machine learning algorithm to establish a mapping relationship model between gas diffusion and penetration behavior and material structure and environmental conditions. The mapping relationship model is used to predict the distribution of weak point areas under different conditions, including: inputting the physical property data into a material physical property database; By querying and analyzing the physical property database of the materials, comprehensive physical property parameters of each layer of materials are obtained; Acquire chemical property data of each layer of material in the multi-layer structure of the battery protection plate, wherein the chemical property data includes a gas permeability coefficient of the material; inputting the chemical property data into a material chemical property database; By querying and analyzing the chemical property database of the materials, comprehensive chemical property parameters of each layer of materials are obtained; Establishing a three-dimensional model of the multilayer structure of the battery protection board according to the comprehensive physical property parameters and the comprehensive chemical property parameters; The simulation model includes a finite element analysis method, through which the spatiotemporal distribution data of the gas diffusion concentration field and the permeation flux field are obtained; Performing local pressure gradient calculation and threshold comparison on the gas diffusion concentration field and the permeation flux field data, identifying the pressure area, marking the pressure area as a potential weak point area, and generating a weak point area distribution map; The predicted distribution results of weak point areas are superimposed and analyzed with the design model of the battery protection board, and material selection and structural optimization are performed on areas with weak point risks, and a strengthened battery protection board design is obtained through iterative optimization; The process parameters are generated according to the enhanced battery protection board design scheme, and the process parameters include the type, thickness, stacking sequence and processing technology of each layer of material to form a process specification document for the manufacture of the battery protection board.

3. The method according to claim 1, wherein: According to the analysis results of the simulation model, air pressure sensors are arranged at key positions and weak points of the multi-layer structure, and a multi-layer structure synchronous monitoring unit is constructed to monitor the changes in pressure data of each layer and transmit them to the early warning center, including: According to the simulation model analysis results of the multi-layer structure of the battery protection plate, the weak areas and key positions of gas diffusion and penetration are identified; Use optimization algorithms to determine the best layout of air pressure sensors; Air pressure sensors are arranged in the weak areas and the key positions to construct an all-round three-dimensional multi-layer structure synchronous monitoring unit; The pressure sensor is used to monitor the changes in the pressure data inside each layer of material in real time; Transmitting the pressure data to a central early warning monitoring center via a wireless sensor network; Receive data from pressure sensors at each layer through the central early warning monitoring center; Comprehensively analyzing the multi-layer pressure data through a data fusion algorithm to evaluate the penetration failure risk of the battery protection plate in real time; If the pressure data exceeds the preset safety threshold, the early warning mechanism is triggered to send risk warnings and location information to the operation and maintenance personnel; Based on historical pressure data and simulation models, machine learning algorithms are used to predict pressure change trends, identify risk areas for penetration failure, and generate early warning probability heat maps.

4. The method according to claim 1, wherein: After constructing a mathematical model describing gas transfer and pressure distribution in a multilayer structure, the method further comprises: By solving the mathematical model, the gas concentration and pressure distribution data in the multilayer structure are obtained; Acquire actually measured air pressure sensor data, compare the air pressure sensor data with the gas concentration and pressure distribution data, and calculate a deviation between the air pressure sensor data and the gas concentration and pressure distribution data; If the deviation exceeds a preset threshold, it is determined that the mathematical model needs to be optimized; If the deviation is within a preset threshold range, it is determined that the mathematical model does not need to be optimized; When it is determined that the mathematical model needs to be optimized, a machine learning algorithm is used to train and adjust the parameters of the mathematical model according to the air pressure sensor data; Through multiple iterations of optimization, the deviation between the output of the mathematical model and the data of the air pressure sensor is continuously reduced until the preset requirements are met; Use the optimized mathematical model to predict the gas concentration and pressure distribution inside the multilayer structure under different environmental conditions; Based on the prediction results, the design parameters and material selection of the multi-layer structure are determined.

5. The method according to claim 1, wherein: The intelligent dynamic pressure feedback system is used to monitor the pressure data of each layer in real time and identify the leakage point. The external image and internal image of the multi-layer structure are collected by using a camera and an endoscope. The external and internal images are analyzed by an image processing module to identify cracks or fractures in each layer of material and output crack defect parameters, including: Obtain real-time pressure data of multi-layer structures, monitor and analyze through intelligent dynamic pressure feedback system, and identify the coordinates of leakage points in each layer of the structure; According to the leak point position coordinates, control the camera and the endoscope to move to the corresponding position area, and collect the external image and the internal image of the corresponding position area; Performing image enhancement processing on the collected external image, segmenting the enhanced external image using an image segmentation algorithm, extracting image blocks of suspected crack areas, and obtaining a set of suspected crack image blocks; Further analyzing the suspected image block according to morphological characteristics of the crack, wherein the morphological characteristics include length, width and direction; Using edge detection algorithm and thinning algorithm, extract the crack contour in the image block to obtain crack contour features, and analyze the crack contour features, including connectivity and curvature, to determine whether the suspected crack image block contains a real crack; Performing image registration on the internal image, aligning the image with a standard internal image, and determining specific positions of components in the internal image; Extracting image blocks of the internal key component area by the internal image segmentation, determining whether the internal key component is fractured by using a fracture recognition algorithm, and obtaining an internal fracture detection result; It also includes: real-time analysis of pressure data of each layer, generation of pressure dynamic map describing pressure changes, analysis of pressure dynamic map, determination of abnormal pressure change areas, including sudden drop or unstable areas, and identification of leakage points; when the image processing module analyzes external and internal images, it identifies linear structures to determine cracks, uses morphological measurement of the extension distance of the crack in the external or internal image to determine the length of the crack, counts the number of pixels on the crack cross section to determine the crack width, analyzes edge features and brightness changes in discontinuous areas, and identifies the crack location and fracture depth.

6. The method according to claim 5, wherein: The real-time analysis of the pressure data of each layer generates a pressure dynamic map describing the pressure change, analyzes the pressure dynamic map, determines the abnormal pressure change area, including the sudden drop or unstable area, and identifies the location of the leak point, including: According to the real-time sampling data of the pressure sensors of each layer in the battery protection board, the sampling time interval is a preset threshold, and a pressure dynamic map reflecting the pressure distribution changes of each layer is generated, and the pressure dynamic map includes a two-dimensional pressure distribution cloud map of each layer of material and a pressure time series curve diagram of each sensor position; By comparing the pressure dynamic graph of each layer of material with the standard pressure distribution template, a feature matching algorithm is used to determine the abnormal degree of pressure distribution, and the area with uneven pressure distribution and abnormal pressure gradient is extracted, and the area with uneven pressure distribution and abnormal pressure gradient is marked as a suspected leakage area; Performing abnormal point detection on the pressure time series data of the suspected leaking area, calculating the pressure change rate by the first-order difference method, and if the change rate exceeds a preset threshold, determining that there is a sudden pressure drop abnormality in the area; Clustering the pressure drop abnormal points using the DBSCAN density clustering algorithm, marking the cluster center as the leakage point, and determining the three-dimensional coordinates of the leakage point in combination with the spatial position of the pressure sensor; Performing spectrum analysis on the pressure change curve of the area whose distance from the leak point is less than a preset threshold, using fast Fourier transform to calculate the spectrum characteristics of the pressure signal, if a spectrum peak appears within the preset frequency band, it is judged that the pressure in the area has periodic fluctuations and belongs to the pressure unstable area; In the pressure dynamic map, the leak point and the area affected by the leak point are highlighted, and a statistical analysis report of the position, area and shape of the leak point is generated; The spatial position information of the leakage point is sent to the endoscope image acquisition module, instructing the endoscope to focus on the suspected leakage area and perform imaging analysis on the cracks and fractures.

7. The method according to claim 5, wherein: When analyzing the external image and the internal image, the image processing module identifies the linear structure to determine the crack, uses morphology to measure the extension distance of the crack in the external or internal image to determine the length of the crack, counts the number of pixels on the crack cross section to determine the crack width, analyzes the edge features and brightness changes of the discontinuous area, and identifies the crack location and fracture depth, including: Converting the internal structure image of the battery protection board collected by the endoscope into a grayscale image, binarizing the grayscale image using an adaptive threshold segmentation algorithm, and extracting the contour and edge information of the internal structure image; The Canny algorithm is used to detect the edge of the binary image, and the linear structure is obtained by setting high and low thresholds; Performing Hough line transform on the detected linear structure, obtaining a transform domain according to the Hough line transform, and judging whether the linear structure is a crack according to the position, number and shape of peaks in the transform domain; If there are two or more parallel peaks in the transform domain and the length of the peaks is greater than a preset pixel threshold, it is determined that a crack exists; Extracting a local image of the area where the crack exists, and using a morphological measurement method to quantitatively analyze the geometric parameters of the crack; By tracking the contour of the crack, an extension distance of the crack in the image is obtained, and a length of the crack is determined according to the extension distance; The number of pixels of the cross section is obtained by summing the grayscale of the cross-sectional image of the crack, and the width of the crack is calculated according to the number of pixels; Performing local threshold segmentation on the grayscale image of the crack region, applying a regional growing algorithm to the inner region of the crack contour, projecting the growing region in the depth direction, and obtaining a depth distribution map of the crack; The external image is obtained by analyzing the image of the outer surface of the battery protection board taken by the CCD camera; Using Roberts algorithm to calculate the gradient of the external image to obtain a gradient amplitude image; Performing local dynamic threshold segmentation on the gradient amplitude image, and extracting areas with gradient values ​​greater than a preset threshold as suspected fracture areas; Calculating a gray level co-occurrence matrix for the gray level image of the suspected fracture area, extracting texture features, and determining whether there is a fracture if the difference in coefficients in the texture features is greater than a preset threshold; Combined with the image taken from the side, the grayscale change of the suspected fracture area in the vertical direction is analyzed, and the depth of the fracture is calculated according to the position and length of the grayscale mutation; Marking the positions of the cracks and the fractures in the internal image and the external image to obtain spatial distribution maps of the cracks and the fractures; According to the length, width and depth parameters of the cracks and fractures, combined with the spatial distribution positions of the cracks and fractures, the severity of the cracks and fractures is judged to form a quantitative analysis report of the crack and fracture defects.

8. The method according to claim 1, wherein: The mechanical property parameters of the combined materials are used to predict the expansion trend and critical size of the crack under the current stress state using a fracture mechanics model, and to analyze the health status of each layer of material, including: If the local pressure value exceeds the yield strength threshold of the material, and the depth of the crack is greater than a preset percentage of the material thickness, it is determined that the material is at risk of overload damage; Performing time-frequency domain analysis on the pressure change data, using wavelet transform to extract the characteristics of the pressure signal, including amplitude, frequency and phase; Using random forest, a classification model between the characteristics of the pressure signal and the structural damage pattern is established to identify different types of damage; Evaluate the crack image processing results, crack extension prediction results and characteristic analysis results of pressure signals to analyze the health status of materials or structures; When the crack size and expansion speed or the characteristics of the pressure signal exceed the preset danger threshold, the system generates a health warning report and takes corresponding control measures according to the danger level; It also includes: using machine learning models to classify image and pressure data, and fusing the classification results to identify preset structural damage patterns, including stress corrosion cracking, fatigue cracks and material fractures, and analyzing the development trend of structural damage based on the identified structural damage patterns to determine the severity of structural damage.

9. The method according to claim 8, wherein: The machine learning model is used to classify the image and pressure data, and the classification results are integrated to identify the preset structural damage mode, including stress corrosion cracking, fatigue cracking and material fracture. According to the identified structural damage mode, the development trend of the structural damage is analyzed to determine the severity of the structural damage, including: Preprocessing the crack defect image and pressure signal data, using median filtering to denoise the crack image, extracting HOG features and Gabor texture features, and using wavelet transform to denoise the pressure signal, extracting statistical features and frequency domain features, to form a high-dimensional feature vector; A nonlinear dimensionality reduction algorithm is used to reduce the dimensionality of high-dimensional features and extract a low-dimensional feature subset that can reflect the internal structure of the sample and the class distinction; Construct a multi-level damage mode labeling system, including stress corrosion cracking, fatigue cracking, ductile fracture and brittle fracture, and annotate the crack images and pressure signal data of the training samples; A crack image classification model is constructed by using a deep convolutional neural network, and the parameters of the crack image classification model are optimized through transfer learning and fine-tuning methods, and the crack image classification model is trained to achieve damage mode classification of the crack image; An integrated learning algorithm is used to construct a pressure signal feature classification model for the pressure signal features, and feature selection and hyperparameter setting of the pressure signal feature classification model are optimized through feature importance analysis and cross-validation to achieve damage pattern classification of the pressure signal; The image classification results and the pressure signal classification results are comprehensively decided, and the multi-source heterogeneous classification results are fused using the evidence fusion module to form the recognition results of the structural damage pattern; Based on the confidence probability of the damage mode classification, a quantitative assessment of the severity of stress corrosion cracking and fatigue cracking is performed; For the stress corrosion cracking, the crack growth rate and the complexity of the crack morphology are analyzed, and the crack growth life is calculated using the Paris model; For the fatigue crack, crack length, crack density and crack orientation are analyzed, and fatigue life is calculated using a cumulative damage module; For the fracture of the material, the fracture morphology, grain size and inclusion distribution are analyzed, and the fracture toughness is calculated using the fracture mechanics method; In combination with the crack growth rate, damage accumulation degree and fracture toughness degradation, a structural damage development trend prediction model is constructed. The support vector machine algorithm is used to establish a nonlinear mapping relationship between the damage parameters and the service time, so as to realize the online prediction of the damage evolution trend and the remaining life prediction.

10. The method according to claim 1, wherein: The method of identifying maintenance and repair areas according to the health status of each layer of material in the multi-layer structure and generating a maintenance report includes the health status assessment results of the multi-layer structure, the location and characteristics of cracks and structural damage, the recommended repair plan and the inspection frequency, including: Based on the results of pressure response analysis, leakage point detection, crack and fracture defect identification of the multi-layer structure of the battery protection board, a multi-parameter feature library of material and structural state is established, including key feature parameters, such as pressure distribution, stress strain, leakage rate, crack size and fracture morphology; Compare and analyze the key characteristic parameters with the safety thresholds of material performance and structural strength, calculate the health factor and safety margin of each layer of material and the overall structure, wherein the health factor is the ratio of the actual value of the key characteristic parameter to the safety threshold, and the safety margin is the ratio of the difference between the safety threshold and the actual value to the safety threshold, and the weighted average of the health factors of each layer of material is used as the health index of the multilayer structure; If the health factor of a layer of material is lower than the preset threshold or the safety margin is lower than the preset threshold, the health status of the layer of material is judged to be sub-healthy, and there is a potential risk of performance degradation; According to the spatial distribution law of the key characteristic parameters, the sub-healthy materials are divided into regions using the K-means clustering algorithm, the optimal clustering number is adaptively determined, and the plane positions and spatial boundaries of multiple sub-healthy regions are obtained; Match the clustering results with the spatial positions of the detected cracks and fracture defects, extract the crack and fracture defects falling within each cluster area, collect the characteristic information of the defects, and modify the health status of each area according to the characteristic information; For weak areas with gas leakage risks, the risk levels are divided into mild, moderate and severe according to the leakage rate, and marked with different colors; For areas where the above-mentioned cracks and fracture defects exist, the risk levels are divided into mild, moderate and severe according to the types, sizes, distribution density and expansion trends of the above-mentioned cracks and fracture defects; For the gas leakage area, silicone rubber sealing and epoxy resin injection are used to repair the leakage point. For moderate and severe leakage, it is recommended to increase the thickness of the sealing material and the number of coating layers; For the areas with the above cracks and fracture defects, resin impregnation, adhesive bonding and metal blending are used for crack repair and fracture bonding. For moderate and severe defects, structural reinforcement or replacement of laminated sheets is recommended; Establishing the inspection frequency according to the risk level of crack and fracture defects; A maintenance report is generated based on the health status assessment results of the multi-layer structure, the leakage rate analysis results, the regional analysis results of the cracks and fracture defects, the repair method results and the established inspection frequency results.

Images