Fault detection method for high-resolution infrared thermal imaging display

Abstract

The invention discloses a fault detection method for high-resolution infrared thermal imaging display, and relates to the technical field of power equipment monitoring. According to the fault detection method for the high-resolution infrared thermal imaging display, a thermal imaging image of the surface of power equipment is obtained through high-resolution infrared thermal imaging equipment, dynamic correction is carried out on a temperature field in combination with environmental parameters and load current, and an interference-free temperature matrix is generated. The method comprises the following steps: modeling a heat conduction relationship between equipment by using a graph neural network, detecting a temperature abnormal region through a U-Net segmentation network, and generating an abnormal region mask and temperature space features; the time sequence correlation between temperature fluctuation and electrical parameters is analyzed through a dynamic time warping algorithm, and continuous faults and instantaneous interference are distinguished. And based on small sample transfer learning, in combination with a pre-training feature extractor and an attention mechanism, fault type classification is carried out, and fault types and confidence coefficients are output. Therefore, efficient and accurate detection of power equipment faults is realized.

Description

Technical Field

[0001] The invention relates to the technical field of power equipment monitoring, in particular to a fault detection method of high-resolution infrared thermal imaging display. Background Art

[0002] With the continuous growth of energy demand, the reliability and safety of power equipment have become increasingly important. As a key equipment in power transmission and distribution, the operating status of the transformer directly affects the stability and safety of the power system. Therefore, transformer fault detection and early warning play a vital role in the power industry. Especially in the rapid development of smart grids and new energy fields, real-time monitoring and fault diagnosis of transformers are particularly urgent. High-resolution infrared thermal imaging technology has become an important tool in fault diagnosis due to its non-contact, high sensitivity and ability to monitor the surface temperature changes of equipment in real time. However, under complex power equipment and environmental conditions, how to accurately and quickly detect faults and determine the type of fault is still a challenge facing current technology.

[0003] At present, traditional fault detection methods mainly rely on sensor data acquisition and empirical models, which usually have the following disadvantages: first, sensor-based monitoring methods are difficult to fully capture the thermal characteristics of equipment and are easily disturbed by environmental noise; second, existing fault diagnosis technologies mostly rely on a single temperature or electrical parameter, and lack comprehensive analysis of multi-dimensional data such as temperature, load, and environment, which limits the accuracy of detection results; in addition, data processing and feature extraction in traditional methods mostly rely on manual design, lacking flexibility and adaptability. Although deep learning and image segmentation technologies have been proposed to improve the accuracy of fault detection, in the real-time monitoring of power equipment, how to effectively integrate information from different data sources and accurately classify them is still a difficult problem to be solved. Summary of the invention

[0004] In view of the deficiencies of the prior art, the present invention provides a high-resolution infrared thermal imaging array fault detection method, which solves the problems of the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a high-resolution infrared thermal imaging display fault detection method, comprising the following steps: S1. Acquire the thermal imaging image of the transformer surface through a high-resolution infrared thermal imaging device, and dynamically correct the infrared temperature field in combination with environmental parameters and transformer load current to generate a de-interference temperature matrix; S2. According to the temperature matrix and the physical topology of the battery pack, the thermal conduction relationship between batteries is modeled through a graph neural network, and the temperature abnormality area is detected in combination with the U-Net segmentation network, and the abnormal area that conforms to the law of thermal conduction is identified, and the abnormal area mask and temperature matrix under the spatial structure constraint are generated. S3. Extract the temperature fluctuation time series data of the abnormal area mask, and synchronously match the real-time current and voltage changes of the battery pack. The correlation between temperature and electrical parameters is quantified through the dynamic time warping algorithm, and continuous faults are distinguished from transient interference. The fault detection index and evolution mode label are output; S4. According to the abnormal area mask, fault detection index and evolution mode label, the battery fault type is classified through small sample transfer learning. The pre-trained feature extractor is combined with the small sample battery fault data to fine-tune the classifier, and the temperature space features and evolution mode labels are weighted and fused in combination with the attention mechanism to output the fault type and its confidence.

[0006] Furthermore, the specific process of dynamically correcting the infrared temperature field is as follows: through the battery load current and voltage data, combined with the ambient temperature and heat dissipation wind speed, the battery load current-temperature characteristic curve is established; the infrared temperature field is dynamically corrected through the Kalman filter to eliminate the temperature error caused by environmental factors and generate interference-free infrared temperature field data; the temperature field data is calculated and corrected to eliminate the temperature rise caused by normal charging and discharging conditions, and the infrared temperature field is dynamically corrected.

[0007] Furthermore, the specific process of generating a de-interference temperature matrix is ​​as follows: performing a multivariate regression analysis on the corrected temperature field data and environmental factors; eliminating the influence of temperature rise under normal working conditions based on the battery load current-temperature characteristic curve; further optimizing the temperature data to remove interference caused by environmental noise and generate a de-interference temperature matrix.

[0008] Furthermore, according to the temperature matrix and the physical topology of the battery pack, the graph neural network is used to model the heat conduction relationship between batteries, and the U-Net segmentation network is used to detect the temperature abnormal areas. The specific process of identifying abnormal areas that conform to the law of heat conduction is as follows: According to the temperature matrix and the physical topology of the battery pack, a graph neural network model of heat conduction between batteries is established to simulate the thermal coupling relationship between each single battery in the battery pack, and the heat flow path between batteries is determined based on the battery position and heat conduction characteristics; the battery pack topology map is combined with the temperature field data, and the heat conduction law between batteries is learned through the graph neural network to optimize the spatial distribution prediction of the temperature field; based on the U-Net segmentation network, the abnormal areas in the temperature field are identified, and combined with the heat conduction law between batteries, those areas with abnormal temperature and drastic fluctuations on the heat conduction path are preferentially identified, especially the temperature difference mutation area between adjacent batteries; the temperature abnormal area is output through the detection results of the temperature abnormal area, and the specific location of the area is calibrated.

[0009] Furthermore, the specific process of generating abnormal area masks and temperature spatial features under spatial structure constraints is as follows: based on the output of the U-Net segmentation network, the spatial structure mask of the temperature abnormal area is generated by combining it with the topological map of the battery pack. The mask marks all abnormal areas where the temperature exceeds the threshold, and further eliminates noise through image post-processing; the spatial features of the abnormal area are extracted, including area, shape factor, and boundary complexity, which are used to describe the geometric characteristics of the abnormal area; based on the spatial distribution characteristics of the temperature field, the spatial temperature features related to the fault are further extracted, including temperature gradient and temperature fluctuation amplitude.

[0010] Furthermore, the correlation between temperature and electrical parameters is quantified by a dynamic time warping algorithm, and the specific process of distinguishing between persistent faults and transient interference is as follows: extracting the time series data of temperature fluctuations from the abnormal area mask, and obtaining the temperature fluctuation characteristics of each abnormal area within a certain time interval by performing time series analysis on the dynamic changes of the temperature abnormal area; synchronously collecting the electrical parameters of the battery pack including real-time current and voltage data, and aligning the electrical parameters with the temperature fluctuation time series data to ensure the time consistency of the temperature data and the electrical data; calculating the time series correlation coefficient of the temperature fluctuation data and the electrical parameters through the dynamic time warping algorithm, and quantifying the time delay relationship between the temperature fluctuation and the electrical parameters; judging the fault type based on the analysis results of the dynamic time warping algorithm and the relationship between the temperature fluctuation and the electrical parameters; if the temperature fluctuation and the electrical parameter changes are synchronized and continuous, it is determined to be a persistent fault; if the temperature fluctuation and the electrical parameter changes are short-lived and irregular, it is determined to be a fluctuation caused by transient interference.

[0011] Furthermore, the specific process of outputting the fault detection index and evolution pattern label is as follows: based on the calculated time series correlation coefficient between the temperature fluctuation data and the electrical parameters, combined with the operating status of the battery pack, a fault detection index is generated; based on the time series characteristics of the temperature fluctuation and the changing trend of the electrical data, a fault evolution pattern label is generated through the evolution pattern classification algorithm to classify the fault into continuous or intermittent.

[0012] Furthermore, the specific process of classifying battery fault types through small sample transfer learning according to the abnormal area mask, fault detection index and evolution mode label is as follows: the abnormal area mask, fault detection index and evolution mode label are used as input features to form a multi-dimensional fault feature vector; through the small sample transfer learning method, the trained network model is migrated and fine-tuned for specific battery types, and the characteristics of different battery faults are mapped to the corresponding category labels through fine-tuning.

[0013] Furthermore, the classifier is fine-tuned through the pre-trained feature extractor in combination with the small sample battery fault data, and the temperature space features and evolution mode labels are weightedly fused in combination with the attention mechanism to output the fault type and its confidence. The specific process is as follows: the abnormal area mask, fault detection index and evolution mode label are feature extracted by the pre-trained feature extractor to obtain a high-dimensional feature vector containing temperature space features and evolution mode labels; the classifier is fine-tuned in combination with the small sample battery fault data, and the network parameters are adjusted to enable the classifier to adapt to the feature mapping of the target battery fault; the attention mechanism is applied to weightedly fuse the temperature space features and evolution mode labels to enhance the contribution of key features in classification; the fine-tuned classifier is used to classify the fault type through the weighted fused feature vector, and the fault confidence is calculated based on the classification result, and finally the battery fault type and its confidence are output.

[0014] The present invention has the following beneficial effects: (1) A fault detection method based on high-resolution infrared thermal imaging display, which uses high-resolution infrared thermal imaging equipment to obtain thermal imaging images of the transformer surface, and combines environmental parameters and transformer load current to dynamically correct the infrared temperature field and generate a de-interference temperature matrix, which can effectively remove interference caused by environmental factors or load current fluctuations, provide more accurate and stable temperature data, and ensure the accuracy of subsequent fault detection. By combining the temperature matrix and the physical topology of the battery pack, the graph neural network is used to model the heat conduction relationship between batteries, so that fault detection can more accurately identify abnormal areas according to the law of heat conduction, generate abnormal area masks and temperature space features under spatial structure constraints, and effectively improve the reliability and accuracy of fault detection.

[0015] (2) A high-resolution infrared thermal imaging display fault detection method, by extracting the temperature fluctuation time series data of the abnormal area mask, and synchronously matching the real-time current and voltage changes of the battery pack, using the dynamic time warping algorithm to quantify the correlation between temperature fluctuations and electrical parameters, outputting fault detection index and evolution pattern labels, quantifying the correlation between temperature fluctuations and electrical parameters, and timely reflecting the severity and change trend of the fault through the fault detection index, providing an important basis for subsequent analysis. The battery fault type is classified through small sample transfer learning, and the classifier is fine-tuned in combination with the pre-trained feature extractor, so that the system can use the transfer learning method to improve the classification performance under limited battery fault data and realize the accurate identification of specific battery faults; by combining the attention mechanism to weightedly fuse the temperature space features and evolution pattern labels, it can increase the focus on key features, optimize the fault classification model, improve the accuracy of fault type identification, and output the corresponding confidence, providing engineers with more reliable fault diagnosis support.

[0016] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The present invention is a flow chart of a fault detection method for a high-resolution infrared thermal imaging display. DETAILED DESCRIPTION

[0018] The embodiment of the present application solves the problems of environmental interference, data inconsistency, and insufficient accuracy in fault type identification in traditional fault detection through a fault detection method based on high-resolution infrared thermal imaging. Specifically, the present invention accurately identifies abnormal surface temperature areas of transformers or battery packs by dynamically correcting the infrared temperature field, combining the physical topology of the battery pack, and efficient heat conduction modeling technology, effectively avoiding detection errors caused by external factors in traditional methods. The correlation between temperature and electrical parameters is quantified through a dynamic time warping algorithm, and the classifier is optimized in combination with small sample transfer learning and attention mechanism, which greatly improves the accuracy of fault type identification and reduces dependence on large amounts of labeled data, further improving the intelligence level and practicality of fault detection.

[0019] The overall idea of ​​the solution in the embodiments of this application is as follows: The thermal imaging image of the transformer surface is obtained through high-resolution infrared thermal imaging equipment, and the infrared temperature field is dynamically corrected in combination with environmental parameters and transformer load current to generate a temperature matrix to eliminate interference.

[0020] According to the temperature matrix and the physical topology of the battery pack, the heat conduction relationship between batteries is modeled through the graph neural network. The temperature abnormality area is detected in combination with the U-Net segmentation network, and the abnormal area that conforms to the law of heat conduction is identified. The abnormal area mask and temperature space characteristics under the spatial structure constraints are generated.

[0021] The temperature fluctuation time series data of the abnormal area mask is extracted, and the real-time current and voltage changes of the battery pack are synchronously matched. The correlation between temperature and electrical parameters is quantified through the dynamic time warping algorithm, and continuous faults are distinguished from transient interferences. The fault detection index and evolution mode label are output.

[0022] According to the abnormal area mask, fault detection index and evolution mode label, the battery fault type is classified through small sample transfer learning. The classifier is fine-tuned with the pre-trained feature extractor combined with the small sample battery fault data, and the temperature space features and evolution mode labels are weightedly fused with the attention mechanism to output the fault type and its confidence.

[0023] See also Figure 1 The embodiment of the present invention provides a technical solution: a fault detection method of high-resolution infrared thermal imaging display, comprising the following steps: S1. Acquire the thermal imaging image of the transformer surface through a high-resolution infrared thermal imaging device, and dynamically correct the infrared temperature field in combination with environmental parameters and transformer load current to generate a de-interference temperature matrix; S2. According to the temperature matrix and the physical topology of the battery pack, use a graph neural network to model the heat conduction relationship between batteries, combine with the U-Net segmentation network to detect the temperature abnormal area, identify the abnormal area that conforms to the law of heat conduction, generate the abnormal area mask and temperature space feature under the spatial structure constraint; S3. Extract the temperature fluctuation time series data of the abnormal area mask, and synchronously match the real-time current and voltage changes of the battery pack. The correlation between temperature and electrical parameters is quantified through the dynamic time warping algorithm, and continuous faults are distinguished from instantaneous interference. The fault detection index and evolution mode label are output; S4. According to the abnormal area mask, fault detection index and evolution mode label, the battery fault type is classified through small sample transfer learning. The pre-trained feature extractor is combined with the small sample battery fault data to fine-tune the classifier, and the temperature space features and evolution mode labels are weighted and fused in combination with the attention mechanism to output the fault type and its confidence.

[0024] In this embodiment, step S1: obtain a thermal imaging image of the transformer surface through a high-resolution infrared thermal imaging device, and correct the interference of the temperature data in real time with the help of environmental factors and load current data. This can eliminate the influence of factors such as temperature fluctuations and external environmental changes, ensure that the generated temperature matrix reflects the actual transformer temperature situation, and provide reliable basic data for subsequent fault detection. Step S2: Use the temperature matrix and the physical layout diagram of the battery pack to use a graph neural network (GNN) to model the thermal conduction relationship between batteries. After that, the temperature abnormality area is accurately located and detected in combination with the U-Net segmentation network. Through the U-Net segmentation network, the area that conforms to the law of heat conduction can be identified, and a "mask" is generated to indicate the location of the abnormal area, and the temperature spatial characteristics of the area are further extracted. This process helps to accurately calibrate the area where the fault occurs and provide key information for subsequent diagnosis. Step S3: By extracting the temperature fluctuation time series data and comparing it with the real-time current and voltage changes of the battery pack, the correlation between temperature and electrical parameters is further analyzed. Through the dynamic time warping (DTW) algorithm, the temperature fluctuation and the change of electrical parameters can be quantified to distinguish between continuous faults and transient interference. In addition, this step further helps to distinguish different types of faults and enhance the accuracy of the detection system by outputting the fault detection index and evolution mode label. Step S4: By using small sample transfer learning technology, combined with a small amount of existing battery fault data and a pre-trained classifier model, the battery fault type is accurately classified. The attention mechanism further helps to focus on the most recognizable information in the temperature space features and evolution mode labels to improve the classification accuracy. The final output is the fault type and its confidence value, indicating the degree of certainty of the fault type. High-resolution infrared thermal imaging equipment: This device presents the surface temperature of the object in the form of a heat map by sensing the infrared radiation radiated from the surface of the object. It is widely used in fault detection to monitor the surface temperature distribution of the equipment and help find thermal anomaly areas. Graph neural network: Graph neural network is a deep learning model that processes graph structured data and can capture the complex relationship between nodes (such as battery cells) in the graph. In the present invention, GNN is used to establish a thermal conduction relationship model between batteries. U-Net is a convolutional neural network (CNN) commonly used in medical image segmentation. It divides the image into different regions by classifying the image at the pixel level. In the present invention, U-Net is used to identify areas with abnormal temperature in transformers or battery packs, thereby locating the source of the fault. Dynamic time warping (DTW) is an algorithm that calculates the similarity between two time series, and finds the best match between the two time series by dynamically adjusting the data points on the time axis. The algorithm is used to quantify the relationship between temperature and electrical parameters (such as current and voltage) and distinguish between continuous faults and transient interference. Small sample transfer learning is a transfer learning technique that aims to migrate a model that has been trained on large-scale data to small sample data and perform fine-tuning.It helps improve the performance of the model in new areas when data is scarce. In the present invention, it is used to accurately classify the type of battery failure with a small amount of battery failure data. The attention mechanism simulates the visual attention process of the human brain, allowing the network to automatically focus on key information in the input data. In the present invention, the attention mechanism helps the model assign appropriate weights to temperature spatial features and evolution pattern labels, improving the accuracy of fault classification.

[0025] Specifically, the specific process of dynamically correcting the infrared temperature field is as follows: through the battery load current and voltage data, combined with the ambient temperature and heat dissipation wind speed, the battery load current-temperature characteristic curve is established; the infrared temperature field is dynamically corrected through the Kalman filter to eliminate the temperature error caused by environmental factors and generate interference-free infrared temperature field data; the temperature field data is calculated and corrected to eliminate the temperature rise caused by normal charging and discharging conditions, and the infrared temperature field is dynamically corrected.

[0026] In this embodiment, first, the battery load current ( ) and battery voltage ( ) data, the impact of the battery on the temperature during operation can be deduced. Since the battery generates heat during the charging and discharging process, this heat will increase with the increase of battery load, thus affecting the infrared measurement temperature. Ambient temperature It is also an important factor in temperature field correction. It will be superimposed with the battery temperature and affect the final measured temperature field. In order to model this relationship, we use a temperature characteristic curve, that is, the relationship between the battery load current and voltage and the battery temperature. This characteristic curve is usually obtained through experiments and can predict temperature changes. In order to eliminate the temperature error caused by environmental factors (such as external climatic conditions), we use a Kalman filter to dynamically correct the temperature field. The core function of the Kalman filter is to reduce the influence of noise by combining prediction and measurement data, and gradually correct the error of temperature data to generate more accurate temperature field data. In this process, we use the Kalman filter to adjust the measured value at the current moment, and combine the battery load, voltage and other parameters to continuously optimize the prediction of the temperature field, and finally eliminate the error and generate interference-free infrared temperature data. The battery will generate temperature rise during normal charging and discharging, which is a natural reaction to the battery load current. In infrared thermal imaging, normal temperature rise may affect fault detection, so the temperature data needs to be corrected. Through the battery load current and temperature characteristic curve, the temperature rise generated during normal charging and discharging is calculated, and it is removed from the actual measured temperature field data, thereby eliminating the temperature deviation generated during normal working. The corrected temperature data can better reflect the actual working state of the battery and reduce the interference of errors. The dynamic temperature correction formula sets the correction formula of the infrared temperature field as: ;in: : Corrected temperature field data. : Infrared temperature field measurement value. :Battery load current. :Battery voltage. : Ambient temperature. : Corrected temperature at the last moment. : Model parameters. Parameter explanation: : The influence coefficient of infrared temperature measurement value on the corrected temperature field, which is used to adjust the influence of temperature measurement error. : The influence coefficient of battery load current and voltage on temperature change, indicating the contribution of load current and voltage to battery temperature. and : Respectively represent the response coefficients of load current and voltage to temperature, which can be adjusted according to the working characteristics of the battery. : The influence coefficient of ambient temperature on the temperature field reflects the interference effect of changes in external ambient temperature on the battery temperature. : The correction coefficient between the last corrected temperature and the current measured temperature ensures that the correction process takes into account the continuity of the time series and the smoothness of historical data.

[0027] Specifically, the specific process of generating a de-interference temperature matrix is ​​as follows: perform a multivariate regression analysis on the corrected temperature field data and environmental factors; based on the battery load current-temperature characteristic curve, eliminate the influence of temperature rise under normal working conditions; further optimize the temperature data to remove interference caused by environmental noise and generate a de-interference temperature matrix.

[0028] In this implementation, the multiple regression analysis of the corrected temperature data and environmental factors: We have completed the dynamic correction of the temperature field through the Kalman filter and generated more accurate temperature data (i.e. ). However, the temperature data is still affected by environmental factors (such as external temperature, humidity, wind speed, etc.) and battery load current. Therefore, we further process the corrected temperature data through multiple regression analysis to eliminate these external interferences. Application of battery load current-temperature characteristic curve: Before performing multiple regression analysis, we must first consider the relationship between battery load current and battery temperature. During operation, the battery will generate additional heat due to the increase in load current. This temperature rise is inevitable under normal working conditions. We need to use the battery load current-temperature characteristic curve To quantify this effect, we can remove it from the analysis. Removal of environmental noise and optimization of temperature data: In order to further optimize the temperature data and remove the interference caused by environmental noise, we introduced environmental factors (such as ambient temperature) into the regression analysis. ,humidity , wind speed ) related items. Through the regression model, we can calculate the relationship with these external factors and remove these interference factors from the temperature data, and finally generate a de-interference temperature matrix. Generate the de-interference temperature matrix: ; Formula explanation: : Temperature matrix after interference removal, : Corrected temperature data. : The value of the battery load current-temperature characteristic curve, which represents the relationship between the battery load current and temperature. : Ambient temperature. :Ambient humidity. :wind speed. : Environmental noise interference factor, Other environment-related variables. : The coefficients of the regression model represent the degree of influence of different factors on the temperature data. These coefficients are optimized through training data. : Indicates the impact factor of environmental noise, corresponding to different interference factors. Relationship between temperature data and battery load current ( ): Relationship between corrected temperature data and battery load current. Load current will directly affect the temperature, so we use The temperature effect of the load current is expressed by the parameter By optimizing the regression coefficient , we remove the effect of battery load current from the temperature data, ensuring that only the part related to environmental interference is retained. Environmental interference removal ( ):Ambient temperature ,humidity and wind speed will interfere with the temperature data. , and , we can accurately remove the influence of these factors on the temperature data and obtain the temperature data after interference removal. ): In addition to the above factors, there may be other environmental noise (electromagnetic interference, light changes) that affect temperature measurement. These interferences are reflected by the environmental noise factor The coefficient Indicates the influence of different noise factors. Through weighted summation, these noises are removed from the temperature data, and finally the interference-free temperature matrix is ​​obtained.

[0029] Specifically, according to the temperature matrix and the physical topology of the battery pack, the graph neural network is used to model the heat conduction relationship between batteries, and the U-Net segmentation network is used to detect temperature abnormal areas. The specific process of identifying abnormal areas that conform to the law of heat conduction is as follows: According to the temperature matrix and the physical topology of the battery pack, a graph neural network model of heat conduction between batteries is established to simulate the thermal coupling relationship between each single battery in the battery pack, and the heat flow path between batteries is determined based on the battery position and heat conduction characteristics; the battery pack topology map is combined with the temperature field data, and the heat conduction law between batteries is learned through the graph neural network to optimize the spatial distribution prediction of the temperature field; based on the U-Net segmentation network, abnormal areas in the temperature field are identified, and combined with the heat conduction law between batteries, those areas with abnormal temperature and drastic fluctuations on the heat conduction path are preferentially identified, especially the temperature difference mutation area between adjacent batteries; through the detection results of the abnormal temperature area, the abnormal temperature area is output, and the specific location of the area is calibrated.

[0030] In this implementation scheme, the heat conduction relationship between batteries is modeled: the heat conduction relationship between batteries is established according to the physical topology of the battery pack. The battery pack consists of multiple single batteries, and there is a thermal coupling effect between the batteries. In order to accurately simulate this heat conduction relationship, we use a graph neural network (GNN) to model it. In a graph neural network, the battery is represented as a node in the graph, and the edges between the nodes represent the heat conduction path between the batteries. We define the weights of these edges according to the location, thermal characteristics, and connection method of the battery, thereby reflecting the heat conduction characteristics between the batteries. The graph neural network learns the law of heat conduction: The graph neural network optimizes the prediction of the spatial distribution of heat flow by iteratively calculating the message passing between nodes. Through the training process, we can obtain the temperature information of each node (battery) to estimate the temperature distribution of the entire battery pack. The goal of this step is to effectively learn the heat conduction law through the graph neural network and optimize the temperature prediction. Identification of abnormal areas: After completing the temperature prediction of the battery pack, we need to detect abnormal areas in the temperature field based on the U-Net segmentation network. U-Net is a neural network commonly used in image segmentation tasks, especially for the segmentation of structured data. Through the U-Net network, we can identify abnormal temperature areas from the temperature field. The input of U-Net is a temperature matrix, which passes the temperature information in the temperature field to the convolution layer, gradually extracts features through the encoder, and then gradually restores the spatial resolution through the decoder, and finally outputs the segmented abnormal area. Prioritize the identification of abnormal areas on the heat conduction path: Since the heat conduction between batteries is regular, when we detect abnormal areas, we will prioritize the identification of those areas that are located on the heat flow path and have abnormal temperatures and drastic fluctuations. In particular, areas with sudden changes in temperature differences between adjacent batteries are usually a precursor to failure. By combining the topological structure of the battery and the law of heat conduction, U-Net can mark these abnormal areas and provide specific location information for further analysis and repair. Generate the final output of the abnormal temperature area: Through the segmentation results of U-Net, we can output a temperature matrix containing the abnormal area and mark the specific location of the abnormal area. The output of the abnormal area includes not only the temperature anomaly, but also the corresponding spatial position and the change information of the heat conduction path.

[0031] Specifically, the specific process of generating abnormal area masks and temperature spatial features under spatial structure constraints is as follows: based on the output of the U-Net segmentation network, the spatial structure mask of the temperature abnormal area is generated by combining it with the topological map of the battery pack. The mask marks all abnormal areas where the temperature exceeds the threshold, and further eliminates noise through image post-processing; the spatial features of the abnormal area are extracted, including area, shape factor, and boundary complexity, which are used to describe the geometric characteristics of the abnormal area; based on the spatial distribution characteristics of the temperature field, the spatial temperature features related to the fault are further extracted, including temperature gradient and temperature fluctuation amplitude.

[0032] In this embodiment, a spatial structure mask is generated: the temperature field of the battery pack is processed by a U-Net segmentation network. U-Net extracts features of different scales through convolution operations, and fuses low-level features with high-level features through jump connections, so that the temperature abnormality area can be accurately identified. The output of the network is a binary image, indicating whether the temperature exceeds the predetermined threshold. On this basis, combined with the topological map of the battery pack, the temperature abnormality area can be combined with the positional relationship of the battery pack to generate a temperature abnormality area mask with spatial structural constraints. The function of the mask is to mark all areas where the temperature exceeds the set threshold and identify the battery or battery pack part that may have a fault. Image post-processing techniques such as morphological operations (corrosion, expansion, etc.) can further clean up the noise and correct the shape of the marked area to make it more accurate and avoid misidentification or over-marking. Extracting spatial features of abnormal areas: After generating the mask, we can extract a series of spatial geometric features from the temperature abnormality area, which can help us further analyze the morphology and size of the abnormal area. These spatial features include: Area: The total area of ​​the abnormal area, which can be used to quantify the size of the area. The area is usually calculated by counting the number of pixels in the mask or by the spatial position of all pixels in the area. Shape factor: used to describe the shape of the abnormal area. The shape factor can measure whether the shape of an area is regular or irregular through a series of parameters (such as aspect ratio, circularity, etc.). For example, the shape factor can be calculated based on the ratio of the perimeter and area of ​​the area. The more regular the shape, the closer its factor value is to a circle. Boundary complexity: measures the complexity of the boundary of the abnormal area. It is usually determined by calculating the curvature of the boundary of the abnormal area or the ratio of the boundary length of the polygon to its area. The more complex the boundary, the more irregular the fault area may be, indicating that there may be more serious temperature anomalies or battery failures. Extract spatial temperature features related to faults: Further, combined with the spatial distribution characteristics of the temperature field, spatial temperature features related to battery failures can be extracted, which can help diagnose the cause of the fault more accurately. Temperature gradient: The temperature gradient refers to the rate of temperature change in space, that is, the amount of temperature change per unit distance. In a battery pack, changes in temperature gradients are often related to uneven heat conduction or faulty areas. By calculating the temperature difference between adjacent pixels, the distribution of temperature gradients can be obtained, and then it can be determined whether the temperature change conforms to the expected heat conduction law. Temperature fluctuation amplitude: The temperature fluctuation amplitude describes the degree of temperature fluctuation in time or space. A large temperature fluctuation amplitude in an abnormal area may mean thermal runaway or battery failure. By calculating the temperature fluctuation amplitude in the temperature abnormal area, it is possible to further evaluate whether the area is in an abnormal state and combine it with electrical parameters to analyze the nature of the fault.

[0033] Specifically, the dynamic time warping algorithm is used to quantify the correlation between temperature and electrical parameters, and the specific process of distinguishing between persistent faults and transient interference is as follows: extract the time series data of temperature fluctuations from the abnormal area mask, and obtain the temperature fluctuation characteristics of each abnormal area within a certain time interval by performing time series analysis on the dynamic changes of the temperature abnormal area; synchronously collect the electrical parameters of the battery pack including real-time current and voltage data, and align the electrical parameters with the temperature fluctuation time series data to ensure the time consistency of the temperature data and the electrical data; calculate the time series correlation coefficient of the temperature fluctuation data and the electrical parameters through the dynamic time warping algorithm, and quantify the time delay relationship between the temperature fluctuation and the electrical parameters; judge the fault type according to the analysis results of the dynamic time warping algorithm and the relationship between the temperature fluctuation and the electrical parameters. If the temperature fluctuation and the electrical parameter changes are synchronized and continuous, it is determined to be a persistent fault. If the temperature fluctuation and the electrical parameter changes are short-lived and irregular, it is determined to be a fluctuation caused by transient interference.

[0034] In this implementation scheme, extracting time series data of temperature fluctuation from the abnormal area mask: extracting time series data of temperature fluctuation from the temperature abnormal area mask generated by the U-Net segmentation network. Each temperature abnormal area has its own specific temperature change pattern, which is usually manifested as fluctuations over time. By continuously monitoring the temperature abnormal area, time series data of temperature fluctuation can be formed. This time series data contains the fluctuation of temperature at different time points and is an important basis for analyzing the failure mode of the battery pack. Synchronous acquisition of electrical parameter data: Synchronous acquisition of electrical parameter data of the battery pack, such as real-time current and voltage. Electrical parameters have a significant impact on the operating state and thermal performance of the battery pack, so it is crucial to compare and analyze electrical data with temperature fluctuation data. In practical applications, the acquisition of electrical parameters (such as current and voltage) and temperature data usually has different time granularity or frequency. In order to ensure the time series consistency of temperature fluctuation data and electrical data, we need to time align the data of the two. A common practice is to align the two types of data to the same time step by interpolation method or by selecting a uniform time interval. Application of Dynamic Time Warping (DTW) algorithm: Dynamic Time Warping (DTW) algorithm is used to calculate the similarity between two sets of time series, especially when there is time offset or misalignment in the time series, DTW can effectively align these sequences and find the best matching path. In this process, the DTW algorithm is used to quantify the correlation between temperature fluctuation time series data and electrical parameter time series data. The core purpose of DTW calculation is to measure the time delay relationship between temperature fluctuation and electrical parameter change, that is, the degree of time alignment between the two. DTW finds the optimal matching path between them by minimizing the distance between the two time series (implemented by dynamic programming). Determine the fault type: continuous fault and instantaneous interference: Based on the correlation coefficient calculated by the DTW algorithm, we can derive the time delay relationship between temperature fluctuation and electrical parameters. If the temperature fluctuation is synchronized and continuous with the change of electrical parameters (that is, the two are well aligned in time, and this correlation remains unchanged for a long time), it can be judged as a continuous fault. Continuous faults may be caused by internal problems of the battery (such as internal short circuit or thermal runaway), which are manifested as continuous and regular temperature fluctuations. If the temperature fluctuation and electrical parameter changes are short-lived and irregular (i.e., the correlation between the two in time is low and there are obvious fluctuations), it can be judged that the fluctuation is caused by transient interference. Transient interference is usually caused by external environmental factors or accidental events (such as electrical noise, transient load changes), and the temperature fluctuation is manifested as sudden and short-lived fluctuations.

[0035] Specifically, the specific process of outputting the fault detection index and evolution pattern label is as follows: based on the calculated time series correlation coefficient between the temperature fluctuation data and the electrical parameters, combined with the operating status of the battery pack, a fault detection index is generated; based on the time series characteristics of the temperature fluctuation and the changing trend of the electrical data, a fault evolution pattern label is generated through the evolution pattern classification algorithm to classify the fault into continuous or intermittent.

[0036] In this embodiment, the time series correlation coefficient between temperature fluctuation and electrical parameters is calculated: the time series correlation coefficient between temperature fluctuation and electrical parameters is calculated by dynamic time warping (DTW) algorithm or other time series analysis methods. This coefficient reflects the synchronization between temperature fluctuation and changes in electrical parameters (such as current and voltage). The formula is as follows: ;in: Represents the time series correlation coefficient. Indicates time point The temperature fluctuation data A measurement value. and The time points The mean and standard deviation of the temperature fluctuation data. Indicates time point The electrical parameters (current, voltage) A measurement value. and The time points The mean and standard deviation of the electrical data on the battery pack. This correlation coefficient is used to quantify the degree of synchronization between temperature fluctuations and changes in electrical parameters, thereby helping to identify whether a fault exists. Generate a fault detection index: By calculating the timing correlation coefficient between temperature fluctuations and electrical parameters, and combining the operating status of the battery pack (such as charging, discharging, standby, etc.), a fault detection index can be generated. This index reflects the severity and possibility of a fault. A weighted synthesis method can be used to combine the timing correlation coefficient with the operating status weight of the battery pack to generate a comprehensive fault detection index. The calculation formula for the fault detection index is as follows: ;in: Stands for Fault Detection Index. Represents the time series correlation coefficient. It is a quantitative indicator of the current operating status of the battery pack, with a value range between 0 and 1, reflecting the current working mode of the battery. and is a weighting coefficient that controls the influence of the timing correlation coefficient and the battery pack status on the fault detection index. The higher the value of the fault detection index, the greater the possibility of the fault; vice versa. By analyzing the timing characteristics of temperature fluctuations and the changing trend of electrical parameters, combined with the evolutionary pattern classification algorithm (support vector machine), the fault can be divided into different evolutionary patterns. Generally, the evolutionary patterns of faults can be divided into continuous faults and intermittent faults. Continuous faults: refers to the continuous trend of temperature fluctuations and electrical parameter changes, and the fault condition will not recover in a short time. Intermittent faults: refers to the periodic or sporadic fluctuations of temperature fluctuations and electrical parameter changes, and the fault state is sometimes present and sometimes absent. The generation of evolutionary pattern labels is based on the timing characteristics of temperature fluctuations (such as fluctuation amplitude, periodicity) and the changing trend of electrical data. The fault mode can be determined by the timing classification algorithm, and an evolutionary pattern label is assigned to each fault event.

[0037] Specifically, the specific process of classifying battery fault types through small sample transfer learning according to abnormal area masks, fault detection indexes and evolutionary mode labels is as follows: the abnormal area masks, fault detection indexes and evolutionary mode labels are used as input features to form a multi-dimensional fault feature vector; through the small sample transfer learning method, the trained network model is migrated and fine-tuned for specific battery types, and the characteristics of different battery faults are mapped to corresponding category labels through fine-tuning.

[0038] In this implementation scheme, a multi-dimensional fault feature vector is generated: the abnormal region mask, fault detection index and evolution mode label are used as input features, and a multi-dimensional information vector is formed through feature fusion. Each input feature provides information on different aspects of battery failure: abnormal region mask: represents the spatial distribution of temperature abnormality area, and provides information on the spatial position and morphology of the fault area. Fault detection index: quantifies the temporal correlation between temperature fluctuation and electrical parameters, reflecting the severity and possibility of the fault. Evolution mode label: describes the evolution mode of the fault, distinguishing between continuous fault and intermittent fault. After combining these three types of feature data, a multi-dimensional fault feature vector can be obtained. This feature vector includes temperature spatial features, temporal features, and fault evolution mode features, providing rich input information for subsequent classification tasks. Few-shot transfer learning: Since the actual battery fault data set often has insufficient samples, the few-shot transfer learning method can be used to classify the fault type. Few-shot transfer learning improves classification performance by migrating the network model that has been trained on a large-scale data set to a new task with less data. At this stage, the process is as follows: Transfer pre-trained model: Select a network model that has been trained on similar tasks. Usually, these models have good learning ability and feature extraction ability on large-scale datasets (for example, using convolutional neural networks (CNN) or graph neural networks (GNN), etc.). These pre-trained models have learned general feature representations, so they can be used for battery fault classification tasks. Fine-tuning: Apply the pre-trained model to the new battery fault classification task through fine-tuning methods. The purpose of fine-tuning is to adjust the parameters of the model based on new small sample data to adapt it to specific battery fault types. This is usually done by local training on a small amount of battery fault data, and the network maps the characteristics of different battery faults to the corresponding category labels by adjusting its weights. During the fine-tuning process, the focus is on retaining the general feature extraction layers in the pre-trained model and training the last few layers in a targeted manner so that the model can adapt to the new fault classification task. Classification process: After fine-tuning with small sample transfer learning, the model is able to classify the battery fault type based on the input fault feature vector (containing information such as abnormal area mask, fault detection index, evolution mode label, etc.). By using transfer learning, the model can achieve good performance on a small amount of training data, thus effectively solving the data scarcity problem in battery fault classification. During the classification process, the model maps the input feature vector to the category label of the fault type based on the learned relationship between the fault features and labels. For example, possible fault types include short circuit fault, overheating fault, internal damage, etc.Finally, the model will output a prediction result of the fault type, indicating whether the battery has failed and determining the specific type of fault. Through such classification, the operating status of the battery can be diagnosed more accurately, helping to take maintenance measures in a timely manner.

[0039] Specifically, the classifier is fine-tuned by combining the pre-trained feature extractor with the small sample battery failure data, and the temperature space features and evolution mode labels are weightedly fused in combination with the attention mechanism to output the fault type and its confidence. The specific process is as follows: the abnormal area mask, fault detection index and evolution mode label are feature extracted by the pre-trained feature extractor to obtain a high-dimensional feature vector containing temperature space features and evolution mode labels; the classifier is fine-tuned in combination with the small sample battery failure data, and the network parameters are adjusted to enable the classifier to adapt to the feature mapping of the target battery failure; the attention mechanism is applied to weightedly fuse the temperature space features and evolution mode labels to enhance the contribution of key features in classification; the fine-tuned classifier is used to classify the fault type through the weighted fused feature vector, and the fault confidence is calculated based on the classification result, and finally the battery fault type and its confidence are output.

[0040] In this embodiment, in the first step of this process, a pre-trained feature extractor is used to extract features from the input data. The goal of the feature extractor is to extract useful feature information from the abnormal region mask, fault detection index and evolution mode label to obtain a high-dimensional feature vector, which mainly includes the following aspects: Abnormal region mask: output by the U-Net segmentation network, representing the spatial distribution of the temperature abnormal region, identifying the battery region where the fault may occur. Fault detection index: the time series correlation coefficient between the temperature fluctuation data and the electrical parameters calculated by the dynamic time warping algorithm, reflecting the time series characteristics of the battery fault. Evolution mode label: represents the evolution trend of the fault and distinguishes between continuous and intermittent fault modes. Multi-level high-dimensional features are extracted from the above input data by pre-training feature extractors (usually deep neural networks, such as convolutional neural networks (CNNs) or graph neural networks (GNNs)). These features include temperature spatial distribution characteristics and fault evolution patterns, which can provide sufficient information for subsequent classification tasks. Fine-tuning the classifier: After obtaining the high-dimensional feature vector, the next step is to fine-tune the classifier using a small sample of battery fault data. The purpose of the fine-tuning process is to apply the feature extraction capability of the pre-trained model to a specific battery fault classification task and adjust the classifier through a small amount of target battery fault data. Small sample learning: Since it is usually difficult to obtain a large number of samples of battery fault data, the fine-tuning process will use the trained network parameters to enable the classifier to adapt to the battery fault features contained in a small amount of data. Objective: Through the fine-tuning process, the weights and biases of the network are adjusted so that the classifier can accurately map the battery fault data to the corresponding fault type. The fine-tuning process is usually carried out on the basis of the pre-trained model, maintaining the generalization ability of the feature extraction part, while refining the classification layer through a small amount of data to ensure that the classifier can accurately identify different battery faults. Apply the attention mechanism for weighted fusion: In order to improve the model's attention to key features, the attention mechanism is applied to weighted fusion of temperature space features and evolution mode labels. The attention mechanism can dynamically assign different weights to different features according to the correlation of the input features, so as to better highlight the role of key features in classification. Temperature space features: Describes the temperature distribution of the battery area, including the heat conduction relationship between batteries, temperature gradients, and temperature fluctuation amplitudes. Evolutionary pattern label: describes the evolution of the fault, indicating whether the fault is persistent or intermittent. Through the attention mechanism, the model automatically learns which features are more important for fault classification and increases their weights accordingly. This weighted fusion method helps improve the accuracy of classification, especially when the input features have different importance. Fault type classification and confidence calculation: In the weighted fusion feature vector after the attention mechanism is applied, the fine-tuned classifier is used to classify the battery fault type. The classifier classifies according to the fault features learned during training, mapping each sample to the corresponding fault type.In addition, the classifier calculates the fault confidence based on the input feature vector. Confidence refers to the degree of trust the model has in its classification results, which is usually obtained by calculating the probability value of each fault type. For example, the model may output a probability value of 0.85 for a fault type of "overheating fault", which means that the model is 85% confident that the battery has an overheating fault. Ultimately, the system outputs the battery fault type and confidence, which provides a clear and reliable basis for diagnosing the battery status. A high confidence prediction usually indicates a higher severity of the battery fault, while a low confidence may indicate a greater uncertainty in the fault type.

[0041] In summary, this application has at least the following effects: The present invention provides a fault detection method for high-resolution infrared thermal imaging display, which aims to improve the fault detection capability of equipment such as transformers or battery packs. The thermal imaging image of the equipment surface is obtained by a high-resolution infrared thermal imaging device, and the temperature field is dynamically corrected in combination with environmental parameters and load current to generate a temperature matrix for interference removal. The heat conduction relationship between the devices is modeled by a graph neural network, and the temperature abnormality area is detected by a U-Net segmentation network to generate the abnormal area mask and temperature spatial features under the spatial structure constraint. Further, the temperature fluctuation time series data of the abnormal area is extracted, and the time series correlation between the temperature and electrical parameters is analyzed by a dynamic time warping algorithm in combination with the real-time electrical parameters of the battery pack, and the continuous fault and instantaneous interference are distinguished, and the fault detection index and evolution mode label are generated. Finally, the small sample transfer learning method is used to fine-tune the classifier in combination with the pre-trained feature extractor, and the temperature spatial features and evolution mode labels are weighted and fused by the attention mechanism to classify the battery fault type, and output the fault type and its confidence. The method realizes efficient and accurate fault detection by comprehensively utilizing high-resolution infrared imaging, graph neural network, dynamic time warping, transfer learning and attention mechanism, and has broad application prospects.

[0042] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0043] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0044] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0045] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0046] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0047] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A high-resolution infrared thermal imaging display fault detection method, characterized in that: The following steps are involved: S1. Obtain thermal imaging images of the transformer surface through high-resolution infrared thermal imaging equipment, and dynamically correct the infrared temperature field in combination with environmental parameters and transformer load current to generate a temperature matrix to eliminate interference; S2. Based on the temperature matrix and the physical topology of the battery pack, the heat conduction relationship between batteries is modeled through the graph neural network, and the temperature abnormality area is detected in combination with the U-Net segmentation network, the abnormal area that conforms to the heat conduction law is identified, and the abnormal area mask and temperature space features under the spatial structure constraint are generated; S3. Extract the temperature fluctuation time series data of the abnormal area mask, and synchronously match the real-time current and voltage changes of the battery pack. Quantify the correlation between temperature and electrical parameters through the dynamic time warping algorithm, distinguish between continuous faults and transient interference, and output the fault detection index and evolution mode label; S4. According to the abnormal area mask, fault detection index and evolution mode label, the battery fault type is classified through small sample transfer learning. The pre-trained feature extractor is combined with the small sample battery fault data to fine-tune the classifier, and the temperature space features and evolution mode labels are weightedly fused with the attention mechanism to output the fault type and its confidence.

2. A high-resolution infrared thermal imaging display fault detection method according to claim 1, characterized in that: The specific process of dynamic correction of infrared temperature field is as follows: The battery load current-temperature characteristic curve is established by combining the battery load current and voltage data with the ambient temperature and cooling wind speed; Dynamically correct the infrared temperature field through the Kalman filter to eliminate the temperature error caused by environmental factors and generate interference-free infrared temperature field data; Calculate and correct the temperature field data, eliminate the temperature rise effect caused by normal charging and discharging conditions, and dynamically correct the infrared temperature field.

3. A high-resolution infrared thermal imaging display fault detection method according to claim 2, characterized in that: The specific process of generating the interference-free temperature matrix is ​​as follows: The corrected temperature field data and environmental factors were subjected to multiple regression analysis; Based on the battery load current-temperature characteristic curve, the temperature rise effect under normal working conditions is eliminated; The temperature data is further optimized to remove the interference caused by environmental noise and generate a de-interference temperature matrix.

4. The method for fault detection of a high-resolution infrared thermal imaging display according to claim 3, characterized in that: According to the temperature matrix and the physical topology of the battery pack, the heat conduction relationship between batteries is modeled through the graph neural network, and the abnormal temperature area is detected in combination with the U-Net segmentation network. The specific process of identifying abnormal areas that conform to the law of heat conduction is as follows: Based on the temperature matrix and the physical topology of the battery pack, a graph neural network model of heat conduction between batteries is established to simulate the thermal coupling relationship between the individual batteries in the battery pack, and determine the heat flow path between batteries based on the battery position and thermal conduction characteristics; Combine the battery pack topology with the temperature field data, learn the heat conduction law between batteries through the graph neural network, and optimize the spatial distribution prediction of the temperature field; Based on the U-Net segmentation network, abnormal areas in the temperature field are identified. Combined with the heat conduction law between batteries, areas with abnormal temperatures and drastic fluctuations on the heat conduction path are preferentially identified, especially areas with sudden temperature differences between adjacent batteries. Through the detection results of the temperature abnormality area, the temperature abnormality area is output and the specific location of the area is calibrated.

5. A method for fault detection of a high-resolution infrared thermal imaging display according to claim 4, characterized in that: The specific process of generating abnormal area masks and temperature spatial features under spatial structure constraints is as follows: Based on the output of the U-Net segmentation network, the spatial structure mask of the temperature abnormality area is generated by combining it with the topological map of the battery pack. The mask marks all abnormal areas where the temperature exceeds the threshold, and further eliminates noise through image post-processing; Extract the spatial features of the abnormal region, including area, shape factor, and boundary complexity, to describe the geometric characteristics of the abnormal region; Based on the spatial distribution characteristics of the temperature field, the spatial temperature characteristics related to the fault are further extracted, including the temperature gradient and the temperature fluctuation amplitude.

6. A method for fault detection of a high-resolution infrared thermal imaging array according to claim 5, characterized in that: The dynamic time warping algorithm is used to quantify the correlation between temperature and electrical parameters and distinguish between sustained faults and transient disturbances as follows: Extract the time series data of temperature fluctuation from the abnormal area mask, and obtain the temperature fluctuation characteristics of each abnormal area within a certain time interval by performing time series analysis on the dynamic changes of the temperature abnormal area; Synchronously collect the electrical parameters of the battery pack, including real-time current and voltage data, and align the electrical parameters with the temperature fluctuation timing data to ensure the timing consistency of the temperature data and the electrical data; The dynamic time warping algorithm is used to calculate the time series correlation coefficient between temperature fluctuation data and electrical parameters, and to quantify the time delay relationship between temperature fluctuation and electrical parameters. According to the analysis results of the dynamic time warping algorithm and the relationship between temperature fluctuations and electrical parameters, the fault type is determined. If the temperature fluctuations and electrical parameter changes are synchronized and continuous, it is determined to be a continuous fault. If the temperature fluctuations and electrical parameter changes are short-lived and irregular, it is determined to be a fluctuation caused by instantaneous interference.

7. A high-resolution infrared thermal imaging display fault detection method according to claim 6, characterized in that: The specific process of outputting the fault detection index and evolution mode label is as follows: Generate a fault detection index based on the calculated time series correlation coefficient of temperature fluctuation data and electrical parameters combined with the operating status of the battery pack; According to the timing characteristics of temperature fluctuations and the changing trend of electrical data, the fault evolution pattern label is generated through the evolution pattern classification algorithm, and the fault is divided into continuous or intermittent type.

8. The method for fault detection of a high-resolution infrared thermal imaging array according to claim 7, characterized in that: The specific process of classifying battery fault types through small sample transfer learning based on abnormal area masks, fault detection indexes, and evolution mode labels is as follows: The abnormal region mask, fault detection index and evolution mode label are used as input features to form a multi-dimensional fault feature vector; Through the small sample transfer learning method, the trained network model is migrated and fine-tuned for specific battery types. Through fine-tuning, the characteristics of different battery failures are mapped to corresponding category labels.

9. A method for fault detection of a high-resolution infrared thermal imaging array according to claim 8, characterized in that: The specific process of outputting the fault type and its confidence level by using the pre-trained feature extractor, fine-tuning the classifier with small sample battery failure data, and weighted fusion of temperature space features and evolution mode labels with the attention mechanism is as follows: The pre-trained feature extractor is used to extract features of the abnormal region mask, fault detection index and evolution mode label to obtain a high-dimensional feature vector containing temperature space features and evolution mode labels. Combined with small sample battery failure data, the classifier is fine-tuned and the network parameters are adjusted to make the classifier adapt to the feature mapping of the target battery failure. The attention mechanism is used to perform weighted fusion of temperature space features and evolution mode labels to enhance the contribution of key features in classification. The fine-tuned classifier is used to classify the fault type through the weighted fused feature vector, and the fault confidence is calculated based on the classification result, and finally the battery fault type and its confidence are output.

Images