Method for thermal fault detection of energy storage devices

Abstract

Embodiments of the present application provide a kind of energy storage equipment thermal fault detection method.The method first obtains the target image of the energy storage equipment to be detected, then the target image is input into thermal fault detection model, obtains the thermal fault information of energy storage equipment, wherein, thermal fault detection model is determined based on the image of multiple energy storage equipment and the corresponding labeled thermal fault information of each image, and thermal fault information includes: fault location and fault type.Application of the technical solution of the present application, based on the image of multiple energy storage equipment and its labeled thermal fault information, thermal fault detection model is trained, and the target image of the energy storage equipment to be detected is analyzed by applying the thermal fault detection model, and the thermal fault of energy storage equipment can be automatically identified based on the thermal distribution characteristics in the image, which helps to improve the detection rate of early micro-thermal anomaly, enhance the accuracy of thermal fault detection and fault positioning ability under complex working conditions.

Description

Technical Field

[0001] This application relates to the fields of deep learning and thermal fault detection technology for energy storage devices, and in particular to a method for thermal fault detection of energy storage devices. Background Technology

[0002] As a key support for building new power systems and promoting the high-proportion consumption of new energy sources, the safe and stable operation of new energy storage devices is directly related to the power grid's reliability and power quality. With the rapid development of new energy sources, the operating conditions of energy storage devices are becoming increasingly complex. Thermal failures caused by internal short circuits, poor contact, overload, and overheating have become major hidden dangers threatening the safety of energy storage systems.

[0003] Existing thermal fault detection for energy storage equipment mainly relies on regular manual inspections and monitoring by fixed-point temperature sensors. For example, manual inspections mainly depend on technicians periodically checking the surface temperature and operating status of the equipment; while fixed-point temperature sensors mainly monitor the temperature at key locations through discretely arranged sensors.

[0004] However, traditional manual inspection and fixed sensor monitoring methods have problems such as slow response, high missed detection rate, and inability to cover hidden thermal anomalies, resulting in low accuracy of thermal fault detection. Summary of the Invention

[0005] This application provides a method for detecting thermal faults in energy storage devices to improve the effectiveness of thermal fault detection.

[0006] In a first aspect, embodiments of this application provide a method for detecting thermal faults in an energy storage device, including:

[0007] Acquire a target image of the energy storage device to be tested;

[0008] The target image is input into the thermal fault detection model to obtain the thermal fault information of the energy storage device to be detected. The thermal fault detection model is determined based on images of multiple energy storage devices and the labeled thermal fault information corresponding to each image. The thermal fault information includes: fault location and fault type.

[0009] In one or more embodiments, before inputting the target image into a thermal fault detection model to obtain thermal fault information of the energy storage device to be detected, the method further includes:

[0010] Acquire images of the multiple energy storage devices and the corresponding labeled thermal fault information for each image;

[0011] The initial thermal fault detection model is trained based on the images of the multiple energy storage devices and the labeled thermal fault information corresponding to each image to obtain the thermal fault detection model.

[0012] In one or more embodiments, the initial thermal fault detection model includes: a feature extraction unit, a feature enhancement unit, and a detection head unit;

[0013] Accordingly, the step of training the initial thermal fault detection model based on the images of the multiple energy storage devices and the labeled thermal fault information corresponding to each image to obtain the thermal fault detection model includes:

[0014] S1, For each energy storage device, the image of the energy storage device is input into the feature extraction unit to obtain multiple initial feature maps;

[0015] S2, input the multiple initial feature maps into the feature enhancement unit to obtain multiple enhanced feature maps;

[0016] S3, input the multiple enhanced feature maps into the detection head unit to obtain predicted thermal fault information;

[0017] S4. Update the parameters of the initial thermal fault detection model based on the predicted thermal fault information and the labeled thermal fault information corresponding to each image. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the thermal fault detection model.

[0018] In one or more embodiments, the feature extraction unit includes: multiple initial feature extraction units, multiple multi-level energy storage device feature extraction units, and a fine energy storage device feature sensing unit;

[0019] Accordingly, for each energy storage device, the image of the energy storage device is input into the feature extraction unit to obtain multiple initial feature maps, including:

[0020] For each energy storage device image, the image of the energy storage device is sequentially input into the first initial feature extraction unit and the second initial feature extraction unit to obtain the first initial feature map;

[0021] The first initial feature map is input into the feature extraction unit of the first multi-level energy storage device to obtain the second initial feature map;

[0022] The second initial feature map is sequentially input into the third initial feature extraction unit and the second multi-level energy storage device feature extraction unit to obtain the third initial feature map;

[0023] The third initial feature map is sequentially input into the fourth initial feature extraction unit and the third multi-level energy storage device feature extraction unit to obtain the fourth initial feature map;

[0024] The fourth initial feature map is sequentially input into the fifth initial feature extraction unit and the fourth multi-level energy storage device feature extraction unit to obtain the fifth initial feature map;

[0025] The fifth initial feature map is input into the feature sensing unit of the fine energy storage device to obtain the sixth initial feature map.

[0026] In one or more embodiments, each multi-level energy storage device feature extraction unit includes: multiple initial feature extraction units, a first slicing unit, multiple depthwise separable convolution units, and a dual-path energy storage feature extraction subunit;

[0027] Accordingly, the step of inputting the first initial feature map into the first multi-level energy storage device feature extraction unit to obtain the second initial feature map includes:

[0028] The first initial feature map is sequentially input into the sixth initial feature extraction unit and the first cutting unit to obtain the seventh initial feature map, the eighth initial feature map, and the ninth initial feature map;

[0029] The seventh initial feature map is input into the first depthwise separable convolutional unit to obtain the tenth initial feature map;

[0030] The eighth initial feature map is input into the second depthwise separable convolutional unit to obtain the eleventh initial feature map;

[0031] The ninth initial feature map is input into the third depthwise separable convolutional unit to obtain the twelfth initial feature map;

[0032] The seventh initial feature map, the eighth initial feature map, and the ninth initial feature map are concatenated to obtain the thirteenth initial feature map;

[0033] The thirteenth initial feature map is sequentially input into the fourth, fifth, and sixth depthwise separable convolutional units to obtain the fourteenth initial feature map.

[0034] The fourteenth initial feature map and the thirteenth initial feature map are superimposed to obtain the fifteenth initial feature map;

[0035] The fifteenth initial feature map is input into the seventh initial feature extraction unit to obtain the sixteenth initial feature map;

[0036] The sixteenth initial feature map is input into the dual-path energy storage feature extraction subunit to obtain the second initial feature map.

[0037] In one or more embodiments, the dual-path energy storage feature extraction subunit includes: multiple initial feature extraction units, a second segmentation unit, a first convolutional block attention unit, and a second convolutional block attention unit;

[0038] Accordingly, inputting the sixteenth initial feature map into the dual-path energy storage feature extraction subunit to obtain the second initial feature map includes:

[0039] The sixteenth initial feature map is input into the eighth initial feature extraction unit and the ninth initial feature extraction unit respectively to obtain the first dual-path energy storage feature map and the second dual-path energy storage feature map;

[0040] The first dual-channel energy storage feature map and the second dual-channel energy storage feature map are superimposed to obtain the third dual-channel energy storage feature map.

[0041] The third dual-path energy storage feature map is input into the tenth initial feature extraction unit to obtain the fourth dual-path energy storage feature map;

[0042] The first dual-channel energy storage feature map, the second dual-channel energy storage feature map, and the fourth dual-channel energy storage feature map are spliced ​​together to obtain the fifth dual-channel energy storage feature map.

[0043] The fifth dual-channel energy storage feature map is input into the second cutting unit to obtain the sixth dual-channel energy storage feature map and the seventh dual-channel energy storage feature map;

[0044] The sixth dual-path energy storage feature map and the seventh dual-path energy storage feature map are respectively input into the first convolutional block attention unit and the second convolutional block attention unit to obtain the eighth dual-path energy storage feature map and the ninth dual-path energy storage feature map;

[0045] The eighth dual-channel energy storage feature map and the ninth dual-channel energy storage feature map are spliced ​​together to obtain the tenth dual-channel energy storage feature map.

[0046] The tenth dual-path energy storage feature map is input into the eleventh initial feature extraction unit to obtain the second initial feature map.

[0047] In one or more embodiments, the fine energy storage device feature sensing unit includes: a first max pooling unit, a first average pooling unit, a first activation unit, a third slicing unit, multiple initial feature extraction units, and multiple depthwise separable convolution units;

[0048] Accordingly, the step of inputting the fifth initial feature map into the feature sensing unit of the fine energy storage device to obtain the sixth initial feature map includes:

[0049] The fifth initial feature map is input into the first max pooling unit, the seventh depthwise separable convolution unit, the eighth depthwise separable convolution unit, and the first average pooling unit, respectively, to obtain the first fine energy storage feature map, the second fine energy storage feature map, the third fine energy storage feature map, and the fourth fine energy storage feature map.

[0050] The second and third refined energy storage feature maps are input into the twelfth and thirteenth initial feature extraction units, respectively, to obtain the fifth and sixth refined energy storage feature maps.

[0051] The fifth and sixth fine energy storage feature maps are stitched together to obtain the seventh fine energy storage feature map.

[0052] The seventh fine energy storage feature map is sequentially input into the fourteenth initial feature extraction unit and the first activation unit to obtain the first fine energy storage weight map;

[0053] The first fine energy storage weight map is input into the third cutting unit to obtain the second fine energy storage weight map and the third fine energy storage weight map;

[0054] The fifth fine energy storage feature map and the second fine energy storage weight map are multiplied element by element to obtain the eighth fine energy storage feature map;

[0055] The sixth fine energy storage feature map and the third fine energy storage weight map are multiplied element by element to obtain the ninth fine energy storage feature map.

[0056] The first fine energy storage feature map, the fourth fine energy storage feature map, the eighth fine energy storage feature map, and the ninth fine energy storage feature map are stitched together to obtain the tenth fine energy storage feature map;

[0057] The tenth fine energy storage feature map is input into the fifteenth initial feature extraction unit to obtain the sixth initial feature map.

[0058] In one or more embodiments, the feature enhancement unit includes: a multi-path feature convolution unit, a first upsampling unit, a second upsampling unit, multiple initial feature extraction units, and multiple energy storage device feature filtering units;

[0059] Accordingly, the step of inputting the plurality of initial feature maps into the feature enhancement unit to obtain a plurality of enhanced feature maps includes:

[0060] The sixth initial feature map is input into the multi-path feature convolution unit to obtain the first enhanced feature map;

[0061] The first enhanced feature map is input into the first upsampling unit to obtain the second enhanced feature map;

[0062] The fourth initial feature map and the second enhanced feature map are concatenated to obtain the third enhanced feature map.

[0063] The third enhanced feature map is input into the feature filtering unit of the first energy storage device to obtain the fourth enhanced feature map;

[0064] The fourth enhanced feature map is input into the second upsampling unit to obtain the fifth enhanced feature map;

[0065] The third initial feature map and the fifth enhanced feature map are concatenated to obtain the sixth enhanced feature map;

[0066] The sixth enhanced feature map is input into the second energy storage device feature filtering unit to obtain the seventh enhanced feature map;

[0067] The seventh enhanced feature map is input into the sixteenth initial feature extraction unit to obtain the eighth enhanced feature map;

[0068] The fourth and eighth enhanced feature maps are concatenated to obtain the ninth enhanced feature map.

[0069] The ninth enhanced feature map is input into the feature filtering unit of the third energy storage device to obtain the tenth enhanced feature map;

[0070] The tenth enhanced feature map is input into the seventeenth initial feature extraction unit to obtain the eleventh enhanced feature map;

[0071] The first enhanced feature map and the eleventh enhanced feature map are concatenated to obtain the twelfth enhanced feature map.

[0072] The twelfth enhanced feature map is input into the fourth energy storage device feature filtering unit to obtain the thirteenth enhanced feature map.

[0073] In one or more embodiments, each energy storage device feature filtering unit includes: a feature coordination subunit, a ninth depthwise separable convolution unit, a batch normalization unit, a second activation unit, a fourth slicing unit, and a plurality of initial feature extraction units;

[0074] Accordingly, inputting the third enhanced feature map into the first energy storage device feature filtering unit to obtain the fourth enhanced feature map includes:

[0075] The third enhanced feature map is input into the eighteenth initial feature extraction unit and the nineteenth initial feature extraction unit respectively to obtain the first filtered feature map and the second filtered feature map;

[0076] The first and second filtered feature maps are superimposed to obtain the third filtered feature map.

[0077] The third filtered feature map is sequentially input into the ninth depthwise separable convolutional unit, the batch normalization unit, and the second activation unit to obtain the first filtered weight map;

[0078] The first filter weight map is input into the fourth cutting unit to obtain the second filter weight map and the third filter weight map;

[0079] The first filter feature map and the second filter weight map are multiplied element by element to obtain the fourth filter feature map;

[0080] The second filter feature map and the third filter weight map are multiplied element by element to obtain the fifth filter feature map;

[0081] The fourth and fifth filter feature maps are multiplied element-wise to obtain the sixth filter feature map.

[0082] The fourth and fifth filter feature maps are added element by element to obtain the seventh filter feature map;

[0083] The third enhanced feature map is input into the feature collaboration subunit to obtain the eighth filtered feature map;

[0084] The third enhanced feature map, the sixth filtered feature map, the seventh filtered feature map, and the eighth filtered feature map are concatenated to obtain the ninth filtered feature map.

[0085] The ninth filtered feature map is input into the twentieth initial feature extraction unit to obtain the fourth enhanced feature map.

[0086] In one or more embodiments, the feature collaboration subunit includes: a second average pooling unit, a third average pooling unit, a second max pooling unit, a third max pooling unit, and a plurality of initial feature extraction units;

[0087] Accordingly, the step of inputting the third enhanced feature map into the feature co-operation subunit to obtain the eighth filtered feature map includes:

[0088] The third enhanced feature map is input into the twenty-first initial feature extraction unit to obtain the first collaborative feature map;

[0089] The first collaborative feature map is sequentially input into the second average pooling unit and the third average pooling unit to obtain the second collaborative feature map;

[0090] The first collaborative feature map and the second collaborative feature map are concatenated to obtain the third collaborative feature map.

[0091] The third collaborative feature map is sequentially input into the second max pooling unit and the third max pooling unit to obtain the fourth collaborative feature map;

[0092] The third and fourth collaborative feature maps are concatenated to obtain the fifth collaborative feature map.

[0093] The fifth collaborative feature map is input into the twenty-second initial feature extraction unit to obtain the sixth collaborative feature map;

[0094] The first collaborative feature map and the sixth collaborative feature map are superimposed to obtain the seventh collaborative feature map.

[0095] The seventh collaborative feature map is input into the twenty-third initial feature extraction unit to obtain the eighth filtered feature map.

[0096] Secondly, embodiments of this application provide a thermal fault detection device for an energy storage device, comprising:

[0097] The acquisition module is used to acquire the target image of the energy storage device to be tested;

[0098] The processing module is used to input the target image into the thermal fault detection model to obtain the thermal fault information of the energy storage device to be detected. The thermal fault detection model is determined based on images of multiple energy storage devices and the labeled thermal fault information corresponding to each image. The thermal fault information includes: fault location and fault type.

[0099] In one or more embodiments, before inputting the target image into the thermal fault detection model to obtain the thermal fault information of the energy storage device to be detected, the processing module is further configured to:

[0100] Acquire images of the multiple energy storage devices and the corresponding labeled thermal fault information for each image;

[0101] The initial thermal fault detection model is trained based on the images of the multiple energy storage devices and the labeled thermal fault information corresponding to each image to obtain the thermal fault detection model.

[0102] In one or more embodiments, the initial thermal fault detection model includes: a feature extraction unit, a feature enhancement unit, and a detection head unit;

[0103] Accordingly, the processing module trains an initial thermal fault detection model based on the images of the multiple energy storage devices and the labeled thermal fault information corresponding to each image, to obtain the thermal fault detection model, specifically used for:

[0104] S1, For each energy storage device, the image of the energy storage device is input into the feature extraction unit to obtain multiple initial feature maps;

[0105] S2, input the multiple initial feature maps into the feature enhancement unit to obtain multiple enhanced feature maps;

[0106] S3, input the multiple enhanced feature maps into the detection head unit to obtain predicted thermal fault information;

[0107] S4. Update the parameters of the initial thermal fault detection model based on the predicted thermal fault information and the labeled thermal fault information corresponding to each image. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the thermal fault detection model.

[0108] In one or more embodiments, the feature extraction unit includes: multiple initial feature extraction units, multiple multi-level energy storage device feature extraction units, and a fine energy storage device feature sensing unit;

[0109] Accordingly, the processing module, for each energy storage device's image, inputs the image of the energy storage device into the feature extraction unit to obtain multiple initial feature maps, specifically used for:

[0110] For each energy storage device image, the image of the energy storage device is sequentially input into the first initial feature extraction unit and the second initial feature extraction unit to obtain the first initial feature map;

[0111] The first initial feature map is input into the feature extraction unit of the first multi-level energy storage device to obtain the second initial feature map;

[0112] The second initial feature map is sequentially input into the third initial feature extraction unit and the second multi-level energy storage device feature extraction unit to obtain the third initial feature map;

[0113] The third initial feature map is sequentially input into the fourth initial feature extraction unit and the third multi-level energy storage device feature extraction unit to obtain the fourth initial feature map;

[0114] The fourth initial feature map is sequentially input into the fifth initial feature extraction unit and the fourth multi-level energy storage device feature extraction unit to obtain the fifth initial feature map;

[0115] The fifth initial feature map is input into the feature sensing unit of the fine energy storage device to obtain the sixth initial feature map.

[0116] In one or more embodiments, each multi-level energy storage device feature extraction unit includes: multiple initial feature extraction units, a first slicing unit, multiple depthwise separable convolution units, and a dual-path energy storage feature extraction subunit;

[0117] Accordingly, the processing module inputs the first initial feature map to the first multi-level energy storage device feature extraction unit to obtain a second initial feature map, specifically for:

[0118] The first initial feature map is sequentially input into the sixth initial feature extraction unit and the first cutting unit to obtain the seventh initial feature map, the eighth initial feature map, and the ninth initial feature map;

[0119] The seventh initial feature map is input into the first depthwise separable convolutional unit to obtain the tenth initial feature map;

[0120] The eighth initial feature map is input into the second depthwise separable convolutional unit to obtain the eleventh initial feature map;

[0121] The ninth initial feature map is input into the third depthwise separable convolutional unit to obtain the twelfth initial feature map;

[0122] The seventh initial feature map, the eighth initial feature map, and the ninth initial feature map are concatenated to obtain the thirteenth initial feature map;

[0123] The thirteenth initial feature map is sequentially input into the fourth, fifth, and sixth depthwise separable convolutional units to obtain the fourteenth initial feature map.

[0124] The fourteenth initial feature map and the thirteenth initial feature map are superimposed to obtain the fifteenth initial feature map;

[0125] The fifteenth initial feature map is input into the seventh initial feature extraction unit to obtain the sixteenth initial feature map;

[0126] The sixteenth initial feature map is input into the dual-path energy storage feature extraction subunit to obtain the second initial feature map.

[0127] In one or more embodiments, the dual-path energy storage feature extraction subunit includes: multiple initial feature extraction units, a second segmentation unit, a first convolutional block attention unit, and a second convolutional block attention unit;

[0128] Accordingly, the processing module inputs the sixteenth initial feature map into the dual-path energy storage feature extraction subunit to obtain the second initial feature map, specifically for:

[0129] The sixteenth initial feature map is input into the eighth initial feature extraction unit and the ninth initial feature extraction unit respectively to obtain the first dual-path energy storage feature map and the second dual-path energy storage feature map;

[0130] The first dual-channel energy storage feature map and the second dual-channel energy storage feature map are superimposed to obtain the third dual-channel energy storage feature map.

[0131] The third dual-path energy storage feature map is input into the tenth initial feature extraction unit to obtain the fourth dual-path energy storage feature map;

[0132] The first dual-channel energy storage feature map, the second dual-channel energy storage feature map, and the fourth dual-channel energy storage feature map are spliced ​​together to obtain the fifth dual-channel energy storage feature map.

[0133] The fifth dual-channel energy storage feature map is input into the second cutting unit to obtain the sixth dual-channel energy storage feature map and the seventh dual-channel energy storage feature map;

[0134] The sixth dual-path energy storage feature map and the seventh dual-path energy storage feature map are respectively input into the first convolutional block attention unit and the second convolutional block attention unit to obtain the eighth dual-path energy storage feature map and the ninth dual-path energy storage feature map;

[0135] The eighth dual-channel energy storage feature map and the ninth dual-channel energy storage feature map are spliced ​​together to obtain the tenth dual-channel energy storage feature map.

[0136] The tenth dual-path energy storage feature map is input into the eleventh initial feature extraction unit to obtain the second initial feature map.

[0137] In one or more embodiments, the fine energy storage device feature sensing unit includes: a first max pooling unit, a first average pooling unit, a first activation unit, a third slicing unit, multiple initial feature extraction units, and multiple depthwise separable convolution units;

[0138] Accordingly, the processing module inputs the fifth initial feature map to the feature sensing unit of the fine energy storage device to obtain a sixth initial feature map, specifically used for:

[0139] The fifth initial feature map is input into the first max pooling unit, the seventh depthwise separable convolution unit, the eighth depthwise separable convolution unit, and the first average pooling unit, respectively, to obtain the first fine energy storage feature map, the second fine energy storage feature map, the third fine energy storage feature map, and the fourth fine energy storage feature map.

[0140] The second and third refined energy storage feature maps are input into the twelfth and thirteenth initial feature extraction units, respectively, to obtain the fifth and sixth refined energy storage feature maps.

[0141] The fifth and sixth fine energy storage feature maps are stitched together to obtain the seventh fine energy storage feature map.

[0142] The seventh fine energy storage feature map is sequentially input into the fourteenth initial feature extraction unit and the first activation unit to obtain the first fine energy storage weight map;

[0143] The first fine energy storage weight map is input into the third cutting unit to obtain the second fine energy storage weight map and the third fine energy storage weight map;

[0144] The fifth fine energy storage feature map and the second fine energy storage weight map are multiplied element by element to obtain the eighth fine energy storage feature map;

[0145] The sixth fine energy storage feature map and the third fine energy storage weight map are multiplied element by element to obtain the ninth fine energy storage feature map.

[0146] The first fine energy storage feature map, the fourth fine energy storage feature map, the eighth fine energy storage feature map, and the ninth fine energy storage feature map are stitched together to obtain the tenth fine energy storage feature map;

[0147] The tenth fine energy storage feature map is input into the fifteenth initial feature extraction unit to obtain the sixth initial feature map.

[0148] In one or more embodiments, the feature enhancement unit includes: a multi-path feature convolution unit, a first upsampling unit, a second upsampling unit, multiple initial feature extraction units, and multiple energy storage device feature filtering units;

[0149] Accordingly, the processing module inputs the plurality of initial feature maps into the feature enhancement unit to obtain a plurality of enhanced feature maps, specifically for:

[0150] The sixth initial feature map is input into the multi-path feature convolution unit to obtain the first enhanced feature map;

[0151] The first enhanced feature map is input into the first upsampling unit to obtain the second enhanced feature map;

[0152] The fourth initial feature map and the second enhanced feature map are concatenated to obtain the third enhanced feature map.

[0153] The third enhanced feature map is input into the feature filtering unit of the first energy storage device to obtain the fourth enhanced feature map;

[0154] The fourth enhanced feature map is input into the second upsampling unit to obtain the fifth enhanced feature map;

[0155] The third initial feature map and the fifth enhanced feature map are concatenated to obtain the sixth enhanced feature map;

[0156] The sixth enhanced feature map is input into the second energy storage device feature filtering unit to obtain the seventh enhanced feature map;

[0157] The seventh enhanced feature map is input into the sixteenth initial feature extraction unit to obtain the eighth enhanced feature map;

[0158] The fourth and eighth enhanced feature maps are concatenated to obtain the ninth enhanced feature map.

[0159] The ninth enhanced feature map is input into the feature filtering unit of the third energy storage device to obtain the tenth enhanced feature map;

[0160] The tenth enhanced feature map is input into the seventeenth initial feature extraction unit to obtain the eleventh enhanced feature map;

[0161] The first enhanced feature map and the eleventh enhanced feature map are concatenated to obtain the twelfth enhanced feature map.

[0162] The twelfth enhanced feature map is input into the fourth energy storage device feature filtering unit to obtain the thirteenth enhanced feature map.

[0163] In one or more embodiments, each energy storage device feature filtering unit includes: a feature coordination subunit, a ninth depthwise separable convolution unit, a batch normalization unit, a second activation unit, a fourth slicing unit, and a plurality of initial feature extraction units;

[0164] Accordingly, the processing module inputs the third enhanced feature map to the first energy storage device feature filtering unit to obtain the fourth enhanced feature map, specifically for:

[0165] The third enhanced feature map is input into the eighteenth initial feature extraction unit and the nineteenth initial feature extraction unit respectively to obtain the first filtered feature map and the second filtered feature map;

[0166] The first and second filtered feature maps are superimposed to obtain the third filtered feature map.

[0167] The third filtered feature map is sequentially input into the ninth depthwise separable convolutional unit, the batch normalization unit, and the second activation unit to obtain the first filtered weight map;

[0168] The first filter weight map is input into the fourth cutting unit to obtain the second filter weight map and the third filter weight map;

[0169] The first filter feature map and the second filter weight map are multiplied element by element to obtain the fourth filter feature map;

[0170] The second filter feature map and the third filter weight map are multiplied element by element to obtain the fifth filter feature map;

[0171] The fourth and fifth filter feature maps are multiplied element-wise to obtain the sixth filter feature map.

[0172] The fourth and fifth filter feature maps are added element by element to obtain the seventh filter feature map;

[0173] The third enhanced feature map is input into the feature collaboration subunit to obtain the eighth filtered feature map;

[0174] The third enhanced feature map, the sixth filtered feature map, the seventh filtered feature map, and the eighth filtered feature map are concatenated to obtain the ninth filtered feature map.

[0175] The ninth filtered feature map is input into the twentieth initial feature extraction unit to obtain the fourth enhanced feature map.

[0176] In one or more embodiments, the feature collaboration subunit includes: a second average pooling unit, a third average pooling unit, a second max pooling unit, a third max pooling unit, and a plurality of initial feature extraction units;

[0177] Accordingly, the processing module inputs the third enhanced feature map into the feature collaboration subunit to obtain the eighth filtered feature map, specifically for:

[0178] The third enhanced feature map is input into the twenty-first initial feature extraction unit to obtain the first collaborative feature map;

[0179] The first collaborative feature map is sequentially input into the second average pooling unit and the third average pooling unit to obtain the second collaborative feature map;

[0180] The first collaborative feature map and the second collaborative feature map are concatenated to obtain the third collaborative feature map.

[0181] The third collaborative feature map is sequentially input into the second max pooling unit and the third max pooling unit to obtain the fourth collaborative feature map;

[0182] The third and fourth collaborative feature maps are concatenated to obtain the fifth collaborative feature map.

[0183] The fifth collaborative feature map is input into the twenty-second initial feature extraction unit to obtain the sixth collaborative feature map;

[0184] The first collaborative feature map and the sixth collaborative feature map are superimposed to obtain the seventh collaborative feature map.

[0185] The seventh collaborative feature map is input into the twenty-third initial feature extraction unit to obtain the eighth filtered feature map.

[0186] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0187] The memory stores computer-executed instructions;

[0188] The processor executes computer execution instructions stored in the memory, such that the processor, when executed, is used to implement the method described in the first aspect and any of the embodiments above.

[0189] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods described in the first aspect and any of the embodiments above.

[0190] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, is used to implement a thermal fault detection method for an energy storage device as described in the first aspect and various possible implementations of the first aspect.

[0191] This application provides a method for detecting thermal faults in energy storage devices. The method first acquires a target image of the energy storage device to be detected, then inputs the target image into a thermal fault detection model to obtain thermal fault information of the energy storage device. The thermal fault detection model is determined based on images of multiple energy storage devices and corresponding labeled thermal fault information for each image. The thermal fault information includes the fault location and fault type. By acquiring a target image of the energy storage device to be detected and inputting it into a thermal fault detection model determined based on multiple energy storage device images and their corresponding labeled thermal fault information, the method can accurately identify the fault location and fault type of the energy storage device from the target image. This improves the accuracy of early thermal fault detection of energy storage devices under complex environments and dynamic operating conditions, and reduces the risk of missed or false detections. Attached Figure Description

[0192] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0193] Figure 1 A flowchart illustrating the thermal fault detection method for energy storage devices provided in this application embodiment. Figure 1 ;

[0194] Figure 2 A flowchart illustrating the thermal fault detection method for energy storage devices provided in this application embodiment. Figure 2 ;

[0195] Figure 3 This is a schematic diagram of the structure of the initial thermal fault detection model provided in the embodiments of this application;

[0196] Figure 4 A flowchart illustrating the thermal fault detection method for energy storage devices provided in this application embodiment. Figure 3 ;

[0197] Figure 5 This is a schematic diagram of the structure of the feature extraction unit provided in the embodiments of this application;

[0198] Figure 6 This is a schematic diagram of the structure of the feature extraction unit for a multi-level energy storage device provided in an embodiment of this application;

[0199] Figure 7 This is a schematic diagram of the structure of the dual-path energy storage feature extraction subunit provided in the embodiments of this application;

[0200] Figure 8 A schematic diagram of the structure of the feature sensing unit for a fine energy storage device provided in an embodiment of this application;

[0201] Figure 9 This is a schematic diagram of the structure of the feature enhancement unit provided in an embodiment of this application;

[0202] Figure 10 This is a schematic diagram of the structure of the energy storage device feature filtering unit provided in the embodiments of this application;

[0203] Figure 11 This is a schematic diagram of the structure of the feature cooperative subunit provided in the embodiments of this application;

[0204] Figure 12 This is a schematic diagram of the structure of the detection head unit provided in the embodiments of this application;

[0205] Figure 13 This is a schematic diagram of the structure of the thermal fault detection model provided in the embodiments of this application;

[0206] Figure 14 A schematic diagram of the structure of the defect detection device for energy storage equipment provided in the embodiments of this application;

[0207] Figure 15 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0208] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0209] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0210] Before introducing the embodiments of this application, the application background of the embodiments of this application will be explained first:

[0211] As a key support for building new power systems and promoting the high-proportion consumption of new energy sources, the safe and stable operation of new energy storage devices is directly related to the power grid's reliability and power quality. With the rapid development of new energy sources, the operating conditions of energy storage devices are becoming increasingly complex. Thermal failures caused by internal short circuits, poor contact, overload, and overheating have become major hidden dangers threatening the safety of energy storage systems.

[0212] Existing methods for detecting thermal faults in energy storage equipment mainly rely on regular manual inspections and monitoring by fixed-point temperature sensors. For example, manual inspections depend on technicians periodically checking the surface temperature and operating status of the equipment; while fixed-point temperature sensors mainly monitor the temperature at key locations through discretely arranged sensors.

[0213] However, traditional manual inspection and fixed sensor monitoring methods have problems such as slow response, high missed detection rate, and inability to cover hidden thermal anomalies, resulting in low accuracy of thermal fault detection.

[0214] The thermal fault detection method for energy storage devices provided in this application aims to solve the aforementioned technical problems of the prior art. The technical concept of this application is as follows: Existing methods suffer from low accuracy in detecting thermal faults in energy storage devices. Thermal fault detection of energy storage devices mainly involves analyzing their appearance or temperature data to identify abnormal areas. If image detection of the energy storage device to be detected can be performed and analyzed in depth, suppressing interfering hot spots caused by environmental reflection and background radiation in the image, and accurately identifying early, minor thermal anomaly features, accurate thermal fault detection can be achieved. Therefore, this application's embodiments consider a thermal fault detection model determined based on multiple images of energy storage devices and their corresponding labeled thermal fault information. This model is applied to the target image of the energy storage device to be detected for thermal fault detection, directly outputting the fault location and fault type of the energy storage device to be detected. This achieves accurate detection of early thermal faults in energy storage devices under complex environments and dynamic operating conditions, reducing the risk of missed or false detections.

[0215] The execution subject of this application embodiment is an electronic device, which can be a terminal device, such as a laptop, desktop computer, or tablet computer, or a server. In practical applications, whether the electronic device is a terminal device or a server can be determined according to the actual situation, and no specific limitation is imposed on it.

[0216] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0217] Figure 1 A flowchart illustrating the thermal fault detection method for energy storage devices provided in this application embodiment. Figure 1 .like Figure 1 As shown, the thermal fault detection method for this energy storage device includes the following steps:

[0218] S101. Obtain the target image of the energy storage device to be tested.

[0219] In this step, in order to perform thermal fault detection on the energy storage device under test, a target image of the energy storage device under test can be acquired first, providing a data basis for subsequent thermal fault detection.

[0220] For example, the energy storage device being tested can be a battery cluster, battery module, cell housing, combiner unit, energy storage converter, or battery cabinet and its associated components in a distributed energy storage device.

[0221] In one possible implementation, the target image is an image that includes thermal state information of the energy storage device to be detected; for example, the target image could be an infrared image.

[0222] Using professional shooting equipment, such as high-definition industrial cameras and energy storage inspection robots equipped with infrared sensing modules, the system can take pictures of the energy storage equipment on site, collect target images of the energy storage equipment to be inspected, and crop the target images according to preset sizes.

[0223] S102. Input the target image into the thermal fault detection model to obtain the thermal fault information of the energy storage device to be detected;

[0224] The thermal fault detection model is determined based on images of multiple energy storage devices and the corresponding labeled thermal fault information for each image. The thermal fault information includes: fault location and fault type.

[0225] In this step, a thermal fault detection model is determined based on images of multiple energy storage devices and the labeled thermal fault information corresponding to each image. The target image is then input into the thermal fault detection model for thermal fault analysis to obtain thermal fault information including fault location and fault type.

[0226] For example, the images of multiple energy storage devices can be sample images obtained by taking pictures of different energy storage devices under different environmental conditions (such as normal temperature, high temperature, low temperature, high humidity, etc.), different shooting angles (such as overhead shots, side shots, close-up detail shots, etc.), and different device operating states (such as no-load, full-load, charge-discharge switching, etc.).

[0227] The labeled thermal fault information for each image can be generated by manual annotation tools or other annotation methods. The annotation content includes the location identification information corresponding to the thermal fault and the fault type label.

[0228] Fault location is used to indicate the spatial location of a thermal fault in an image of an energy storage device.

[0229] The fault type is used to indicate the category of thermal fault, such as abnormal overheating, local high temperature, overall overheating, etc., or further subdivided into cell hot spots, module thermal imbalance, connection part overheating, and energy storage converter overheating, etc.

[0230] In one possible implementation, the thermal fault detection model can be a deep learning model that predetermines images of multiple energy storage devices and corresponding labeled thermal fault information. The target image of the energy storage device to be detected is input into the thermal fault detection model for deep learning feature analysis, and the output is a detection image of the energy storage device to be detected, including a detection box of the heating area. The coordinate position of the detection box of the heating area is the fault location, and the fault type is marked on the detection box of the heating area.

[0231] In addition, the overlap between the detection frame of the energy storage device area to be tested and the detection frame of the heat generation area will be measured. As a criterion for determining thermal failures in energy storage devices, its calculation formula is as follows:

[0232]

[0233] Among them, the detection box for the energy storage device under test is used to indicate the area where the energy storage device under test is located, and the detection box for the overheating area is used to indicate the abnormal overheating area.

[0234] When overlap If the temperature is below the preset threshold, it can be determined that the energy storage device under test has not experienced abnormal heating; otherwise, it can be determined that the energy storage device under test has experienced abnormal heating.

[0235] The thermal fault detection method for energy storage devices provided in this application first acquires a target image of the energy storage device to be detected, and then inputs the target image into a thermal fault detection model to obtain thermal fault information of the energy storage device. The thermal fault detection model is determined based on images of multiple energy storage devices and the corresponding labeled thermal fault information for each image. The thermal fault information includes the fault location and fault type. In this embodiment, by acquiring a target image of the energy storage device to be detected and inputting it into a thermal fault detection model determined based on multiple energy storage device images and their corresponding labeled thermal fault information, the fault location and fault type of the energy storage device can be accurately identified from the target image. This improves the accuracy of early thermal fault detection of energy storage devices under complex environments and dynamic operating conditions, and reduces the risk of missed or false detections.

[0236] Based on the above embodiments, Figure 2 A flowchart illustrating the thermal fault detection method for energy storage devices provided in this application embodiment. Figure 2 .like Figure 2 As shown, prior to S120 above, the thermal fault detection method for this energy storage device further includes the following steps:

[0237] S201. Obtain images of multiple energy storage devices and the corresponding labeled thermal fault information for each image.

[0238] In this step, images of multiple energy storage devices and the corresponding labeled defect information for each image are acquired, providing input samples and supervision targets for subsequent model training.

[0239] In one possible implementation, images of multiple energy storage devices can be acquired by infrared thermal imaging devices, inspection cameras, or fusion imaging terminals under different ambient temperatures, different charging and discharging conditions, and different device postures. The images of multiple energy storage devices can include thermal distribution maps of key components such as cells, battery modules, busbars, and energy storage converter housings.

[0240] The labeled thermal fault information for each image can be determined using a labeling tool. For example, first, the labeling tool is used to select the energy storage device and the abnormal heating area in each image to form a bounding box. Then, a corresponding fault type label is assigned to each bounding box to obtain an energy storage device image dataset with labeled thermal fault information. The fault type label can correspond to thermal fault attributes such as local overheating, abnormal temperature rise, poor contact, or abnormal heat diffusion, so that the labeled thermal fault information includes both location constraints and type constraints.

[0241] In addition, the image dataset of energy storage devices can be uniformly cropped to a preset size, such as 640×640 pixels, and data augmentation processing can be performed on the uniformly sized image dataset of energy storage devices to obtain an optimized energy storage device dataset, so as to improve the generalization ability of the model and increase the sample diversity of the dataset. The data augmentation processing includes: flipping, rotation transformation and noise addition.

[0242] The optimized energy storage equipment dataset is divided into a training set, a validation set, and a test set according to a preset ratio, such as 8:1:1. The training set is used to train the model, the validation set is used to verify the detection accuracy of the thermal fault detection model, and the test set is used to test the performance of the thermal fault detection model.

[0243] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0244] S202. Train the initial thermal fault detection model based on images of multiple energy storage devices and the labeled thermal fault information corresponding to each image to obtain the thermal fault detection model.

[0245] In this step, images of multiple energy storage devices are input into the initial thermal fault detection model for thermal fault identification. Predicted thermal fault information corresponding to the images of multiple energy storage devices is output. The model parameters of the initial thermal fault detection model are updated based on the predicted thermal fault information and the labeled thermal fault information corresponding to each image, thereby obtaining the thermal fault detection model.

[0246] In one possible implementation, Figure 3 This is a schematic diagram of the structure of the initial thermal fault detection model provided in an embodiment of this application. Figure 3 As shown, the initial thermal fault detection model includes: a feature extraction unit, a feature enhancement unit, and a detection head unit;

[0247] For example, the feature extraction unit is used to perform layered characterization of the image of the energy storage device, and extract thermal texture features, local hot spot features and overall thermal field distribution features related to the thermal anomaly area; the feature enhancement unit is used to perform cross-scale fusion and feature filtering enhancement processing on the extracted feature data; the detection head unit is used to convert the enhanced feature data into thermal fault information including fault location and fault type.

[0248] Accordingly, Figure 4 A flowchart illustrating the thermal fault detection method for energy storage devices provided in this application embodiment. Figure 3 .like Figure 4 As shown, one possible implementation of S202 above includes the following steps:

[0249] S1, For each energy storage device, the image of the energy storage device is input into the feature extraction unit to obtain multiple initial feature maps;

[0250] For example, for each energy storage device image, the current energy storage device image is input into the feature extraction unit for hierarchical feature extraction, extracting thermal texture features, local hotspot features and overall thermal field distribution features related to the thermal anomaly area, and outputting multiple initial feature maps.

[0251] In one possible implementation, Figure 5 This is a schematic diagram of the structure of the feature extraction unit provided in an embodiment of this application. Figure 5 As shown, the feature extraction unit includes: multiple initial feature extraction units, multiple multi-level energy storage device feature extraction (MESDF) units, and a precision energy storage device feature perception (PESDFP) unit;

[0252] The multiple initial feature extraction units include: a first initial feature extraction unit, a second initial feature extraction unit, a third initial feature extraction unit, a fourth initial feature extraction unit, and a fifth initial feature extraction unit;

[0253] The multi-level energy storage device feature extraction unit includes: a first multi-level energy storage device feature extraction unit, a second multi-level energy storage device feature extraction unit, a third multi-level energy storage device feature extraction unit, and a fourth multi-level energy storage device feature extraction unit.

[0254] In one possible implementation, the initial feature extraction unit can be a CBS (Conv+Batch Normalization+SiLU) module, which includes convolutional layers, batch normalization layers, and activation function layers.

[0255] Accordingly, one possible implementation of S1 above includes the following steps:

[0256] S11. For each energy storage device image, the energy storage device image is sequentially input into the first initial feature extraction unit and the second initial feature extraction unit to obtain the first initial feature map;

[0257] For example, the image X1 of the energy storage device is input into the first CBS module for preliminary feature extraction to obtain an initial feature map X2. Then, the initial feature map X2 is input into the second CBS module for feature extraction to obtain a first initial feature map X3.

[0258] S12. Input the first initial feature map into the first multi-level energy storage device feature extraction unit to obtain the second initial feature map;

[0259] For example, the first initial feature map X3 is input into the first MESDF unit for multi-level feature extraction to obtain the second initial feature map X4.

[0260] S13. Input the second initial feature map sequentially into the third initial feature extraction unit and the second multi-level energy storage device feature extraction unit to obtain the third initial feature map;

[0261] For example, the second initial feature map X4 is input into the third CBS module for feature extraction to obtain the initial feature map X5. Then, the initial feature map X5 is input into the second MESDF unit for multi-level feature extraction to obtain the third initial feature map X6.

[0262] S14. Input the third initial feature map into the fourth initial feature extraction unit and the third multi-level energy storage device feature extraction unit in sequence to obtain the fourth initial feature map;

[0263] For example, the third initial feature map X6 is input into the fourth CBS module for feature extraction to obtain the initial feature map X7. Then, the initial feature map X7 is input into the third MESDF unit for multi-level feature extraction to obtain the fourth initial feature map X8.

[0264] S15. Input the fourth initial feature map into the fifth initial feature extraction unit and the fourth multi-level energy storage device feature extraction unit in sequence to obtain the fifth initial feature map;

[0265] For example, the fourth initial feature map X8 is input into the fifth CBS module for feature extraction to obtain the initial feature map X9. Then, the initial feature map X9 is input into the fourth MESDF unit for multi-level feature extraction to obtain the fifth initial feature map X10.

[0266] S16. Input the fifth initial feature map into the feature sensing unit of the fine energy storage device to obtain the sixth initial feature map.

[0267] For example, the fifth initial feature map X10 is input into the PESDFP unit for refined feature extraction to obtain the sixth initial feature map X11.

[0268] For example, if the image size of an energy storage device is C×H×W, where C is the number of channels in the image, H is the height of the image, and W is the width of the image.

[0269] For the above S1, let's take an example where the image size of the input energy storage device image X1 is 3×640×640:

[0270] 1) Input the image X1 of the energy storage device into the first CBS module to obtain an initial feature map X2 with a size of 64×320×320. Then input the initial feature map X2 into the second CBS module to obtain a first initial feature map X3 with a size of 128×160×160. Input the first initial feature map X3 into the first MESDF unit to obtain a second initial feature map X4 with a size of 128×160×160. Input the second initial feature map X4 into the third CBS module to obtain an initial feature map X5 with a size of 256×80×80. Then input the initial feature map X5 into the second MESDF unit to obtain a third initial feature map X6 with a size of 256×80×80.

[0271] 2) Input the third initial feature map X6 into the fourth CBS module to obtain an initial feature map X7 with a size of 512×40×40. Then input the initial feature map X7 into the third MESDF unit to obtain a fourth initial feature map X8 with a size of 512×40×40.

[0272] 3) Input the fourth initial feature map X8 into the fifth CBS module to obtain an initial feature map X9 with a size of 512×20×20. Then input the initial feature map X9 into the fourth MESDF unit to obtain a fifth initial feature map X10 with a size of 512×20×20. Input the fifth initial feature map X10 into the PESDFP unit to obtain a sixth initial feature map X11 with a size of 512×20×20.

[0273] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0274] In one possible implementation, Figure 6 This is a schematic diagram of the structure of a feature extraction unit for a multi-level energy storage device provided in an embodiment of this application. (See attached diagram.) Figure 6 As shown, each Multi-level Energy Storage Device Feature Extraction (MESDF) unit includes: multiple initial feature extraction units, a first slicing unit, multiple depthwise separable convolution units, and a dual-path energy storage feature extraction subunit;

[0275] For example, the multi-level energy storage device feature extraction unit dynamically mines the multi-level features of the energy storage device through a progressive structure of initial feature extraction preprocessing, multi-scale depth separable convolution, feature splicing and residual fusion, so as to achieve efficient fusion of shallow features and deep semantic features.

[0276] Among them, the multiple initial feature extraction units include: a sixth initial feature extraction unit and a seventh initial feature extraction unit;

[0277] The multiple depthwise separable convolutional units include: a first depthwise separable convolutional unit, a second depthwise separable convolutional unit, a third depthwise separable convolutional unit, a fourth depthwise separable convolutional unit, a fifth depthwise separable convolutional unit, and a sixth depthwise separable convolutional unit.

[0278] Accordingly, one possible implementation of S12 above includes the following steps:

[0279] S121. Input the first initial feature map into the sixth initial feature extraction unit and the first cutting unit in sequence to obtain the seventh initial feature map, the eighth initial feature map, and the ninth initial feature map;

[0280] For example, the first initial feature map X3 is input to the sixth CBS module for feature extraction to obtain the initial feature map H2. Then, the initial feature map H2 is input to the first cutting unit for split operation to obtain the seventh initial feature map H3, the eighth initial feature map H4, and the ninth initial feature map H5.

[0281] S122. Input the seventh initial feature map into the first depthwise separable convolutional unit to obtain the tenth initial feature map;

[0282] For example, if the first depthwise separable convolutional unit is a DSConv3×3 depthwise separable convolution, then the seventh initial feature map H3 is input into the DSConv3×3 depthwise separable convolution to obtain the tenth initial feature map H6.

[0283] S123. Input the eighth initial feature map into the second depthwise separable convolutional unit to obtain the eleventh initial feature map;

[0284] For example, if the second depthwise separable convolutional unit is a DSConv5×5 depthwise separable convolution, then the eighth initial feature map H4 is input into the DSConv5×5 depthwise separable convolution to obtain the eleventh initial feature map H7.

[0285] S124. Input the ninth initial feature map into the third depthwise separable convolutional unit to obtain the twelfth initial feature map;

[0286] For example, if the third depthwise separable convolutional unit is a DSConv7×7 depthwise separable convolution, then the ninth initial feature map H5 is input into the DSConv7×7 depthwise separable convolution to obtain the twelfth initial feature map H8.

[0287] S125. The seventh initial feature map, the eighth initial feature map, and the ninth initial feature map are concatenated to obtain the thirteenth initial feature map.

[0288] For example, the Concat operation is used to concatenate the seventh initial feature map H3, the eighth initial feature map H4, and the ninth initial feature map H5 to obtain the thirteenth initial feature map H9.

[0289] S126. Input the thirteenth initial feature map sequentially into the fourth, fifth, and sixth depthwise separable convolutional units to obtain the fourteenth initial feature map.

[0290] For example, if the fourth depthwise separable convolutional unit is a DSConv3×3 depthwise separable convolution, the fifth depthwise separable convolutional unit is a DSConv5×5 depthwise separable convolution, and the sixth depthwise separable convolutional unit is a DSConv7×7 depthwise separable convolution, then the thirteenth initial feature map H9 is input into the DSConv3×3 depthwise separable convolution to obtain the initial feature map H10. Then, the initial feature map H10 is input into the DSConv5×5 depthwise separable convolution to obtain the initial feature map H11. Then, the initial feature map H11 is input into the DSConv7×7 depthwise separable convolution to obtain the fourteenth initial feature map H12.

[0291] S127. Perform feature overlay processing on the fourteenth initial feature map and the thirteenth initial feature map to obtain the fifteenth initial feature map;

[0292] For example, the fourteenth initial feature map H12 and the thirteenth initial feature map H9 are superimposed to obtain the fifteenth initial feature map H13.

[0293] S128. Input the fifteenth initial feature map into the seventh initial feature extraction unit to obtain the sixteenth initial feature map;

[0294] For example, the fifteenth initial feature map H13 is input into the seventh CBS module for feature extraction to obtain the sixteenth initial feature map H14.

[0295] S129. Input the sixteenth initial feature map into the dual-path energy storage feature extraction sub-unit to obtain the second initial feature map.

[0296] For example, the sixteenth initial feature map H14 is input into the dual-path energy storage feature extraction sub-unit to obtain the second initial feature map X4.

[0297] For example, regarding S12 above, let's take the first initial feature map X3 with a size of 128×160×160 as an example for illustration:

[0298] 1) Input the first initial feature map X3 into the sixth CBS module to obtain an initial feature map H2 with a size of 132×160×160. Then input the initial feature map H2 into the first cutting unit to obtain the seventh initial feature map H3 with a size of 44×160×160, the eighth initial feature map H4 with a size of 44×160×160, and the ninth initial feature map H5 with a size of 44×160×160.

[0299] 2) Input the seventh initial feature map H3 into a DSConv3×3 depthwise separable convolution to obtain the tenth initial feature map H6 with a size of 44×160×160; input the eighth initial feature map H4 into a DSConv5×5 depthwise separable convolution to obtain the eleventh initial feature map H7 with a size of 44×160×160; input the ninth initial feature map H5 into a DSConv7×7 depthwise separable convolution to obtain the twelfth initial feature map H8 with a size of 44×160×160; use the Concat operation to concatenate the seventh initial feature map H3, the eighth initial feature map H4, and the ninth initial feature map H5 to obtain the thirteenth initial feature map H9 with a size of 132×160×160;

[0300] 3) Input the thirteenth initial feature map H9 into the DSConv3×3 depthwise separable convolution to obtain the initial feature map H10 with a size of 256×160×160. Then input the initial feature map H10 into the DSConv5×5 depthwise separable convolution to obtain the initial feature map H11 with a size of 512×160×160. Then input the initial feature map H11 into the DSConv7×7 depthwise separable convolution to obtain the fourteenth initial feature map H12 with a size of 132×160×160.

[0301] 4) Add the fourteenth initial feature map H12 and the thirteenth initial feature map H9 element by element to obtain the fifteenth initial feature map H13 with a size of 132×160×160; input the fifteenth initial feature map H13 into the seventh CBS module for feature extraction to obtain the sixteenth initial feature map H14 with a size of 128×160×160; input the sixteenth initial feature map H14 into the dual-path energy storage feature extraction sub-unit to obtain the second initial feature map X4 with a size of 128×160×160.

[0302] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0303] In one possible implementation, Figure 7 This is a schematic diagram of the structure of the dual-path energy storage feature extraction subunit provided in an embodiment of this application. Figure 7 As shown, the dual-path energy storage feature extraction subunit includes: multiple initial feature extraction units, a second cutting unit, a first convolutional block attention unit, and a second convolutional block attention unit;

[0304] For example, the dual-path energy storage feature extraction submodule uses dual-path CBS parallelism, feature superposition enhancement, and Convolutional Block Attention Module (CBAM) mechanism to strengthen the weight allocation of key features of energy storage devices, further improving feature representation capabilities and adaptability to subsequent tasks.

[0305] The initial feature extraction units include: the eighth initial feature extraction unit, the ninth initial feature extraction unit, the tenth initial feature extraction unit, and the eleventh initial feature extraction unit.

[0306] Accordingly, one possible implementation of S129 above includes the following steps:

[0307] Step 1: Input the sixteenth initial feature map into the eighth initial feature extraction unit and the ninth initial feature extraction unit respectively to obtain the first dual-path energy storage feature map and the second dual-path energy storage feature map;

[0308] For example, the sixteenth initial feature map H14 is input to the eighth CBS module for feature extraction to obtain the first dual-channel energy storage feature map F1, and the sixteenth initial feature map H14 is input to the ninth CBS module for feature extraction to obtain the second dual-channel energy storage feature map F2.

[0309] Step 2: Perform feature overlay processing on the first dual-path energy storage feature map and the second dual-path energy storage feature map to obtain the third dual-path energy storage feature map;

[0310] For example, the first dual-channel energy storage feature map F1 and the second dual-channel energy storage feature map F2 are superimposed to obtain the third dual-channel energy storage feature map F3.

[0311] Step 3: Input the third dual-path energy storage feature map into the tenth initial feature extraction unit to obtain the fourth dual-path energy storage feature map;

[0312] For example, the third dual-channel energy storage feature map F3 is input into the tenth CBS module for feature extraction to obtain the fourth dual-channel energy storage feature map F4.

[0313] Step 4: The first dual-channel energy storage feature map, the second dual-channel energy storage feature map, and the fourth dual-channel energy storage feature map are spliced ​​together to obtain the fifth dual-channel energy storage feature map.

[0314] For example, the Concat operation is used to stitch together the first dual-channel energy storage feature map F1, the second dual-channel energy storage feature map F2, and the fourth dual-channel energy storage feature map F4 to obtain the fifth dual-channel energy storage feature map F5.

[0315] Step 5: Input the fifth dual-path energy storage feature map into the second cutting unit to obtain the sixth dual-path energy storage feature map and the seventh dual-path energy storage feature map;

[0316] For example, the fifth dual-channel energy storage feature map F5 is input to the second cutting unit for split operation to obtain the sixth dual-channel energy storage feature map F6 and the seventh dual-channel energy storage feature map F7.

[0317] Step 6: Input the sixth dual-path energy storage feature map and the seventh dual-path energy storage feature map into the first convolutional block attention unit and the second convolutional block attention unit, respectively, to obtain the eighth dual-path energy storage feature map and the ninth dual-path energy storage feature map;

[0318] For example, the sixth dual-channel energy storage feature map F6 is input to the first CBAM attention unit to obtain the eighth dual-channel energy storage feature map F8, and the seventh dual-channel energy storage feature map F7 is input to the second CBAM attention unit to obtain the ninth dual-channel energy storage feature map F9.

[0319] Step 7: Combine the eighth dual-channel energy storage feature map and the ninth dual-channel energy storage feature map to obtain the tenth dual-channel energy storage feature map.

[0320] For example, the Concat operation is used to stitch together the eighth dual-channel energy storage feature map F8 and the ninth dual-channel energy storage feature map F9 to obtain the tenth dual-channel energy storage feature map F10.

[0321] Step 8: Input the tenth dual-path energy storage feature map into the eleventh initial feature extraction unit to obtain the second initial feature map.

[0322] For example, the tenth dual-channel energy storage feature map F10 is input into the eleventh CBS module for feature extraction to obtain the second initial feature map X4.

[0323] For example, regarding S129 above, let's take the sixteenth initial feature map H14, which has a size of 128×160×160, as an example for illustration:

[0324] 1) Input the sixteenth initial feature map H14 into the eighth CBS module to obtain the first dual-channel energy storage feature map F1 with a size of 128×160×160, and input the sixteenth initial feature map H14 into the ninth CBS module to obtain the second dual-channel energy storage feature map F2 with a size of 128×160×160; add the first dual-channel energy storage feature map F1 and the second dual-channel energy storage feature map F2 element by element to obtain the third dual-channel energy storage feature map F3 with a size of 128×160×160; input the third dual-channel energy storage feature map F3 into the tenth CBS module to obtain the fourth dual-channel energy storage feature map F4 with a size of 256×160×160; use the Concat operation to splice the first dual-channel energy storage feature map F1, the second dual-channel energy storage feature map F2, and the fourth dual-channel energy storage feature map F4 to obtain the fifth dual-channel energy storage feature map F5 with a size of 512×160×160;

[0325] 2) Input the fifth dual-channel energy storage feature map F5 into the second cutting unit for a split operation to obtain the sixth dual-channel energy storage feature map F6 and the seventh dual-channel energy storage feature map F7, both with a size of 256×160×160. Input the sixth dual-channel energy storage feature map F6 into the first CBAM attention unit to obtain the eighth dual-channel energy storage feature map F8, with a size of 256×160×160. Input the seventh dual-channel energy storage feature map F7 into the second CBAM attention unit to obtain the ninth dual-channel energy storage feature map F9, with a size of 256×160×160.

[0326] 3) The eighth dual-channel energy storage feature map F8 and the ninth dual-channel energy storage feature map F9 are spliced ​​together using the Concat operation to obtain the tenth dual-channel energy storage feature map F10 with a size of 512×160×160; the tenth dual-channel energy storage feature map F10 is input into the eleventh CBS module to obtain the second initial feature map X4 with a size of 128×160×160.

[0327] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0328] In one possible implementation, Figure 8This is a schematic diagram of the structure of the Precision Energy Storage Device Feature Perception Module (PESDFP) unit provided in an embodiment of this application. Figure 8 As shown, the feature sensing unit of the fine energy storage device includes: a first max pooling unit, a first average pooling unit, a first activation unit, a third cutting unit, multiple initial feature extraction units, and multiple depthwise separable convolution units;

[0329] For example, the fine energy storage device feature sensing unit dynamically calibrates the key feature response of the energy storage device through average pooling, max pooling, depthwise separable convolution, and Softmax activation function, avoiding information redundancy and weakening of key details, suppressing background interference, and effectively reducing the false alarm rate.

[0330] The initial feature extraction units include: the twelfth initial feature extraction unit, the thirteenth initial feature extraction unit, the fourteenth initial feature extraction unit, and the fifteenth initial feature extraction unit;

[0331] Multiple depthwise separable convolutional units include: a seventh depthwise separable convolutional unit and an eighth depthwise separable convolutional unit.

[0332] Accordingly, one possible implementation of S16 above includes the following steps:

[0333] S161. Input the fifth initial feature map into the first max pooling unit, the seventh depthwise separable convolution unit, the eighth depthwise separable convolution unit, and the first average pooling unit respectively to obtain the first fine energy storage feature map, the second fine energy storage feature map, the third fine energy storage feature map, and the fourth fine energy storage feature map.

[0334] For example, the fifth initial feature map X10 is input into the first max pooling unit for max pooling to obtain the first fine energy storage feature map O1;

[0335] The seventh depthwise separable convolutional unit is a DSConv5×5 depthwise separable convolution. Then, the fifth initial feature map X10 is input into the DSConv5×5 depthwise separable convolution to obtain the second fine energy storage feature map O2.

[0336] The eighth depthwise separable convolutional unit is a DSConv7×7 depthwise separable convolution. The fifth initial feature map X10 is then input into the DSConv7×7 depthwise separable convolution to obtain the third fine energy storage feature map O3.

[0337] The fifth initial feature map X10 is input into the first average pooling unit for average pooling to obtain the fourth fine energy storage feature map O4.

[0338] S162. Input the second and third refined energy storage feature maps into the twelfth and thirteenth initial feature extraction units, respectively, to obtain the fifth and sixth refined energy storage feature maps.

[0339] For example, the second fine energy storage feature map O2 is input to the twelfth CBS module for feature extraction to obtain the fifth fine energy storage feature map O5, and the third fine energy storage feature map O3 is input to the thirteenth CBS module for feature extraction to obtain the sixth fine energy storage feature map O6.

[0340] S163. The fifth and sixth fine energy storage feature maps are spliced ​​together to obtain the seventh fine energy storage feature map.

[0341] For example, the fifth fine energy storage feature map O5 and the sixth fine energy storage feature map O6 are spliced ​​together using the Concat operation to obtain the seventh fine energy storage feature map O7.

[0342] S164. Input the seventh fine energy storage feature map into the fourteenth initial feature extraction unit and the first activation unit in sequence to obtain the first fine energy storage weight map;

[0343] For example, the seventh fine energy storage feature map O7 is input to the fourteenth CBS module for feature extraction to obtain the fine energy storage feature map O8, and the fine energy storage feature map O8 is input to the first activation unit and activated by the Softmax activation function to obtain the first fine energy storage weight map Q1.

[0344] S165. Input the first fine energy storage weight map into the third cutting unit to obtain the second fine energy storage weight map and the third fine energy storage weight map;

[0345] For example, the first fine energy storage weight map Q1 is input to the third cutting unit for a split operation. Through the cutting process, the second fine energy storage weight map Q2 and the third fine energy storage weight map Q3 are obtained.

[0346] S166. Multiply the fifth fine energy storage feature map and the second fine energy storage weight map element by element to obtain the eighth fine energy storage feature map.

[0347] For example, the fifth fine energy storage feature map O5 and the second fine energy storage weight map Q2 are multiplied element by element to obtain the eighth fine energy storage feature map O9.

[0348] S167. Multiply the sixth fine energy storage feature map and the third fine energy storage weight map element by element to obtain the ninth fine energy storage feature map.

[0349] For example, the sixth fine energy storage feature map O6 and the third fine energy storage weight map Q3 are multiplied element by element to obtain the ninth fine energy storage feature map O10.

[0350] S168. The first, fourth, eighth, and ninth fine energy storage feature maps are spliced ​​together to obtain the tenth fine energy storage feature map.

[0351] For example, the first fine energy storage feature map O1, the fourth fine energy storage feature map O4, the eighth fine energy storage feature map O9, and the ninth fine energy storage feature map O10 are spliced ​​together using the Concat operation to obtain the tenth fine energy storage feature map O11.

[0352] S169. Input the tenth fine energy storage feature map into the fifteenth initial feature extraction unit to obtain the sixth initial feature map.

[0353] For example, the tenth fine energy storage feature map O11 is input into the fifteenth CBS module for feature extraction to obtain the sixth initial feature map X11.

[0354] For example, regarding S16 above, let's take the fifth initial feature map X10, which has a size of 512×20×20, as an example for illustration:

[0355] 1) Input the fifth initial feature map X10 into the first max pooling unit, and use max pooling with a kernel size of 5×5, a stride of 1, and a padding of 2 to obtain a first fine energy storage feature map O1 with a size of 512×20×20; input the fifth initial feature map X10 into a DSConv5×5 depthwise separable convolution to obtain a second fine energy storage feature map O2 with a size of 1024×20×20; input the fifth initial feature map X10 into a DSConv7×7 depthwise separable convolution to obtain a third fine energy storage feature map O3 with a size of 1024×20×20; input the fifth initial feature map X10 into the first average pooling unit, and use average pooling with a kernel size of 5×5, a stride of 1, and a padding of 2 to obtain a fourth fine energy storage feature map O4;

[0356] 2) Input the second fine energy storage feature map O2 into the twelfth CBS module to obtain the fifth fine energy storage feature map O5 with a size of 512×20×20; input the third fine energy storage feature map O3 into the thirteenth CBS module to obtain the sixth fine energy storage feature map O6 with a size of 512×20×20; use the Concat operation to stitch the fifth fine energy storage feature map O5 and the sixth fine energy storage feature map O6 to obtain the seventh fine energy storage feature map O7 with a size of 1024×20×20;

[0357] 3) Input the seventh fine energy storage feature map O7 into the fourteenth CBS module to obtain a fine energy storage feature map O8 with a size of 1024×20×20. Then, input the fine energy storage feature map O8 into the first activation unit and perform activation processing through the Softmax activation function to obtain a first fine energy storage weight map Q1 with a size of 1024×20×20. Input the first fine energy storage weight map Q1 into the third cutting unit for Split operation to obtain a second fine energy storage weight map Q2 with a size of 512×20×20 and a third fine energy storage weight map Q3 with a size of 512×20×20.

[0358] 4) Multiply the fifth fine energy storage feature map O5 and the second fine energy storage weight map Q2 element by element to obtain the eighth fine energy storage feature map O9 with a size of 512×20×20; multiply the sixth fine energy storage feature map O6 and the third fine energy storage weight map Q3 element by element to obtain the ninth fine energy storage feature map O10 with a size of 512×20×20.

[0359] 5) The first fine energy storage feature map O1, the fourth fine energy storage feature map O4, the eighth fine energy storage feature map O9, and the ninth fine energy storage feature map O10 are stitched together using the Concat operation to obtain the tenth fine energy storage feature map O11 with a size of 2048×20×20; the tenth fine energy storage feature map O11 is input into the fifteenth CBS module for feature extraction to obtain the sixth initial feature map X11.

[0360] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0361] S2, input multiple initial feature maps into the feature enhancement unit to obtain multiple enhanced feature maps;

[0362] For example, multiple initial feature maps are input into the feature enhancement unit, cross-scale fusion and feature filtering are performed on the initial feature maps, and features under different receptive fields are convolutionally mapped, upsampled and concatenated to obtain multiple enhanced feature maps.

[0363] In one possible implementation, Figure 9 This is a schematic diagram of the feature enhancement unit provided in an embodiment of this application. Figure 9 As shown, the feature enhancement unit includes: a multi-path feature convolution unit, a first upsampling unit, a second upsampling unit, multiple initial feature extraction units, and multiple Energy Storage Device Feature Filtering Module (ESDFF) units;

[0364] Among them, the multiple initial feature extraction units include: the sixteenth initial feature extraction unit and the seventeenth initial feature extraction unit;

[0365] The multiple energy storage device feature filtering units include: a first energy storage device feature filtering unit, a second energy storage device feature filtering unit, a third energy storage device feature filtering unit, and a fourth energy storage device feature filtering unit.

[0366] Accordingly, one possible implementation of S2 above includes the following steps:

[0367] S21. Input the sixth initial feature map into the multi-path feature convolution unit to obtain the first enhanced feature map;

[0368] For example, the sixth initial feature map X11 is input into a multi-path feature convolution unit for feature extraction to obtain the first enhanced feature map X12.

[0369] S22. Input the first enhanced feature map into the first upsampling unit to obtain the second enhanced feature map;

[0370] For example, the first enhanced feature map X12 is input to the first upsampling unit and upsampled through the UpSample operation to obtain the second enhanced feature map X13.

[0371] S23. The fourth initial feature map and the second enhanced feature map are concatenated to obtain the third enhanced feature map;

[0372] For example, the Concat operation is used to concatenate the fourth initial feature map X8 and the second enhanced feature map X13 to obtain the third enhanced feature map X14.

[0373] S24. Input the third enhanced feature map into the feature filtering unit of the first energy storage device to obtain the fourth enhanced feature map;

[0374] For example, the third enhanced feature map X14 is input into the first ESDFF unit for feature filtering to obtain the fourth enhanced feature map X15.

[0375] S25. Input the fourth enhanced feature map into the second upsampling unit to obtain the fifth enhanced feature map;

[0376] For example, the fourth enhanced feature map X15 is input to the second upsampling unit and upsampled through the UpSample operation to obtain the fifth enhanced feature map X16.

[0377] S26. The third initial feature map and the fifth enhanced feature map are concatenated to obtain the sixth enhanced feature map;

[0378] For example, the Concat operation is used to concatenate the third initial feature map X6 and the fifth enhanced feature map X16 to obtain the sixth enhanced feature map X17.

[0379] S27. Input the sixth enhanced feature map into the feature filtering unit of the second energy storage device to obtain the seventh enhanced feature map;

[0380] For example, the sixth enhanced feature map X17 is input into the second ESDFF unit for feature filtering to obtain the seventh enhanced feature map X18.

[0381] S28. Input the seventh enhanced feature map into the sixteenth initial feature extraction unit to obtain the eighth enhanced feature map;

[0382] For example, the seventh enhanced feature map X18 is input into the sixteenth CBS module to obtain the eighth enhanced feature map X19.

[0383] S29. The fourth and eighth enhanced feature maps are concatenated to obtain the ninth enhanced feature map.

[0384] For example, the fourth enhanced feature map X15 and the eighth enhanced feature map X19 are concatenated using the Concat operation to obtain the ninth enhanced feature map X20.

[0385] S210. Input the ninth enhanced feature map into the third energy storage device feature filtering unit to obtain the tenth enhanced feature map;

[0386] For example, the ninth enhanced feature map X20 is input into the third ESDFF unit for feature filtering to obtain the tenth enhanced feature map X21.

[0387] S211. Input the tenth enhanced feature map into the seventeenth initial feature extraction unit to obtain the eleventh enhanced feature map;

[0388] For example, the tenth enhanced feature map X21 is input into the seventeenth CBS module to obtain the eleventh enhanced feature map X22.

[0389] S212. The first enhanced feature map and the eleventh enhanced feature map are concatenated to obtain the twelfth enhanced feature map.

[0390] For example, the first enhanced feature map X12 and the eleventh enhanced feature map X22 are concatenated using the Concat operation to obtain the twelfth enhanced feature map X23.

[0391] S213. Input the twelfth enhanced feature map into the fourth energy storage device feature filtering unit to obtain the thirteenth enhanced feature map.

[0392] For example, the twelfth enhanced feature map X23 is input into the fourth ESDFF unit for feature filtering to obtain the thirteenth enhanced feature map X24.

[0393] For example, regarding S2 above, let's take the third initial feature map X6 with an input size of 256×80×80, the fourth initial feature map X8 with a size of 512×40×40, and the sixth initial feature map X11 with a size of 512×20×20 as examples for illustration:

[0394] 1) Input the sixth initial feature map X11 into the multi-path feature convolution unit to obtain the first enhanced feature map X12 with a size of 512×20×20; input the first enhanced feature map X12 into the first upsampling unit to obtain the second enhanced feature map X13 with a size of 512×40×40;

[0395] 2) The fourth initial feature map X8 and the second enhanced feature map X13 are concatenated using the Concat operation to obtain the third enhanced feature map X14 with a size of 1024×40×40; the third enhanced feature map X14 is input into the first ESDFF unit to obtain the fourth enhanced feature map X15 with a size of 512×40×40; the fourth enhanced feature map X15 is input into the second upsampling unit to obtain the fifth enhanced feature map X16 with a size of 512×80×80.

[0396] 3) The third initial feature map X6 and the fifth enhanced feature map X16 are concatenated using the Concat operation to obtain the sixth enhanced feature map X17 with a size of 768×80×80; the sixth enhanced feature map X17 is input into the second ESDFF unit to obtain the seventh enhanced feature map X18 with a size of 256×80×80.

[0397] 4) Input the seventh enhanced feature map X18 into the sixteenth CBS module to obtain the eighth enhanced feature map X19 with a size of 256×40×40; use the Concat operation to concatenate the fourth enhanced feature map X15 and the eighth enhanced feature map X19 to obtain the ninth enhanced feature map X20 with a size of 768×40×40; input the ninth enhanced feature map X20 into the third ESDFF unit to obtain the tenth enhanced feature map X21 with a size of 512×40×40;

[0398] 5) Input the tenth enhanced feature map X21 into the seventeenth CBS module to obtain the eleventh enhanced feature map X22 with a size of 512×20×20; use the Concat operation to concatenate the first enhanced feature map X12 and the eleventh enhanced feature map X22 to obtain the twelfth enhanced feature map X23 with a size of 1024×20×20; input the twelfth enhanced feature map X23 into the fourth ESDFF unit to obtain the thirteenth enhanced feature map X24 with a size of 1024×20×20.

[0399] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0400] In one possible implementation, Figure 10 This is a schematic diagram of the structure of a feature filtering unit for an energy storage device provided in an embodiment of this application. Figure 10 As shown, each Energy Storage Device FeatureFiltering Module (ESDFF) unit includes: a feature collaboration subunit, a ninth depthwise separable convolution unit, a batch normalization unit, a second activation unit, a fourth slicing unit, and multiple initial feature extraction units;

[0401] For example, the feature filtering unit of the energy storage device uses parallel feature extraction to obtain two types of filter feature maps, and then obtains preliminary fusion features through element-wise addition and depthwise separable convolution. After that, a dynamic weight map is generated, and then different weight branches are obtained through a cutting operation. The feature maps are weighted and interactively fused to achieve dynamic screening and enhancement of key fault features.

[0402] The initial feature extraction units include: the eighteenth initial feature extraction unit, the nineteenth initial feature extraction unit, and the twentieth initial feature extraction unit.

[0403] Accordingly, one possible implementation of S24 above includes the following steps:

[0404] S241. Input the third enhanced feature map into the eighteenth initial feature extraction unit and the nineteenth initial feature extraction unit respectively to obtain the first filtered feature map and the second filtered feature map;

[0405] For example, the third enhanced feature map X14 is input to the eighteenth CBS module for feature extraction to obtain the first filtered feature map U2, and the third enhanced feature map X14 is input to the nineteenth CBS module for feature extraction to obtain the second filtered feature map U3.

[0406] S242. Perform feature overlay processing on the first and second filtered feature maps to obtain the third filtered feature map;

[0407] For example, the first filter feature map U2 and the second filter feature map U3 are superimposed to obtain the third filter feature map U4.

[0408] S243. The third filter feature map is sequentially input into the ninth depthwise separable convolutional unit, the batch normalization unit, and the second activation unit to obtain the first filter weight map.

[0409] For example, if the ninth depthwise separable convolutional unit is a DSConv1×1 depthwise separable convolution, then the third filter feature map U4 is input into the DSConv1×1 depthwise separable convolution to obtain the filter feature map U5. Then, the filter feature map U5 is input into the batch normalization unit for batch normalization processing to obtain the filter feature map U6. After that, the filter feature map U6 is input into the second activation unit and activated by the Softmax activation function to obtain the first filter weight map V1.

[0410] S244. Input the first filter weight map into the fourth cutting unit to obtain the second filter weight map and the third filter weight map;

[0411] For example, the first filter weight map V1 is input to the fourth cutting unit to perform a split operation, and the second filter weight map V2 and the third filter weight map V3 are obtained through the cutting process.

[0412] S245. Multiply the first filter feature map and the second filter weight map element by element to obtain the fourth filter feature map.

[0413] For example, the first filter feature map U2 and the second filter weight map V2 are multiplied element-wise to obtain the fourth filter feature map U7.

[0414] S246. Multiply the second filter feature map and the third filter weight map element by element to obtain the fifth filter feature map.

[0415] For example, the second filter feature map U3 and the third filter weight map V3 are multiplied element-wise to obtain the fifth filter feature map U8.

[0416] S247. Multiply the fourth and fifth filter feature maps element by element to obtain the sixth filter feature map.

[0417] For example, the fourth filter feature map U7 and the fifth filter feature map U8 are multiplied element-wise to obtain the sixth filter feature map U9.

[0418] S248. Add the fourth and fifth filter feature maps element by element to obtain the seventh filter feature map.

[0419] For example, the fourth filter feature map U7 and the fifth filter feature map U8 are added element by element to obtain the seventh filter feature map U10.

[0420] S249. Input the third enhanced feature map into the feature co-operation subunit to obtain the eighth filtered feature map;

[0421] For example, the third enhanced feature map X14 is input into the feature collaboration subunit for feature collaboration processing to obtain the eighth filtered feature map U11.

[0422] S2410. The third enhanced feature map, the sixth filtered feature map, the seventh filtered feature map, and the eighth filtered feature map are concatenated to obtain the ninth filtered feature map.

[0423] For example, the third enhanced feature map X14, the sixth filtered feature map U9, the seventh filtered feature map U10, and the eighth filtered feature map U11 are concatenated using the Concat operation to obtain the ninth filtered feature map U12.

[0424] S2411. Input the ninth filtered feature map into the twentieth initial feature extraction unit to obtain the fourth enhanced feature map.

[0425] For example, the ninth filtered feature map U12 is input into the twentieth CBS module to obtain the fourth enhanced feature map X15.

[0426] For example, regarding S24 above, let's take the third enhanced feature map X14, which has a size of 1024×40×40, as an example for illustration:

[0427] 1) Input the third enhanced feature map X14 into the eighteenth CBS module to obtain the first filter feature map U2 with a size of 1024×40×40, and input the third enhanced feature map X14 into the nineteenth CBS module to obtain the second filter feature map U3 with a size of 1024×40×40;

[0428] 2) Element-wise sum the first filter feature map U2 and the second filter feature map U3 to obtain a third filter feature map U4 with a size of 1024×40×40; input the third filter feature map U4 into a DSConv1×1 depthwise separable convolution to obtain a filter feature map U5 with a size of 2048×40×40; input the filter feature map U5 into a batch normalization unit to obtain a filter feature map U6 with a size of 2048×40×40; input the filter feature map U6 into a second activation unit to obtain a first filter weight map V1 with a size of 2048×40×40; input the first filter weight map V1 into a fourth cutting unit to obtain a second filter weight map V2 with a size of 1024×40×40 and a third filter weight map V3 with a size of 1024×40×40.

[0429] 3) Multiply the first filter feature map U2 and the second filter weight map V2 element-wise to obtain the fourth filter feature map U7 with a size of 1024×40×40; multiply the second filter feature map U3 and the third filter weight map V3 element-wise to obtain the fifth filter feature map U8 with a size of 1024×40×40; multiply the fourth filter feature map U7 and the fifth filter feature map U8 element-wise to obtain the sixth filter feature map U9 with a size of 1024×40×40; add the fourth filter feature map U7 and the fifth filter feature map U8 element-wise to obtain the seventh filter feature map U10 with a size of 1024×40×40.

[0430] 4) Input the third enhanced feature map X14 into the feature co-operation sub-unit to obtain the eighth filter feature map U11 with a size of 1024×40×40; use the Concat operation to concatenate the third enhanced feature map X14, the sixth filter feature map U9, the seventh filter feature map U10, and the eighth filter feature map U11 to obtain the ninth filter feature map U12 with a size of 4096×40×40; input the ninth filter feature map U12 into the twentieth CBS module to obtain the fourth enhanced feature map X15.

[0431] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0432] In one possible implementation, Figure 11 This is a schematic diagram of the structure of the feature cooperative subunit provided in an embodiment of this application. For example... Figure 11 As shown, the feature collaboration subunit includes: a second average pooling unit, a third average pooling unit, a second max pooling unit, a third max pooling unit, and multiple initial feature extraction units;

[0433] For example, the feature collaboration subunit uses multi-level average pooling and max pooling concatenation operations, combined with feature addition, to further enhance the model's ability to maintain feature collaboration and structural information under different thermal field distribution patterns, thereby effectively improving the model's adaptability and generalization performance under dynamic operating conditions.

[0434] The initial feature extraction units include: the twenty-first initial feature extraction unit, the twenty-second initial feature extraction unit, and the twenty-third initial feature extraction unit.

[0435] Accordingly, one possible implementation of S249 above includes the following steps:

[0436] Step 1: Input the third enhanced feature map into the twenty-first initial feature extraction unit to obtain the first collaborative feature map;

[0437] For example, the third enhanced feature map X14 is input into the twenty-first CBS module for feature extraction to obtain the first collaborative feature map B1.

[0438] Step 2: Input the first collaborative feature map into the second average pooling unit and the third average pooling unit in sequence to obtain the second collaborative feature map;

[0439] For example, the first collaborative feature map B1 is input into the second average pooling unit to obtain the collaborative feature map B2, and then the collaborative feature map B2 is input into the third average pooling unit to obtain the second collaborative feature map B3.

[0440] Step 3: Concatenate the first and second collaborative feature maps to obtain the third collaborative feature map;

[0441] For example, the Concat operation is used to concatenate the first collaborative feature map B1 and the second collaborative feature map B3 to obtain the third collaborative feature map B4.

[0442] Step 4: Input the third collaborative feature map into the second max pooling unit and the third max pooling unit in sequence to obtain the fourth collaborative feature map;

[0443] For example, the third collaborative feature map B4 is input into the second max pooling unit to obtain the collaborative feature map B5, and then the collaborative feature map B5 is input into the third max pooling unit to obtain the fourth collaborative feature map B6.

[0444] Step 5: Concatenate the third and fourth collaborative feature maps to obtain the fifth collaborative feature map;

[0445] For example, the third collaborative feature map B4 and the fourth collaborative feature map B5 are concatenated using the Concat operation to obtain the fifth collaborative feature map B7.

[0446] Step 6: Input the fifth collaborative feature map into the twenty-second initial feature extraction unit to obtain the sixth collaborative feature map;

[0447] For example, the fifth collaborative feature map B7 is input into the twenty-second CBS module for feature extraction to obtain the sixth collaborative feature map B8.

[0448] Step 7: Perform feature overlay processing on the first collaborative feature map and the sixth collaborative feature map to obtain the seventh collaborative feature map;

[0449] For example, the first collaborative feature map B1 and the sixth collaborative feature map B8 are superimposed to obtain the seventh collaborative feature map B9.

[0450] Step 8: Input the seventh collaborative feature map into the twenty-third initial feature extraction unit to obtain the eighth filtered feature map.

[0451] For example, the seventh collaborative feature map B9 is input into the twenty-third CBS module for feature extraction to obtain the eighth filtered feature map U11.

[0452] For example, regarding S249 above, let's take the third enhanced feature map X14, which has a size of 1024×40×40, as an example for illustration:

[0453] 1) Input the third enhanced feature map X14 into the twenty-first CBS module to obtain a first collaborative feature map B1 with a size of 1024×40×40; input the first collaborative feature map B1 into the second average pooling unit and use average pooling with a kernel size of 3×3, a stride of 1, and a padding of 1 to obtain a collaborative feature map B2 with a size of 1024×40×40; then input the collaborative feature map B2 into the third average pooling unit and use average pooling with a kernel size of 3×3, a stride of 1, and a padding of 1 to obtain a second collaborative feature map B3 with a size of 1024×40×40; use the Concat operation to concatenate the first collaborative feature map B1 and the second collaborative feature map B3 to obtain a third collaborative feature map B4 with a size of 2048×40×40.

[0454] 2) Input the third collaborative feature map B4 into the second max pooling unit and use max pooling with a kernel size of 3×3, a stride of 1, and a padding of 1 to obtain a collaborative feature map B5 with a size of 2048×40×40; input the collaborative feature map B5 into the third max pooling unit and use max pooling with a kernel size of 3×3, a stride of 1, and a padding of 1 to obtain a fourth collaborative feature map B6 with a size of 2048×40×40; use the Concat operation to concatenate the third collaborative feature map B4 and the fourth collaborative feature map B5 to obtain a fifth collaborative feature map B7 with a size of 4096×40×40; input the fifth collaborative feature map B7 into the twenty-second CBS module to obtain a sixth collaborative feature map B8 with a size of 1024×40×40;

[0455] 3) Perform feature overlay processing on the first collaborative feature map B1 and the sixth collaborative feature map B8 to obtain the seventh collaborative feature map B9 with a size of 1024×40×40; input the seventh collaborative feature map B9 into the twenty-third CBS module to obtain the eighth filter feature map U11 with a size of 1024×40×40.

[0456] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0457] S3, input multiple enhanced feature maps into the detection head unit to obtain predicted thermal fault information;

[0458] For example, the detection head unit is used to convert multiple enhanced feature maps into prediction results corresponding to fault locations and fault types, and output predicted thermal fault information.

[0459] In one possible implementation, Figure 12 This is a schematic diagram of the detection head unit provided in an embodiment of this application. Figure 12 As shown, the detection head unit includes: a first detection head, a second detection head, and a third detection head;

[0460] For example, one possible implementation of S3 above includes the following steps:

[0461] Step 1: Input the seventh enhanced feature map X18 into the first detection head to obtain feature map X25, and deform feature map X25 through the Reshape operation to obtain feature map X28;

[0462] Step 2: Input the tenth enhanced feature map X21 into the second detection head to obtain feature map X26, and deform feature map X26 through the Reshape operation to obtain feature map X29;

[0463] Step 3: Input the thirteenth enhanced feature map X24 into the third detection head to obtain feature map X27, and deform feature map X27 through the Reshape operation to obtain feature map X30;

[0464] Step 4: Use the Concat operation to stitch together feature map X28, feature map X29, and feature map X30 to obtain the initial thermal fault information X31;

[0465] Step 5: The initial thermal fault information X31 is filtered through a non-maximum suppression algorithm to select high-confidence prediction boxes, thus obtaining the predicted thermal fault information.

[0466] Each row of the predicted thermal fault information corresponds to a prediction result, which includes the corresponding prediction box coordinates, confidence score, and fault category label.

[0467] For example, regarding S2 above, let's take the seventh enhanced feature map X18 with an input size of 256×80×80, the tenth enhanced feature map X21 with a size of 512×40×40, and the thirteenth enhanced feature map X24 with a size of 1024×20×20 as examples for illustration:

[0468] 1) Input the seventh enhanced feature map X18 into the first detection head to obtain a feature map X25 with a size of Na(5+Nc)×80×80, and deform the feature map X25 by the Reshape operation to obtain a feature map X28 with a size of 6400Na(5+Nc);

[0469] 2) Input the tenth enhanced feature map X21 into the second detection head to obtain a feature map X26 with a size of Na(5+Nc)×40×40. Then, perform a Reshape operation on the feature map X26 to obtain a feature map X29 with a size of 1600Na(5+Nc).

[0470] 3) Input the thirteenth enhanced feature map X24 into the third detection head to obtain a feature map X27 with a size of Na(5+Nc)×20×20. Then, deform the feature map X27 through the Reshape operation to obtain a feature map X30 with a size of 400Na(5+Nc).

[0471] 4) The Concat operation is used to concatenate feature maps X28, X29 and X30 to obtain initial thermal fault information X31 with a size of 8400Na(5+Nc). Then, the initial thermal fault information X31 is filtered for high-confidence prediction boxes through the non-maximum suppression algorithm to obtain predicted thermal fault information with a size of N×(5+Nc).

[0472] Where Na represents the number of anchor boxes (a hyperparameter), Nc represents the number of label types (a hyperparameter), 5+Nc represents the number of parameters of the predicted box, including the x-coordinate of the center point of the predicted box, the y-coordinate of the center point, the width w, the height h, the confidence score, and the number of class labels, and N represents the number of predicted boxes after processing by the non-maximum suppression algorithm.

[0473] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0474] S4. Update the parameters of the initial thermal fault detection model based on the predicted thermal fault information and the labeled thermal fault information corresponding to each image. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the thermal fault detection model.

[0475] For example, the initial thermal fault detection model is trained and optimized by supervised learning based on the predicted thermal fault information and the labeled thermal fault information corresponding to each image. The parameters are continuously updated by the backpropagation algorithm until the preset training stopping condition is met, and the thermal fault detection model is obtained.

[0476] In one possible implementation, an initial thermal fault detection model can be trained based on the training set of the aforementioned optimized energy storage device dataset, and the training effect of the initial thermal fault detection model can be evaluated using the validation set of the optimized energy storage device dataset. Finally, the generalization ability of the trained thermal fault detection model can be verified using the test set of the optimized energy storage device dataset.

[0477] In one possible implementation, Figure 13This is a schematic diagram of the structure of the thermal fault detection model provided in an embodiment of this application. Figure 13 As shown, the thermal fault detection model specifically includes: a feature extraction unit, a feature enhancement unit, and a detection head unit;

[0478] For example, the feature extraction unit includes: multiple initial feature extraction units, multiple multi-level energy storage device feature extraction (MESDF) units, and a precision energy storage device feature perception (PESDFP) unit; wherein, the multiple initial feature extraction units include: a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, and a fifth CBS module, and the multiple multi-level energy storage device feature extraction units include: a first MESDF unit, a second MESDF unit, a third MESDF unit, and a fourth MESDF unit;

[0479] The feature enhancement unit includes: a multi-path feature convolution unit, a first upsampling unit, a second upsampling unit, multiple initial feature extraction units, and multiple energy storage device feature filtering (ESDFF) units; wherein, the multiple initial feature extraction units include: a sixteenth CBS module and a seventeenth CBS module, and the multiple energy storage device feature filtering units include: a first ESDFF unit, a second ESDFF unit, a third ESDFF unit, and a fourth ESDFF unit;

[0480] The detection head unit includes: a first detection head, a second detection head, and a third detection head.

[0481] It is understandable that the thermal fault detection model and the initial thermal fault detection model have the same model structure, the difference being that the model parameters are different. The parameters of the initial thermal fault detection model are updated until the preset training stopping condition is met, thus obtaining the thermal fault detection model.

[0482] In addition, when applying the thermal fault detection model, the target image of the energy storage device to be tested is directly input into the thermal fault detection model to obtain the thermal fault information of the energy storage device to be tested.

[0483] The thermal fault detection method for energy storage devices provided in this application first acquires images of multiple energy storage devices and corresponding labeled thermal fault information for each image. Then, it trains an initial thermal fault detection model based on the images of the multiple energy storage devices and the labeled thermal fault information for each image to obtain a thermal fault detection model. In this embodiment, training the initial thermal fault detection model based on images of multiple energy storage devices and the labeled thermal fault information for each image enables it to learn local hotspot details and overall thermal distribution features in the images, allowing the thermal fault detection model obtained based on the initial thermal fault detection model to accurately output thermal fault information.

[0484] Based on the above embodiments, the following are embodiments of the apparatus involved in this application:

[0485] Figure 14 This is a schematic diagram of the structure of a thermal fault detection device for an energy storage device provided in an embodiment of this application. Figure 14 As shown, the thermal fault detection device 1400 of the energy storage device includes:

[0486] The acquisition module 1401 is used to acquire the target image of the energy storage device to be tested;

[0487] The processing module 1402 is used to input the target image into the thermal fault detection model to obtain the thermal fault information of the energy storage device to be detected. The thermal fault detection model is determined based on images of multiple energy storage devices and the labeled thermal fault information corresponding to each image. The thermal fault information includes: fault location and fault type.

[0488] In one or more embodiments, before inputting the target image into the thermal fault detection model to obtain the thermal fault information of the energy storage device to be detected, the processing module 1402 is further configured to:

[0489] Acquire images of multiple energy storage devices and corresponding labeled thermal fault information for each image;

[0490] The initial thermal fault detection model is trained by using images of multiple energy storage devices and the corresponding labeled thermal fault information for each image, thus obtaining the thermal fault detection model.

[0491] In one or more embodiments, the initial thermal fault detection model includes: a feature extraction unit, a feature enhancement unit, and a detection head unit;

[0492] Accordingly, the processing module 1402 trains the initial thermal fault detection model based on images of multiple energy storage devices and the labeled thermal fault information corresponding to each image, to obtain a thermal fault detection model, specifically used for:

[0493] S1, For each energy storage device, the image of the energy storage device is input into the feature extraction unit to obtain multiple initial feature maps;

[0494] S2, input multiple initial feature maps into the feature enhancement unit to obtain multiple enhanced feature maps;

[0495] S3, input multiple enhanced feature maps into the detection head unit to obtain predicted thermal fault information;

[0496] S4. Update the parameters of the initial thermal fault detection model based on the predicted thermal fault information and the labeled thermal fault information corresponding to each image. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the thermal fault detection model.

[0497] In one or more embodiments, the feature extraction unit includes: multiple initial feature extraction units, multiple multi-level energy storage device feature extraction units, and a fine energy storage device feature sensing unit;

[0498] Correspondingly, the processing module 1402, for each energy storage device's image, inputs the image of the energy storage device into the feature extraction unit to obtain multiple initial feature maps, specifically used for:

[0499] For each energy storage device image, the images of the energy storage devices are sequentially input into the first initial feature extraction unit and the second initial feature extraction unit to obtain the first initial feature map;

[0500] The first initial feature map is input into the feature extraction unit of the first multi-level energy storage device to obtain the second initial feature map;

[0501] The second initial feature map is sequentially input into the third initial feature extraction unit and the second multi-level energy storage device feature extraction unit to obtain the third initial feature map;

[0502] The third initial feature map is sequentially input into the fourth initial feature extraction unit and the third multi-level energy storage device feature extraction unit to obtain the fourth initial feature map;

[0503] The fourth initial feature map is sequentially input into the fifth initial feature extraction unit and the fourth multi-level energy storage device feature extraction unit to obtain the fifth initial feature map.

[0504] The fifth initial feature map is input into the feature sensing unit of the fine energy storage device to obtain the sixth initial feature map.

[0505] In one or more embodiments, each multi-level energy storage device feature extraction unit includes: multiple initial feature extraction units, a first slicing unit, multiple depthwise separable convolution units, and a dual-path energy storage feature extraction subunit;

[0506] Correspondingly, the processing module 1402 inputs the first initial feature map to the first multi-level energy storage device feature extraction unit to obtain the second initial feature map, specifically used for:

[0507] The first initial feature map is sequentially input into the sixth initial feature extraction unit and the first cutting unit to obtain the seventh initial feature map, the eighth initial feature map, and the ninth initial feature map;

[0508] The seventh initial feature map is input into the first depthwise separable convolutional unit to obtain the tenth initial feature map;

[0509] The eighth initial feature map is input into the second depthwise separable convolutional unit to obtain the eleventh initial feature map;

[0510] The ninth initial feature map is input into the third depthwise separable convolutional unit to obtain the twelfth initial feature map;

[0511] The seventh, eighth, and ninth initial feature maps are concatenated to obtain the thirteenth initial feature map.

[0512] The thirteenth initial feature map is sequentially input into the fourth, fifth, and sixth depthwise separable convolutional units to obtain the fourteenth initial feature map.

[0513] The fourteenth and thirteenth initial feature maps are superimposed to obtain the fifteenth initial feature map.

[0514] The fifteenth initial feature map is input into the seventh initial feature extraction unit to obtain the sixteenth initial feature map;

[0515] The sixteenth initial feature map is input into the dual-path energy storage feature extraction sub-unit to obtain the second initial feature map.

[0516] In one or more embodiments, the dual-path energy storage feature extraction subunit includes: multiple initial feature extraction units, a second segmentation unit, a first convolutional block attention unit, and a second convolutional block attention unit;

[0517] Correspondingly, the processing module 1402 inputs the sixteenth initial feature map into the dual-path energy storage feature extraction subunit to obtain the second initial feature map, which is specifically used for:

[0518] The sixteenth initial feature map is input into the eighth initial feature extraction unit and the ninth initial feature extraction unit respectively to obtain the first dual-path energy storage feature map and the second dual-path energy storage feature map;

[0519] The first dual-path energy storage feature map and the second dual-path energy storage feature map are superimposed to obtain the third dual-path energy storage feature map;

[0520] The third dual-path energy storage feature map is input into the tenth initial feature extraction unit to obtain the fourth dual-path energy storage feature map;

[0521] The first dual-channel energy storage feature map, the second dual-channel energy storage feature map, and the fourth dual-channel energy storage feature map are stitched together to obtain the fifth dual-channel energy storage feature map.

[0522] The fifth dual-path energy storage feature map is input into the second cutting unit to obtain the sixth dual-path energy storage feature map and the seventh dual-path energy storage feature map;

[0523] The sixth dual-path energy storage feature map and the seventh dual-path energy storage feature map are input into the first convolutional block attention unit and the second convolutional block attention unit, respectively, to obtain the eighth dual-path energy storage feature map and the ninth dual-path energy storage feature map;

[0524] The eighth dual-channel energy storage feature map and the ninth dual-channel energy storage feature map are spliced ​​together to obtain the tenth dual-channel energy storage feature map.

[0525] The tenth dual-path energy storage feature map is input into the eleventh initial feature extraction unit to obtain the second initial feature map.

[0526] In one or more embodiments, the fine energy storage device feature sensing unit includes: a first max pooling unit, a first average pooling unit, a first activation unit, a third slicing unit, multiple initial feature extraction units, and multiple depthwise separable convolution units;

[0527] Correspondingly, the processing module 1402 inputs the fifth initial feature map to the fine energy storage device feature sensing unit to obtain the sixth initial feature map, which is specifically used for:

[0528] The fifth initial feature map is input into the first max pooling unit, the seventh depthwise separable convolution unit, the eighth depthwise separable convolution unit, and the first average pooling unit, respectively, to obtain the first fine energy storage feature map, the second fine energy storage feature map, the third fine energy storage feature map, and the fourth fine energy storage feature map;

[0529] The second and third refined energy storage feature maps are input into the twelfth and thirteenth initial feature extraction units, respectively, to obtain the fifth and sixth refined energy storage feature maps.

[0530] The fifth and sixth fine energy storage feature maps are stitched together to obtain the seventh fine energy storage feature map.

[0531] The seventh fine energy storage feature map is sequentially input into the fourteenth initial feature extraction unit and the first activation unit to obtain the first fine energy storage weight map;

[0532] The first fine energy storage weight map is input into the third cutting unit to obtain the second fine energy storage weight map and the third fine energy storage weight map;

[0533] The fifth fine energy storage feature map and the second fine energy storage weight map are multiplied element by element to obtain the eighth fine energy storage feature map.

[0534] The sixth fine energy storage feature map and the third fine energy storage weight map are multiplied element by element to obtain the ninth fine energy storage feature map.

[0535] The first, fourth, eighth, and ninth fine energy storage feature maps are stitched together to obtain the tenth fine energy storage feature map.

[0536] The tenth fine energy storage feature map is input into the fifteenth initial feature extraction unit to obtain the sixth initial feature map.

[0537] In one or more embodiments, the feature enhancement unit includes: a multi-path feature convolution unit, a first upsampling unit, a second upsampling unit, multiple initial feature extraction units, and multiple energy storage device feature filtering units;

[0538] Correspondingly, the processing module 1402 inputs multiple initial feature maps into the feature enhancement unit to obtain multiple enhanced feature maps, specifically used for:

[0539] The sixth initial feature map is input into the multi-path feature convolution unit to obtain the first enhanced feature map;

[0540] The first enhanced feature map is input into the first upsampling unit to obtain the second enhanced feature map;

[0541] The fourth initial feature map and the second enhanced feature map are concatenated to obtain the third enhanced feature map;

[0542] The third enhanced feature map is input into the feature filtering unit of the first energy storage device to obtain the fourth enhanced feature map;

[0543] The fourth enhanced feature map is input into the second upsampling unit to obtain the fifth enhanced feature map;

[0544] The third initial feature map and the fifth enhanced feature map are concatenated to obtain the sixth enhanced feature map.

[0545] The sixth enhanced feature map is input into the feature filtering unit of the second energy storage device to obtain the seventh enhanced feature map;

[0546] The seventh enhanced feature map is input into the sixteenth initial feature extraction unit to obtain the eighth enhanced feature map;

[0547] The fourth and eighth enhanced feature maps are concatenated to obtain the ninth enhanced feature map;

[0548] The ninth enhanced feature map is input into the feature filtering unit of the third energy storage device to obtain the tenth enhanced feature map;

[0549] The tenth enhanced feature map is input into the seventeenth initial feature extraction unit to obtain the eleventh enhanced feature map;

[0550] The first and eleventh enhanced feature maps are concatenated to obtain the twelfth enhanced feature map.

[0551] The twelfth enhanced feature map is input into the fourth energy storage device feature filtering unit to obtain the thirteenth enhanced feature map.

[0552] In one or more embodiments, each energy storage device feature filtering unit includes: a feature coordination subunit, a ninth depthwise separable convolution unit, a batch normalization unit, a second activation unit, a fourth slicing unit, and a plurality of initial feature extraction units;

[0553] Correspondingly, the processing module 1402 inputs the third enhanced feature map to the first energy storage device feature filtering unit to obtain the fourth enhanced feature map, which is specifically used for:

[0554] The third enhanced feature map is input into the eighteenth initial feature extraction unit and the nineteenth initial feature extraction unit respectively to obtain the first filtered feature map and the second filtered feature map;

[0555] The first and second filtered feature maps are superimposed to obtain the third filtered feature map.

[0556] The third filtered feature map is sequentially input into the ninth depthwise separable convolutional unit, the batch normalization unit, and the second activation unit to obtain the first filtered weight map.

[0557] The first filter weight map is input into the fourth cutting unit to obtain the second filter weight map and the third filter weight map;

[0558] The first filter feature map and the second filter weight map are multiplied element by element to obtain the fourth filter feature map.

[0559] The second filter feature map and the third filter weight map are multiplied element by element to obtain the fifth filter feature map;

[0560] The fourth and fifth filter feature maps are multiplied element-wise to obtain the sixth filter feature map.

[0561] The fourth and fifth filter feature maps are added element by element to obtain the seventh filter feature map.

[0562] The third enhanced feature map is input into the feature co-operation sub-unit to obtain the eighth filtered feature map;

[0563] The third enhanced feature map, the sixth filtered feature map, the seventh filtered feature map, and the eighth filtered feature map are concatenated to obtain the ninth filtered feature map.

[0564] The ninth filtered feature map is input into the twentieth initial feature extraction unit to obtain the fourth enhanced feature map.

[0565] In one or more embodiments, the feature collaboration subunit includes: a second average pooling unit, a third average pooling unit, a second max pooling unit, a third max pooling unit, and a plurality of initial feature extraction units;

[0566] Correspondingly, processing module 1402 inputs the third enhanced feature map into the feature co-operation subunit to obtain the eighth filtered feature map, which is specifically used for:

[0567] The third enhanced feature map is input into the twenty-first initial feature extraction unit to obtain the first collaborative feature map;

[0568] The first collaborative feature map is sequentially input into the second average pooling unit and the third average pooling unit to obtain the second collaborative feature map;

[0569] The first and second collaborative feature maps are concatenated to obtain the third collaborative feature map.

[0570] The third collaborative feature map is sequentially input into the second max pooling unit and the third max pooling unit to obtain the fourth collaborative feature map;

[0571] The third and fourth collaborative feature maps are concatenated to obtain the fifth collaborative feature map;

[0572] The fifth collaborative feature map is input into the twenty-second initial feature extraction unit to obtain the sixth collaborative feature map;

[0573] The first and sixth collaborative feature maps are superimposed to obtain the seventh collaborative feature map.

[0574] The seventh collaborative feature map is input into the twenty-third initial feature extraction unit to obtain the eighth filtered feature map.

[0575] The thermal fault detection device for energy storage equipment provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0576] Figure 15 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 15 As shown, the electronic device 1500 includes: a processor 1501, a memory 1502, and a bus 1503;

[0577] The memory 1502 is used to store the computer-executed instructions of the processor 1501;

[0578] The processor 1501 is configured to execute the technical solutions of any of the foregoing method embodiments by executing computer execution instructions.

[0579] The specific implementation process of processor 1501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0580] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0581] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0582] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0583] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0584] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0585] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0586] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0587] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0588] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0589] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0590] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory (RAM), magnetic disks, or optical disks.

[0591] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0592] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for detecting thermal faults in an energy storage device, characterized in that, include: Acquire a target image of the energy storage device to be tested; The target image is input into the thermal fault detection model to obtain the thermal fault information of the energy storage device to be detected. The thermal fault detection model is determined based on images of multiple energy storage devices and the labeled thermal fault information corresponding to each image. The thermal fault information includes: fault location and fault type.

2. The method according to claim 1, characterized in that, Before inputting the target image into the thermal fault detection model to obtain the thermal fault information of the energy storage device under test, the method further includes: Acquire images of the multiple energy storage devices and the corresponding labeled thermal fault information for each image; The initial thermal fault detection model is trained based on the images of the multiple energy storage devices and the labeled thermal fault information corresponding to each image to obtain the thermal fault detection model.

3. The method according to claim 2, characterized in that, The initial thermal fault detection model includes: a feature extraction unit, a feature enhancement unit, and a detection head unit; Accordingly, the step of training the initial thermal fault detection model based on the images of the multiple energy storage devices and the labeled thermal fault information corresponding to each image to obtain the thermal fault detection model includes: S1, For each energy storage device, the image of the energy storage device is input into the feature extraction unit to obtain multiple initial feature maps; S2, input the multiple initial feature maps into the feature enhancement unit to obtain multiple enhanced feature maps; S3, input the multiple enhanced feature maps into the detection head unit to obtain predicted thermal fault information; S4. Update the parameters of the initial thermal fault detection model based on the predicted thermal fault information and the labeled thermal fault information corresponding to each image. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the thermal fault detection model.

4. The method according to claim 3, characterized in that, The feature extraction unit includes: multiple initial feature extraction units, multiple multi-level energy storage device feature extraction units, and a fine energy storage device feature sensing unit; Accordingly, for each energy storage device, the image of the energy storage device is input into the feature extraction unit to obtain multiple initial feature maps, including: For each energy storage device image, the image of the energy storage device is sequentially input into the first initial feature extraction unit and the second initial feature extraction unit to obtain the first initial feature map; The first initial feature map is input into the feature extraction unit of the first multi-level energy storage device to obtain the second initial feature map; The second initial feature map is sequentially input into the third initial feature extraction unit and the second multi-level energy storage device feature extraction unit to obtain the third initial feature map; The third initial feature map is sequentially input into the fourth initial feature extraction unit and the third multi-level energy storage device feature extraction unit to obtain the fourth initial feature map; The fourth initial feature map is sequentially input into the fifth initial feature extraction unit and the fourth multi-level energy storage device feature extraction unit to obtain the fifth initial feature map; The fifth initial feature map is input into the feature sensing unit of the fine energy storage device to obtain the sixth initial feature map.

5. The method according to claim 4, characterized in that, Each multi-level energy storage device feature extraction unit includes: multiple initial feature extraction units, a first cutting unit, multiple depthwise separable convolution units, and a dual-path energy storage feature extraction subunit; Accordingly, the step of inputting the first initial feature map into the first multi-level energy storage device feature extraction unit to obtain the second initial feature map includes: The first initial feature map is sequentially input into the sixth initial feature extraction unit and the first cutting unit to obtain the seventh initial feature map, the eighth initial feature map, and the ninth initial feature map; The seventh initial feature map is input into the first depthwise separable convolutional unit to obtain the tenth initial feature map; The eighth initial feature map is input into the second depthwise separable convolutional unit to obtain the eleventh initial feature map; The ninth initial feature map is input into the third depthwise separable convolutional unit to obtain the twelfth initial feature map; The seventh initial feature map, the eighth initial feature map, and the ninth initial feature map are concatenated to obtain the thirteenth initial feature map; The thirteenth initial feature map is sequentially input into the fourth, fifth, and sixth depthwise separable convolutional units to obtain the fourteenth initial feature map. The fourteenth initial feature map and the thirteenth initial feature map are superimposed to obtain the fifteenth initial feature map; The fifteenth initial feature map is input into the seventh initial feature extraction unit to obtain the sixteenth initial feature map; The sixteenth initial feature map is input into the dual-path energy storage feature extraction subunit to obtain the second initial feature map.

6. The method according to claim 5, characterized in that, The dual-path energy storage feature extraction subunit includes: multiple initial feature extraction units, a second segmentation unit, a first convolutional block attention unit, and a second convolutional block attention unit; Accordingly, inputting the sixteenth initial feature map into the dual-path energy storage feature extraction subunit to obtain the second initial feature map includes: The sixteenth initial feature map is input into the eighth initial feature extraction unit and the ninth initial feature extraction unit respectively to obtain the first dual-path energy storage feature map and the second dual-path energy storage feature map; The first dual-channel energy storage feature map and the second dual-channel energy storage feature map are superimposed to obtain the third dual-channel energy storage feature map. The third dual-path energy storage feature map is input into the tenth initial feature extraction unit to obtain the fourth dual-path energy storage feature map; The first dual-channel energy storage feature map, the second dual-channel energy storage feature map, and the fourth dual-channel energy storage feature map are spliced ​​together to obtain the fifth dual-channel energy storage feature map. The fifth dual-channel energy storage feature map is input into the second cutting unit to obtain the sixth dual-channel energy storage feature map and the seventh dual-channel energy storage feature map; The sixth dual-path energy storage feature map and the seventh dual-path energy storage feature map are respectively input into the first convolutional block attention unit and the second convolutional block attention unit to obtain the eighth dual-path energy storage feature map and the ninth dual-path energy storage feature map; The eighth dual-channel energy storage feature map and the ninth dual-channel energy storage feature map are spliced ​​together to obtain the tenth dual-channel energy storage feature map. The tenth dual-path energy storage feature map is input into the eleventh initial feature extraction unit to obtain the second initial feature map.

7. The method according to claim 6, characterized in that, The precision energy storage device feature sensing unit includes: a first max pooling unit, a first average pooling unit, a first activation unit, a third slicing unit, multiple initial feature extraction units, and multiple depthwise separable convolution units; Accordingly, the step of inputting the fifth initial feature map into the feature sensing unit of the fine energy storage device to obtain the sixth initial feature map includes: The fifth initial feature map is input into the first max pooling unit, the seventh depthwise separable convolution unit, the eighth depthwise separable convolution unit, and the first average pooling unit, respectively, to obtain the first fine energy storage feature map, the second fine energy storage feature map, the third fine energy storage feature map, and the fourth fine energy storage feature map. The second and third refined energy storage feature maps are input into the twelfth and thirteenth initial feature extraction units, respectively, to obtain the fifth and sixth refined energy storage feature maps. The fifth and sixth fine energy storage feature maps are stitched together to obtain the seventh fine energy storage feature map. The seventh fine energy storage feature map is sequentially input into the fourteenth initial feature extraction unit and the first activation unit to obtain the first fine energy storage weight map; The first fine energy storage weight map is input into the third cutting unit to obtain the second fine energy storage weight map and the third fine energy storage weight map; The fifth fine energy storage feature map and the second fine energy storage weight map are multiplied element by element to obtain the eighth fine energy storage feature map; The sixth fine energy storage feature map and the third fine energy storage weight map are multiplied element by element to obtain the ninth fine energy storage feature map. The first fine energy storage feature map, the fourth fine energy storage feature map, the eighth fine energy storage feature map, and the ninth fine energy storage feature map are stitched together to obtain the tenth fine energy storage feature map; The tenth fine energy storage feature map is input into the fifteenth initial feature extraction unit to obtain the sixth initial feature map.

8. The method according to any one of claims 4-7, characterized in that, The feature enhancement unit includes: a multi-path feature convolution unit, a first upsampling unit, a second upsampling unit, multiple initial feature extraction units, and multiple energy storage device feature filtering units; Accordingly, the step of inputting the plurality of initial feature maps into the feature enhancement unit to obtain a plurality of enhanced feature maps includes: The sixth initial feature map is input into the multi-path feature convolution unit to obtain the first enhanced feature map; The first enhanced feature map is input into the first upsampling unit to obtain the second enhanced feature map; The fourth initial feature map and the second enhanced feature map are concatenated to obtain the third enhanced feature map. The third enhanced feature map is input into the feature filtering unit of the first energy storage device to obtain the fourth enhanced feature map; The fourth enhanced feature map is input into the second upsampling unit to obtain the fifth enhanced feature map; The third initial feature map and the fifth enhanced feature map are concatenated to obtain the sixth enhanced feature map; The sixth enhanced feature map is input into the second energy storage device feature filtering unit to obtain the seventh enhanced feature map; The seventh enhanced feature map is input into the sixteenth initial feature extraction unit to obtain the eighth enhanced feature map; The fourth and eighth enhanced feature maps are concatenated to obtain the ninth enhanced feature map. The ninth enhanced feature map is input into the feature filtering unit of the third energy storage device to obtain the tenth enhanced feature map; The tenth enhanced feature map is input into the seventeenth initial feature extraction unit to obtain the eleventh enhanced feature map; The first enhanced feature map and the eleventh enhanced feature map are concatenated to obtain the twelfth enhanced feature map. The twelfth enhanced feature map is input into the fourth energy storage device feature filtering unit to obtain the thirteenth enhanced feature map.

9. The method according to claim 8, characterized in that, Each energy storage device feature filtering unit includes: a feature coordination subunit, a ninth depthwise separable convolution unit, a batch normalization unit, a second activation unit, a fourth slicing unit, and multiple initial feature extraction units; Accordingly, inputting the third enhanced feature map into the first energy storage device feature filtering unit to obtain the fourth enhanced feature map includes: The third enhanced feature map is input into the eighteenth initial feature extraction unit and the nineteenth initial feature extraction unit respectively to obtain the first filtered feature map and the second filtered feature map; The first and second filtered feature maps are superimposed to obtain the third filtered feature map. The third filtered feature map is sequentially input into the ninth depthwise separable convolutional unit, the batch normalization unit, and the second activation unit to obtain the first filtered weight map; The first filter weight map is input into the fourth cutting unit to obtain the second filter weight map and the third filter weight map; The first filter feature map and the second filter weight map are multiplied element by element to obtain the fourth filter feature map; The second filter feature map and the third filter weight map are multiplied element by element to obtain the fifth filter feature map; The fourth and fifth filter feature maps are multiplied element-wise to obtain the sixth filter feature map. The fourth and fifth filter feature maps are added element by element to obtain the seventh filter feature map; The third enhanced feature map is input into the feature collaboration subunit to obtain the eighth filtered feature map; The third enhanced feature map, the sixth filtered feature map, the seventh filtered feature map, and the eighth filtered feature map are concatenated to obtain the ninth filtered feature map. The ninth filtered feature map is input into the twentieth initial feature extraction unit to obtain the fourth enhanced feature map.

10. The method according to claim 9, characterized in that, The feature collaboration subunit includes: a second average pooling unit, a third average pooling unit, a second max pooling unit, a third max pooling unit, and multiple initial feature extraction units; Accordingly, the step of inputting the third enhanced feature map into the feature co-operation subunit to obtain the eighth filtered feature map includes: The third enhanced feature map is input into the twenty-first initial feature extraction unit to obtain the first collaborative feature map; The first collaborative feature map is sequentially input into the second average pooling unit and the third average pooling unit to obtain the second collaborative feature map; The first collaborative feature map and the second collaborative feature map are concatenated to obtain the third collaborative feature map. The third collaborative feature map is sequentially input into the second max pooling unit and the third max pooling unit to obtain the fourth collaborative feature map; The third and fourth collaborative feature maps are concatenated to obtain the fifth collaborative feature map. The fifth collaborative feature map is input into the twenty-second initial feature extraction unit to obtain the sixth collaborative feature map; The first collaborative feature map and the sixth collaborative feature map are superimposed to obtain the seventh collaborative feature map. The seventh collaborative feature map is input into the twenty-third initial feature extraction unit to obtain the eighth filtered feature map.

Images