Method and apparatus for detecting defects in energy storage devices

Abstract

Embodiments of the present application provide a kind of defect detection method and device of energy storage equipment. The method first obtains the target image of the energy storage equipment to be detected, then the target image is input into the energy storage equipment defect detection model, and the defect information of the energy storage equipment to be detected is obtained, wherein the energy storage equipment defect detection model is determined based on the image of multiple energy storage equipment and the corresponding labeled defect information of each image, and the defect information includes defect type and defect level. By using the technical solution of the present application, the target image of the energy storage equipment to be detected is analyzed by using the energy storage equipment defect detection model trained based on the image of multiple energy storage equipment and its labeled defect information, which can effectively improve the accuracy and stability of micro defect identification, realize accurate defect detection of the energy storage equipment to be detected, and is beneficial to realize efficient inspection and subsequent targeted operation and maintenance management of energy storage equipment.

Description

Technical Field

[0001] This application relates to the field of computer vision and deep learning image processing technology, and in particular to a method and apparatus for defect detection of energy storage devices. Background Technology

[0002] With the rapid development of new energy sources, new energy storage devices are widely used in scenarios such as distribution network self-synchronizing voltage sources, distributed energy collaborative aggregation, and power grid frequency and voltage regulation. Their safe and stable operation is directly related to the reliability of power supply and power quality. However, under complex operating conditions, energy storage devices are prone to local overheating or even thermal runaway due to internal short circuits, poor contact, insulation aging, and thermal management failures, leading to equipment damage, power loss, and even major safety accidents such as fires. Therefore, defect detection of energy storage devices is necessary.

[0003] Current technologies mainly rely on regular manual inspections or monitoring by fixed-point sensors. For example, manual inspections depend on maintenance personnel observing and judging the appearance of equipment, connection points, and abnormal signs, while fixed sensors obtain temperature, voltage, or local status data through preset points to provide fault warnings.

[0004] However, in the scenario of energy storage equipment inspection, the field of view of energy storage equipment is large while the defect scale is small. Existing technologies are prone to losing information about minute defects when performing defect detection, resulting in low accuracy of defect detection. Summary of the Invention

[0005] This application provides a defect detection method and apparatus for energy storage devices to improve the accuracy of defect detection.

[0006] In a first aspect, embodiments of this application provide a defect detection method for an energy storage device, comprising:

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

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

[0009] In one or more embodiments, before inputting the target image into the energy storage device defect detection model to obtain the defect 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 defect information for each image;

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

[0012] In one or more embodiments, the initial defect detection model includes: a small target perception multi-scale feature extraction network unit, a dynamic environment adaptation feature fusion network unit, an asymmetric spatial refinement unit, and a detection head unit;

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

[0014] S1, For each energy storage device image, the image of the energy storage device is input into the small target perception multi-scale feature extraction network unit to obtain multiple small target perception feature maps;

[0015] S2, input the multiple small target perception feature maps into the dynamic environment adaptation feature fusion network unit to obtain the target environment adaptation fusion feature map;

[0016] S3, input the target environment adaptation fusion feature map into the asymmetric spatial refinement unit to obtain the target spatial refinement feature map;

[0017] S4, input the refined feature map of the target space into the detection head unit to obtain the estimated defect information;

[0018] S5, update the parameters of the initial defect detection model according to the estimated defect information and the labeled defect information corresponding to each image, repeat steps S1-S5 until the preset training stopping condition is met, and obtain the energy storage device defect detection model.

[0019] In one or more embodiments, the small target perception multi-scale feature extraction network unit includes: multiple adaptive small target enhancement detection units, fast spatial pyramid pooling units, multiple initial feature extraction units, and multiple deep feature extraction units;

[0020] Accordingly, for each energy storage device, the image of the energy storage device is input into the small target perception multi-scale feature extraction network unit to obtain multiple small target perception feature maps, including:

[0021] 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 first depth feature extraction unit for feature extraction to obtain the first small target perception feature map;

[0022] The first small target perception feature map and the image of the energy storage device are input into the first adaptive small target enhancement and detection unit for feature enhancement to obtain the second small target perception feature map.

[0023] The first small target perception feature map is sequentially input into the second initial feature extraction unit and the second deep feature extraction unit for feature extraction to obtain the third small target perception feature map.

[0024] The second small target perception feature map and the third small target perception feature map are input into the second adaptive small target enhancement and detection unit for feature enhancement to obtain the fourth small target perception feature map.

[0025] The third small target perception feature map is sequentially input into the third initial feature extraction unit and the fast spatial pyramid pooling unit for feature extraction and feature aggregation to obtain the fifth small target perception feature map.

[0026] The fourth and fifth small target perception feature maps are input into the third adaptive small target enhancement detection unit for feature enhancement to obtain the sixth small target perception feature map.

[0027] In one or more embodiments, each adaptive small target enhancement detection unit includes: a cutting unit, a first convolutional unit, a second convolutional unit, a spatial structure perception feature calibration unit, a first dilated convolutional unit, and a second dilated convolutional unit;

[0028] Accordingly, the step of inputting the first small target perception feature map and the image of the energy storage device into the first adaptive small target enhancement and detection unit for feature enhancement to obtain the second small target perception feature map includes:

[0029] The first small target perception feature map and the image of the energy storage device are superimposed and fused to obtain the seventh small target perception feature map.

[0030] The seventh small target perception feature map is input into the cutting unit to obtain the eighth small target perception feature map, the ninth small target perception feature map, and the tenth small target perception feature map;

[0031] The eighth small target perception feature map is input into the first convolutional unit to obtain the eleventh small target perception feature map;

[0032] The ninth small target perception feature map is input into the second convolutional unit to obtain the twelfth small target perception feature map;

[0033] The tenth small target perception feature map is input into the spatial structure perception feature calibration unit to obtain the thirteenth small target perception feature map;

[0034] The eleventh small target perception feature map, the twelfth small target perception feature map, and the thirteenth small target perception feature map are stitched together to obtain the fourteenth small target perception feature map.

[0035] The fourteenth small target perception feature map is input into the first dilated convolution unit and the second dilated convolution unit respectively to obtain the second small target perception feature map.

[0036] In one or more embodiments, the spatial structure-aware feature calibration unit includes: a horizontal average pooling unit, a vertical average pooling unit, a third convolution unit, a fourth convolution unit, a first activation unit, and a second activation unit;

[0037] Accordingly, the step of inputting the tenth small target perception feature map into the spatial structure perception feature calibration unit to obtain the thirteenth small target perception feature map includes:

[0038] The tenth small target perception feature map is input into the horizontal average pooling unit and the vertical average pooling unit respectively to obtain the sixteenth small target perception feature map;

[0039] The sixteenth small target perception feature map is sequentially input into the third convolutional unit, the first activation unit, the fourth convolutional unit, and the second activation unit to obtain the attention weight map;

[0040] The thirteenth small target perception feature map is obtained by multiplying the tenth small target perception feature map and the attention weight map.

[0041] In one or more embodiments, the dynamic environment adaptation feature fusion network unit includes: multiple depthwise separable convolutional units, multiple dynamic environment-aware feature adaptive correction units, a first average pooling unit, a fifth convolutional unit, a downsampling unit, a normalization and activation unit, a third dilated convolutional unit, and an upsampling unit.

[0042] Accordingly, the step of inputting the plurality of small target perception feature maps into the dynamic environment adaptation feature fusion network unit to obtain the target environment adaptation fusion feature map includes:

[0043] The second small target perception feature map is sequentially input into the first depthwise separable convolutional unit, the first dynamic environment perception feature adaptive correction unit, the first average pooling unit, the fifth convolutional unit, and the downsampling unit to obtain the first environment adaptive fusion feature map;

[0044] The fourth small target perception feature map is sequentially input into the second depthwise separable convolutional unit, the second dynamic environment perception feature adaptive correction unit, and the normalization and activation unit to obtain the second environment adaptive fusion feature map;

[0045] The sixth small target perception feature map is sequentially input into the third dynamic environment perception feature adaptive correction unit, the third dilated convolution unit, and the upsampling unit to obtain the third environment adaptive fusion feature map;

[0046] The first environmental adaptation fusion feature map, the second environmental adaptation fusion feature map, and the third environmental adaptation fusion feature map are stitched together to obtain the target environmental adaptation fusion feature map.

[0047] In one or more embodiments, each dynamic environment-aware feature adaptive correction unit includes: a sixth convolutional unit, a third depthwise separable convolutional unit, a batch normalization unit, a second average pooling unit, a third activation unit, a dynamic convolutional unit, and an efficient multi-scale attention unit.

[0048] Accordingly, the step of sequentially inputting the second small target perception feature map into the first depthwise separable convolutional unit, the first dynamic environment perception feature adaptive correction unit, the first average pooling unit, the fifth convolutional unit, and the downsampling unit to obtain the first environment-adaptive fusion feature map includes:

[0049] The second small target perception feature map is input into the first depthwise separable convolutional unit to obtain the fourth environment-adaptive fusion feature map.

[0050] The fourth environmental adaptation fusion feature map is sequentially input into the sixth convolutional unit and the batch normalization unit to obtain the fifth environmental adaptation fusion feature map.

[0051] The fourth environment-adaptive fusion feature map is sequentially input into the third depthwise separable convolutional unit and the second average pooling unit to obtain the sixth environment-adaptive fusion feature map.

[0052] The fifth and sixth environmental adaptation fusion feature maps are overlaid and then input into the third activation unit to obtain the environmental adaptation fusion weight map.

[0053] The sixth environment adaptation fusion feature map is sequentially input into the dynamic convolutional unit and the efficient multi-scale attention unit to obtain the seventh environment adaptation fusion feature map;

[0054] The seventh environmental adaptation fusion feature map and the environmental adaptation fusion weight map are multiplied to obtain the eighth environmental adaptation fusion feature map;

[0055] The eighth environment adaptation fusion feature map is sequentially input into the first average pooling unit, the fifth convolution unit, and the downsampling unit to obtain the first environment adaptation fusion feature map.

[0056] In one or more embodiments, the asymmetric spatial refinement unit includes: a first asymmetric convolutional layer, a second asymmetric convolutional layer, a third average pooling unit, a fourth average pooling unit, a max pooling unit, a seventh convolutional unit, and an eighth convolutional unit.

[0057] Accordingly, the step of inputting the target environment adaptation fusion feature map into the asymmetric spatial refinement unit to obtain the target spatial refinement feature map includes:

[0058] The target environment adaptation fusion feature map is sequentially input into the first asymmetric convolutional layer and the second asymmetric convolutional layer to extract features in the horizontal and vertical directions, thereby obtaining the first spatial refinement feature map;

[0059] The first spatially refined feature map is sequentially input into the third average pooling unit, the max pooling unit, the fourth average pooling unit, and the seventh convolutional unit to obtain the second spatially refined feature map.

[0060] The first spatial refinement feature map and the second spatial refinement feature map are superimposed to obtain the third spatial refinement feature map;

[0061] The third spatial refinement feature map is input into the eighth convolutional unit to obtain the target spatial refinement feature map.

[0062] Secondly, embodiments of this application provide a defect detection device for an energy storage device, comprising:

[0063] The acquisition module is used to acquire the target image of the energy storage device to be tested;

[0064] The processing module is used to input the target image into the energy storage device defect detection model to obtain the defect information of the energy storage device to be detected. The energy storage device defect detection model is determined based on images of multiple energy storage devices and the labeled defect information corresponding to each image. The defect information includes: defect type and defect level.

[0065] In one or more embodiments, before inputting the target image into the energy storage device defect detection model to obtain the defect information of the energy storage device to be detected, the processing module is further configured to:

[0066] Acquire images of the multiple energy storage devices and the corresponding labeled defect information for each image;

[0067] The initial defect detection model is trained based on the images of the multiple energy storage devices and the labeled defect information corresponding to each image to obtain the energy storage device defect detection model.

[0068] In one or more embodiments, the initial defect detection model includes: a small target perception multi-scale feature extraction network unit, a dynamic environment adaptation feature fusion network unit, an asymmetric spatial refinement unit, and a detection head unit;

[0069] Accordingly, the processing module trains an initial defect detection model based on the images of the multiple energy storage devices and the labeled defect information corresponding to each image to obtain the energy storage device defect detection model, specifically for:

[0070] S1, For each energy storage device image, the image of the energy storage device is input into the small target perception multi-scale feature extraction network unit to obtain multiple small target perception feature maps;

[0071] S2, input the multiple small target perception feature maps into the dynamic environment adaptation feature fusion network unit to obtain the target environment adaptation fusion feature map;

[0072] S3, input the target environment adaptation fusion feature map into the asymmetric spatial refinement unit to obtain the target spatial refinement feature map;

[0073] S4, input the refined feature map of the target space into the detection head unit to obtain the estimated defect information;

[0074] S5, update the parameters of the initial defect detection model according to the estimated defect information and the labeled defect information corresponding to each image, repeat steps S1-S5 until the preset training stopping condition is met, and obtain the energy storage device defect detection model.

[0075] In one or more embodiments, the small target perception multi-scale feature extraction network unit includes: multiple adaptive small target enhancement detection units, fast spatial pyramid pooling units, multiple initial feature extraction units, and multiple deep feature extraction units;

[0076] Accordingly, the processing module, for each energy storage device image, inputs the image of the energy storage device into the small target perception multi-scale feature extraction network unit to obtain multiple small target perception feature maps, specifically used for:

[0077] 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 first depth feature extraction unit for feature extraction to obtain the first small target perception feature map;

[0078] The first small target perception feature map and the image of the energy storage device are input into the first adaptive small target enhancement and detection unit for feature enhancement to obtain the second small target perception feature map.

[0079] The first small target perception feature map is sequentially input into the second initial feature extraction unit and the second deep feature extraction unit for feature extraction to obtain the third small target perception feature map.

[0080] The second small target perception feature map and the third small target perception feature map are input into the second adaptive small target enhancement and detection unit for feature enhancement to obtain the fourth small target perception feature map.

[0081] The third small target perception feature map is sequentially input into the third initial feature extraction unit and the fast spatial pyramid pooling unit for feature extraction and feature aggregation to obtain the fifth small target perception feature map.

[0082] The fourth and fifth small target perception feature maps are input into the third adaptive small target enhancement detection unit for feature enhancement to obtain the sixth small target perception feature map.

[0083] In one or more embodiments, each adaptive small target enhancement detection unit includes: a cutting unit, a first convolutional unit, a second convolutional unit, a spatial structure perception feature calibration unit, a first dilated convolutional unit, and a second dilated convolutional unit;

[0084] Accordingly, the processing module inputs the first small target perception feature map and the image of the energy storage device into the first adaptive small target enhancement and detection unit for feature enhancement to obtain the second small target perception feature map, specifically for:

[0085] The first small target perception feature map and the image of the energy storage device are superimposed and fused to obtain the seventh small target perception feature map.

[0086] The seventh small target perception feature map is input into the cutting unit to obtain the eighth small target perception feature map, the ninth small target perception feature map, and the tenth small target perception feature map;

[0087] The eighth small target perception feature map is input into the first convolutional unit to obtain the eleventh small target perception feature map;

[0088] The ninth small target perception feature map is input into the second convolutional unit to obtain the twelfth small target perception feature map;

[0089] The tenth small target perception feature map is input into the spatial structure perception feature calibration unit to obtain the thirteenth small target perception feature map;

[0090] The eleventh small target perception feature map, the twelfth small target perception feature map, and the thirteenth small target perception feature map are stitched together to obtain the fourteenth small target perception feature map.

[0091] The fourteenth small target perception feature map is input into the first dilated convolution unit and the second dilated convolution unit respectively to obtain the second small target perception feature map.

[0092] In one or more embodiments, the spatial structure-aware feature calibration unit includes: a horizontal average pooling unit, a vertical average pooling unit, a third convolution unit, a fourth convolution unit, a first activation unit, and a second activation unit;

[0093] Correspondingly, the processing module inputs the tenth small target perception feature map into the spatial structure perception feature calibration unit to obtain the thirteenth small target perception feature map, specifically for:

[0094] The tenth small target perception feature map is input into the horizontal average pooling unit and the vertical average pooling unit respectively to obtain the sixteenth small target perception feature map;

[0095] The sixteenth small target perception feature map is sequentially input into the third convolutional unit, the first activation unit, the fourth convolutional unit, and the second activation unit to obtain the attention weight map;

[0096] The thirteenth small target perception feature map is obtained by multiplying the tenth small target perception feature map and the attention weight map.

[0097] In one or more embodiments, the dynamic environment adaptation feature fusion network unit includes: multiple depthwise separable convolutional units, multiple dynamic environment-aware feature adaptive correction units, a first average pooling unit, a fifth convolutional unit, a downsampling unit, a normalization and activation unit, a third dilated convolutional unit, and an upsampling unit.

[0098] Accordingly, the processing module inputs the plurality of small target perception feature maps into the dynamic environment adaptation feature fusion network unit to obtain a target environment adaptation fusion feature map, specifically for:

[0099] The second small target perception feature map is sequentially input into the first depthwise separable convolutional unit, the first dynamic environment perception feature adaptive correction unit, the first average pooling unit, the fifth convolutional unit, and the downsampling unit to obtain the first environment adaptive fusion feature map;

[0100] The fourth small target perception feature map is sequentially input into the second depthwise separable convolutional unit, the second dynamic environment perception feature adaptive correction unit, and the normalization and activation unit to obtain the second environment adaptive fusion feature map;

[0101] The sixth small target perception feature map is sequentially input into the third dynamic environment perception feature adaptive correction unit, the third dilated convolution unit, and the upsampling unit to obtain the third environment adaptive fusion feature map;

[0102] The first environmental adaptation fusion feature map, the second environmental adaptation fusion feature map, and the third environmental adaptation fusion feature map are stitched together to obtain the target environmental adaptation fusion feature map.

[0103] In one or more embodiments, each dynamic environment-aware feature adaptive correction unit includes: a sixth convolutional unit, a third depthwise separable convolutional unit, a batch normalization unit, a second average pooling unit, a third activation unit, a dynamic convolutional unit, and an efficient multi-scale attention unit.

[0104] Accordingly, the processing module sequentially inputs the second small target perception feature map into the first depthwise separable convolutional unit, the first dynamic environment perception feature adaptive correction unit, the first average pooling unit, the fifth convolutional unit, and the downsampling unit to obtain the first environment-adaptive fusion feature map, specifically used for:

[0105] The second small target perception feature map is input into the first depthwise separable convolutional unit to obtain the fourth environment-adaptive fusion feature map.

[0106] The fourth environmental adaptation fusion feature map is sequentially input into the sixth convolutional unit and the batch normalization unit to obtain the fifth environmental adaptation fusion feature map.

[0107] The fourth environment-adaptive fusion feature map is sequentially input into the third depthwise separable convolutional unit and the second average pooling unit to obtain the sixth environment-adaptive fusion feature map.

[0108] The fifth and sixth environmental adaptation fusion feature maps are overlaid and then input into the third activation unit to obtain the environmental adaptation fusion weight map.

[0109] The sixth environment adaptation fusion feature map is sequentially input into the dynamic convolutional unit and the efficient multi-scale attention unit to obtain the seventh environment adaptation fusion feature map;

[0110] The seventh environmental adaptation fusion feature map and the environmental adaptation fusion weight map are multiplied to obtain the eighth environmental adaptation fusion feature map;

[0111] The eighth environment adaptation fusion feature map is sequentially input into the first average pooling unit, the fifth convolution unit, and the downsampling unit to obtain the first environment adaptation fusion feature map.

[0112] In one or more embodiments, the asymmetric spatial refinement unit includes: a first asymmetric convolutional layer, a second asymmetric convolutional layer, a third average pooling unit, a fourth average pooling unit, a max pooling unit, a seventh convolutional unit, and an eighth convolutional unit.

[0113] Accordingly, the processing module inputs the target environment adaptation fusion feature map into the asymmetric spatial refinement unit to obtain a target spatial refinement feature map, specifically for:

[0114] The target environment adaptation fusion feature map is sequentially input into the first asymmetric convolutional layer and the second asymmetric convolutional layer to extract features in the horizontal and vertical directions, thereby obtaining the first spatial refinement feature map;

[0115] The first spatially refined feature map is sequentially input into the third average pooling unit, the max pooling unit, the fourth average pooling unit, and the seventh convolutional unit to obtain the second spatially refined feature map.

[0116] The first spatial refinement feature map and the second spatial refinement feature map are superimposed to obtain the third spatial refinement feature map;

[0117] The third spatial refinement feature map is input into the eighth convolutional unit to obtain the target spatial refinement feature map.

[0118] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0119] The memory stores computer-executed instructions;

[0120] 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.

[0121] 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.

[0122] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, is used to implement a defect detection method for an energy storage device as described in the first aspect and various possible implementations of the first aspect.

[0123] This application provides a method and apparatus for defect detection of energy storage devices. The method first acquires a target image of the energy storage device to be detected, then inputs the target image into an energy storage device defect detection model to obtain defect information of the energy storage device. The energy storage device defect detection model is determined based on images of multiple energy storage devices and the corresponding labeled defect information for each image. The defect information includes defect type and defect level. By acquiring a target image of the energy storage device to be detected and inputting it into an energy storage device defect detection model determined based on multiple energy storage device images and their corresponding labeled defect information, the method can directly output the defect type and defect level of the energy storage device to be detected, thereby improving the accuracy of defect detection in complex energy storage inspection scenarios and reducing the probability of missed and false detections. Attached Figure Description

[0124] 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.

[0125] Figure 1 A flowchart illustrating the defect detection method for energy storage devices provided in this application embodiment. Figure 1 ;

[0126] Figure 2 A flowchart illustrating the defect detection method for energy storage devices provided in this application embodiment. Figure 2 ;

[0127] Figure 3 This is a schematic diagram of the structure of the initial defect detection model provided in the embodiments of this application;

[0128] Figure 4 A flowchart illustrating the defect detection method for energy storage devices provided in this application embodiment. Figure 3 ;

[0129] Figure 5 This is a schematic diagram of the structure of a small target perception multi-scale feature extraction network unit provided in an embodiment of this application;

[0130] Figure 6 This is a schematic diagram of the structure of the adaptive small target enhancement detection unit provided in the embodiments of this application;

[0131] Figure 7 A schematic diagram of the spatial structure sensing feature calibration unit provided in the embodiments of this application;

[0132] Figure 8 A schematic diagram of the structure of the dynamic environment adaptation feature fusion network unit provided in the embodiments of this application;

[0133] Figure 9A schematic diagram of the structure of the dynamic environment perception feature adaptive correction unit provided in the embodiments of this application;

[0134] Figure 10 This is a schematic diagram of the structure of the asymmetric spatial refinement unit provided in the embodiments of this application;

[0135] Figure 11 This is a schematic diagram of the structure of the energy storage device defect detection model provided in the embodiments of this application;

[0136] Figure 12 A schematic diagram of the structure of the defect detection device for energy storage equipment provided in the embodiments of this application;

[0137] Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0138] 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

[0139] 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.

[0140] Before introducing the embodiments of this application, the application background of the embodiments of this application will be explained first:

[0141] With the rapid development of new energy sources, new energy storage devices are widely used in scenarios such as distribution network self-synchronizing voltage sources, distributed energy collaborative aggregation, and power grid frequency and voltage regulation. Their safe and stable operation is directly related to the reliability of power supply and power quality. However, under complex operating conditions, energy storage devices are prone to local overheating or even thermal runaway due to internal short circuits, poor contact, insulation aging, and thermal management failures, leading to equipment damage, power loss, and even major safety accidents such as fires. Therefore, defect detection of energy storage devices is necessary.

[0142] Current technologies mainly rely on regular manual inspections and monitoring by fixed sensors. For example, manual inspections depend on maintenance personnel to observe and judge the appearance of the equipment, connection points, and abnormal signs, while fixed sensors obtain temperature, voltage, or local status data through preset points to provide fault warnings.

[0143] However, in the scenario of energy storage equipment inspection, the field of view of energy storage equipment is large while the defect scale is small. Existing technologies are prone to losing information about minute defects when performing defect detection, resulting in low accuracy of defect detection.

[0144] The defect 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 defect detection during energy storage device inspections. Defect detection of energy storage devices mainly involves observing the appearance of the device or analyzing its status data to identify abnormal areas or data and determine defect information. If image detection of the energy storage device and in-depth image analysis can be performed to accurately identify minute defect features, accurate defect detection can be achieved. Therefore, this application's embodiments consider constructing an energy storage device defect detection model based on multiple energy storage device images and their corresponding labeled defect information. This model is used to perform defect detection on the target image of the energy storage device under inspection, directly outputting the defect type and defect level of the device, thus achieving accurate defect detection of energy storage devices in complex energy storage inspection scenarios.

[0145] 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.

[0146] 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.

[0147] Figure 1 A flowchart illustrating the defect detection method for energy storage devices provided in this application embodiment. Figure 1 .like Figure 1 As shown, the defect detection method for this energy storage device includes the following steps:

[0148] S101. Obtain the target image of the energy storage device to be tested.

[0149] In this step, a target image of the energy storage device to be tested is acquired, providing a data foundation for subsequent defect detection of the energy storage device.

[0150] For example, energy storage devices, as the target objects for defect detection during inspections, can specifically be battery modules, power conversion systems (PCS), combiner boxes, or other devices that perform energy storage, conversion, and combination functions within an energy storage power station.

[0151] In one possible implementation, the target image is the image data corresponding to the energy storage device to be inspected. It can be a single frame image or at least one frame from multiple continuously acquired frames. The target image can include the overall appearance information and / or local area information of the energy storage device to be inspected for subsequent defect detection.

[0152] In addition, target images of the energy storage devices to be inspected can be acquired through industrial cameras, surveillance cameras, or visual acquisition modules integrated into intelligent inspection robots.

[0153] S102. Input the target image into the energy storage device defect detection model to obtain the defect information of the energy storage device to be detected;

[0154] The energy storage device defect detection model is determined based on images of multiple energy storage devices and the corresponding labeled defect information for each image. The defect information includes: defect type and defect level.

[0155] In this step, a defect detection model for energy storage devices is determined based on images of multiple energy storage devices and the corresponding labeled defect information for each image. The target image is then input into the defect detection model for defect analysis to obtain the defect type and defect level of the energy storage device to be detected.

[0156] For example, images of multiple energy storage devices may include images of different devices such as battery modules, energy storage converters, and combiner boxes, captured from different shooting angles, under different lighting conditions, with different environmental interferences, and with different defect morphologies.

[0157] The defect information labeling is used to characterize the defect type and defect level corresponding to images of multiple energy storage devices.

[0158] Defect type is a category attribute used to characterize the category to which the detected anomaly belongs. For example, the condition of the battery module's tabs is classified as normal, oxidized, or broken; the condition of the wiring terminals is classified as normal, loose, or overheated and discolored; and the appearance of the equipment is classified as normal, bulging, or leaking.

[0159] Defect level is a graded attribute used to characterize the severity of a defect. For example, defect levels can be divided into mild, moderate, and severe levels. Mild level corresponds to the initial stage of local anomaly, moderate level corresponds to the expansion of the anomaly range or the enhancement of structural impact, and severe level corresponds to a significant risk of structural failure.

[0160] In one possible implementation, the energy storage device defect detection model can be a deep learning model that predetermines images of multiple energy storage devices and corresponding labeled defect information. By performing feature analysis on the input target image, it outputs a detection image of the energy storage device to be detected, including a bounding box. The detection image is labeled with the defect location, defect type, and confidence score. The defect level is further determined by comparing the confidence score with a preset threshold.

[0161] This application provides a defect detection method for energy storage devices. First, a target image of the energy storage device to be detected is acquired. Then, the target image is input into an energy storage device defect detection model to obtain defect information of the energy storage device to be detected. The energy storage device defect detection model is determined based on images of multiple energy storage devices and the corresponding labeled defect information for each image. The defect information includes: defect type and defect level. In this embodiment, by acquiring a target image of the energy storage device to be detected and inputting it into an energy storage device defect detection model determined based on multiple energy storage device images and their corresponding labeled defect information for defect detection, the defect type and defect level of the energy storage device to be detected can be directly output, thereby improving the accuracy of defect detection in complex energy storage inspection scenarios and reducing the probability of missed and false detections.

[0162] Based on the above embodiments, Figure 2 A flowchart illustrating the defect detection method for energy storage devices provided in this application embodiment. Figure 2 .like Figure 2 As shown, prior to step S120 above, the defect detection method for this energy storage device further includes the following steps:

[0163] S201. Obtain images of multiple energy storage devices and the corresponding defect information for each image.

[0164] In this step, images of multiple energy storage devices are acquired to provide input samples for subsequent model training, and labeled defect information corresponding to each image is obtained to provide supervision targets for model training.

[0165] In one possible implementation, intelligent inspection robots equipped with high-definition industrial cameras and fixed-point monitoring terminals can be used to acquire multi-angle, high-resolution images of key energy storage devices inside the energy storage power station, including battery modules, energy storage converters, combiner boxes, and circuit breakers, under different time periods such as strong daylight and supplemental nightlight, different shooting angles, and different environmental interferences such as reflections and dust, thereby obtaining images of multiple energy storage devices.

[0166] Furthermore, images of multiple energy storage devices can be manually annotated in a two-tiered manner or annotated using an open-source annotation platform. The annotation strategy employs a nested bounding box approach: First, for each image, the functional units of the energy storage device within that image are selected and annotated as a whole; then, within the annotated bounding box of the energy storage device, specific key components or defect areas are precisely selected, resulting in an image dataset of energy storage devices with annotated defect information; finally, a bilinear interpolation algorithm can be used to standardize the size of all images in the energy storage device image dataset, uniformly adjusting them to a preset pixel size, such as 640×640 pixels, to adapt to the input requirements of subsequent models.

[0167] To ensure the model can handle the complex and varied actual operating conditions of energy storage power stations, the construction of the energy storage equipment image dataset focused on collecting samples under challenging scenarios. For example, it simulated corrosion scenarios by capturing images of battery cabinets in high-humidity, high-salt-fog coastal environments, and simulated glare interference scenarios by capturing images of PCS cabinets under direct sunlight. The final energy storage equipment image dataset covers different types of energy storage devices, including lithium iron phosphate battery clusters and flow battery tanks. Furthermore, the defect types in the energy storage equipment image dataset include: tab defects, terminal defects, bulging, and leakage. The dataset also includes images under normal operating conditions.

[0168] In one possible implementation, the image dataset of energy storage devices can be randomly divided into a training set, a validation set, and a test set according to a preset ratio, such as 80%:10%:10%. The training set is used to train the model, the validation set is used to verify the defect detection accuracy of the trained energy storage device defect detection model, and the test set is used to test the performance of the energy storage device defect detection model.

[0169] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0170] S202. Train the initial defect detection model based on images of multiple energy storage devices and the labeled defect information corresponding to each image to obtain the energy storage device defect detection model.

[0171] In this step, images of multiple energy storage devices are input into the initial defect detection model for defect identification, and the defect information corresponding to multiple energy storage devices is output. The model parameters of the initial defect detection model are updated based on the defect information corresponding to multiple energy storage devices and the labeled defect information corresponding to each image, thereby obtaining the energy storage device defect detection model.

[0172] In one possible implementation, Figure 3 This is a schematic diagram of the structure of the initial defect detection model provided in an embodiment of this application. Figure 3 As shown, the initial defect detection model includes: Small-Object Aware Multi-scale Feature Extraction Backbone (SOAMFE), Dynamic Environment-Adaptive Feature Fusion Neck (DEAFFN), Asymmetric Spatial Refinement Module (ASR), and a detection head unit;

[0173] Accordingly, Figure 4 A flowchart illustrating the defect detection method for energy storage devices provided in this application embodiment. Figure 3 .like Figure 4 As shown, one possible implementation of step S202 above includes the following steps:

[0174] S1, For each energy storage device image, input the energy storage device image into the small target perception multi-scale feature extraction network unit to obtain multiple small target perception feature maps;

[0175] For example, for each energy storage device image, the current energy storage device image is input into the small target perception multi-scale feature extraction network unit to perform fine feature extraction of small targets in the energy storage device image, to prevent the loss of features of small parts in feature extraction, and to enhance the features of small targets, so as to retain the key spatial location information of small target features while suppressing background noise, and finally obtain multiple small target perception feature maps.

[0176] In one possible implementation, Figure 5 This is a schematic diagram of the structure of a small target perception multi-scale feature extraction network unit provided in an embodiment of this application. (See attached diagram.) Figure 5 As shown, the Small-ObjectAware Multi-scale Feature Extraction Backbone (SOAMFE) unit includes: multiple Adaptive Small-Object Detection Enhancement Module (ASODEM) units, Spatial Pyramid Pooling-Fast (SPPF) units, multiple initial feature extraction units, and multiple deep feature extraction units;

[0177] The multiple adaptive small target enhancement detection units include: a first adaptive small target enhancement detection unit, a second adaptive small target enhancement detection unit, and a third adaptive small target enhancement detection unit;

[0178] The multiple initial feature extraction units include: a first initial feature extraction unit, a second initial feature extraction unit, and a third initial feature extraction unit;

[0179] The multiple deep feature extraction units include: a first deep feature extraction unit and a second deep feature extraction unit.

[0180] 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, and the deep feature extraction unit can be a C2f (C2with better gradient Flow) module.

[0181] Accordingly, one possible implementation of step S1 above includes the following steps:

[0182] S11. For the image of each energy storage device, the image of the energy storage device is sequentially input into the first initial feature extraction unit and the first deep feature extraction unit for feature extraction to obtain the first small target perception feature map;

[0183] For example, the image F1 of the energy storage device is input into the first CBS module for preliminary feature extraction and downsampling to obtain a small target perception feature map F2. Then, the small target perception feature map F2 is input into the first C2f module for deep feature extraction to obtain a first target perception feature map F3.

[0184] S12. Input the first small target perception feature map and the image of the energy storage device into the first adaptive small target enhancement and detection unit for feature enhancement to obtain the second small target perception feature map;

[0185] For example, the first small target perception feature map F3 and the image F1 of the energy storage device are input into the first ASODEM. The semantic information of the first small target perception feature map F3 is used to perform spatial calibration and enhancement on the image F1 of the energy storage device to obtain the second small target perception feature map F4.

[0186] S13. Input the first small target perception feature map into the second initial feature extraction unit and the second deep feature extraction unit in sequence to extract features and obtain the third small target perception feature map.

[0187] For example, the first small target perception feature map F3 is input into the second CBS module for secondary downsampling to obtain the small target perception feature map F5. Then, the small target perception feature map F5 is input into the second C2f module for deep feature extraction to obtain the third target perception feature map F6.

[0188] S14. Input the second small target perception feature map and the third small target perception feature map into the second adaptive small target enhancement detection unit for feature enhancement to obtain the fourth small target perception feature map.

[0189] For example, the second small target perception feature map F4 and the third small target perception feature map F6 are input into the second ASODEM for two-way feature fusion to expand the receptive field and correct the localization deviation, resulting in the fourth small target perception feature map F7.

[0190] S15. Input the third small target perception feature map into the third initial feature extraction unit and the fast spatial pyramid pooling unit in sequence for feature extraction and feature aggregation to obtain the fifth small target perception feature map.

[0191] For example, the third small target perception feature map F6 is input into the third CBS module for three downsampling operations to obtain the small target perception feature map F8. Then, the small target perception feature map F8 is input into the SPPF unit for feature aggregation. By aggregating global context information, the fifth target perception feature map F9 is obtained.

[0192] S16. Input the fourth and fifth small target perception feature maps into the third adaptive small target enhancement detection unit for feature enhancement to obtain the sixth small target perception feature map.

[0193] For example, the fourth small target perception feature map F7 and the fifth target perception feature map F9 are input into the third ASODEM for feature calibration and enhancement to obtain the sixth small target perception feature map F10.

[0194] For example, regarding step S1 above, let's take an image F1 input to the energy storage device with an image size of 640×640 pixels and 3 channels as an example for explanation:

[0195] 1) The image F1 of the energy storage device is input into the first CBS module for downsampling convolution with a stride of 2 to obtain a small target perception feature map F2 (320×320 pixels, 32 channels). Then, the small target perception feature map F2 is input into the first C2f module for depth feature extraction to obtain a first target perception feature map F3 (320×320 pixels, 64 channels). Then, the first small target perception feature map F3 is input into the second CBS module for downsampling to obtain a small target perception feature map F5 (160×160 pixels, 64 channels). Next, the small target perception feature map F5 is input into the second C2f module for deep feature extraction to obtain the third target perception feature map F6 (size 160×160 pixels, 128 channels). Then, the third small target perception feature map F6 is input into the third CBS module for three downsampling operations to obtain the small target perception feature map F8 (size 80×80 pixels, 128 channels). Then, the small target perception feature map F8 is input into the SPPF unit to aggregate global context information to obtain the fifth target perception feature map F9 (size 80×80 pixels, 256 channels).

[0196] 2) Input the first small target perception feature map F3 and the image F1 of the energy storage device into the first ASODEM to obtain the second small target perception feature map F4 (320×320 pixels, 64 channels). Then, input the second small target perception feature map F4 and the third small target perception feature map F6 into the second ASODEM to obtain the fourth small target perception feature map F7 (160×160 pixels, 128 channels). After that, input the fourth small target perception feature map F7 and the fifth target perception feature map F9 into the third ASODEM for feature calibration and enhancement to obtain the sixth small target perception feature map F10 (80×80 pixels, 256 channels).

[0197] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0198] In one possible implementation, Figure 6 This is a schematic diagram of the structure of the adaptive small target enhancement detection unit provided in an embodiment of this application. Figure 6 As shown, each adaptive small target enhancement detection unit includes: a cutting unit, a first convolutional unit, a second convolutional unit, a spatial structure perception feature calibration unit, a first dilated convolutional unit, and a second dilated convolutional unit;

[0199] For example, the adaptive small target enhancement detection unit adopts a multi-branch feature decoupling and cross-scale perception architecture. It utilizes the orthogonal Spatial Structure Perception-Feature Calibration Unit (SSP-FCU) to extract the array arrangement features of small targets such as battery tabs and terminals in a refined manner. While maintaining high-resolution details, it captures long-distance contextual information and effectively resists occlusion interference, which can significantly improve the model's ability to express and locate features of weak and small targets.

[0200] Accordingly, one possible implementation of step S12 above includes the following steps:

[0201] S121. The first small target perception feature map and the image of the energy storage device are superimposed and fused to obtain the seventh small target perception feature map.

[0202] For example, the first small target perception feature map F3 and the image F1 of the energy storage device are simultaneously input into the Add operation for feature superposition and fusion to obtain the seventh small target perception feature map A1.

[0203] S122. Input the seventh small target perception feature map into the cutting unit to obtain the eighth small target perception feature map, the ninth small target perception feature map, and the tenth small target perception feature map;

[0204] For example, the seventh small target perception feature map A1 is input to the cutting unit for split operation to obtain the eighth small target perception feature map A2, the ninth small target perception feature map A3, and the tenth small target perception feature map A4.

[0205] S123. Input the eighth small target perception feature map into the first convolutional unit to obtain the eleventh small target perception feature map;

[0206] For example, if the first convolutional unit is a Conv3*3 convolution, then the eighth small target perception feature map A2 is input into the Conv3*3 convolution to extract the basic local texture features of the energy storage device, and the eleventh small target perception feature map A5 is obtained.

[0207] S124. Input the ninth small target perception feature map into the second convolutional unit to obtain the twelfth small target perception feature map;

[0208] For example, if the second convolutional unit is a Conv1*1 convolution, then the ninth small target perception feature map A3 is input into the Conv1*1 convolution to perform cross-channel information interaction and feature integration, and the twelfth small target perception feature map A6 is obtained.

[0209] S125. Input the tenth small target perception feature map into the spatial structure perception feature calibration unit to obtain the thirteenth small target perception feature map;

[0210] For example, the tenth small target perception feature map A4 is input into the SSP-FCU unit, and its internal orthogonal pooling mechanism is used to capture spatial location information to obtain the thirteenth small target perception feature map A7.

[0211] S126. The eleventh small target perception feature map, the twelfth small target perception feature map, and the thirteenth small target perception feature map are spliced ​​together to obtain the fourteenth small target perception feature map.

[0212] For example, the Concat operation is used to concatenate the eleventh small target perception feature map A5, the twelfth small target perception feature map A6, and the thirteenth small target perception feature map A7 to obtain the fourteenth small target perception feature map A8.

[0213] S127. Input the fourteenth small target perception feature map into the first dilated convolution unit and the second dilated convolution unit respectively to obtain the second small target perception feature map.

[0214] For example, the first dilated convolutional unit is a DilatedConv3*3 dilated convolution with an inflation rate of 2, and the second dilated convolutional unit is a DilatedConv3*3 dilated convolution with an inflation rate of 4. Then, the fourteenth small target perception feature map A8 is input into the DilatedConv3*3 dilated convolution with an inflation rate of 2 to capture mid-range context information, resulting in the small target perception feature map A9. The fourteenth small target perception feature map A8 is input into the DilatedConv3*3 dilated convolution with an inflation rate of 4 to capture long-range context information, resulting in the small target perception feature map A10. Then, the small target perception feature maps A9 and A10 are simultaneously input into the Add operation for feature superposition and fusion to obtain the second small target perception feature map F4.

[0215] For example, regarding step S12 above, let's take the input image F1 (640×640 pixels, 3 channels) of the energy storage device and the first small target perception feature map F3 (320×320 pixels, 64 channels) as an example for explanation:

[0216] 1) Perform a convolution transformation with a stride of 2 on the image F1 of the energy storage device to match the spatial dimension. Input the first small target perception feature map F3 and the image F1 of the energy storage device into the Add operation to perform feature superposition and fusion, and obtain the seventh small target perception feature map A1 (size is 320×320 pixels, 64 channels).

[0217] 2) Input the seventh small target perception feature map A1 into the cutting unit for split operation to obtain three branches: the eighth small target perception feature map A2 (32 channels), the ninth small target perception feature map A3 (16 channels), and the tenth small target perception feature map A4 (16 channels).

[0218] 3) In branch one, the eighth small target perception feature map A2 is input into a Conv3*3 convolution to obtain the eleventh small target perception feature map A5 (320×320 pixels, 32 channels); in branch two, the ninth small target perception feature map A3 is input into a Conv1*1 convolution to obtain the twelfth small target perception feature map A6 (320×320 pixels, 16 channels); in branch three, the tenth small target perception feature map A4 is input into the SSP-FCU unit to obtain the thirteenth small target perception feature map A7 (320×320 pixels, 16 channels).

[0219] 4) The eleventh small target perception feature map A5, the twelfth small target perception feature map A6, and the thirteenth small target perception feature map A7 are concatenated using the Concat operation to obtain the fourteenth small target perception feature map A8 (size 320×320 pixels, 64 channels). Then, it is input in parallel into a DilatedConv3*3 dilated convolution with a dilation rate of 2 and a DilatedConv3*3 dilated convolution with a dilation rate of 4 to obtain small target perception feature maps A9 and A10. Then, the small target perception feature maps A9 and A10 are added together to obtain the second small target perception feature map F4 (size 320×320 pixels, 64 channels).

[0220] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0221] In one possible implementation, Figure 7 This is a schematic diagram of the spatial structure sensing feature calibration unit provided in an embodiment of this application. Figure 7 As shown, the spatial structure perception feature calibration unit includes: a horizontal average pooling unit, a vertical average pooling unit, a third convolutional unit, a fourth convolutional unit, a first activation unit, and a second activation unit.

[0222] For example, the spatial structure perception feature calibration unit is used to capture the unique array arrangement structure of the energy storage battery module by utilizing global pooling operations in orthogonal directions, and generate attention weights with spatial position sensitivity.

[0223] Accordingly, one possible implementation of step S125 above includes the following steps:

[0224] Step 1: Input the tenth small target perception feature map into the horizontal average pooling unit and the vertical average pooling unit respectively to obtain the sixteenth small target perception feature map;

[0225] For example, the tenth small target perception feature map A4 is input into the horizontal average pooling unit, the feature is compressed along the horizontal direction, the position information of the width dimension is extracted, and the horizontal feature vector S1 is obtained. The tenth small target perception feature map A4 is input into the vertical average pooling unit, the feature is compressed along the vertical direction, the position information of the height dimension is extracted, and the vertical feature vector S2 is obtained. Then, the horizontal feature vector S1 and the vertical feature vector S2 are concatenated by the Concat operation to obtain the sixteenth small target perception feature map S3.

[0226] Step 2: Input the sixteenth small target perception feature map into the third convolutional unit, the first activation unit, the fourth convolutional unit, and the second activation unit in sequence to obtain the attention weight map;

[0227] For example, the third and fourth convolutional units are both Conv1*1 convolutions, the first activation unit is a ReLU activation function, and the second activation unit is a Sigmoid activation function. Then, the sixteenth small target perception feature map S3 is input into the Conv1*1 convolution to perform feature dimensionality reduction and channel fusion to obtain the small target perception feature map S4. Then, the small target perception feature map S4 is input into the ReLU activation function to increase the non-linear feature expression capability to obtain the small target perception feature map S5.

[0228] Next, the small target perception feature map S5 is input into the Conv1*1 convolution to restore the feature dimension to the same number of channels as the original, resulting in the small target perception feature map S6. Then, the small target perception feature map S6 is input into the Sigmoid activation function to obtain the attention weight map S7.

[0229] Step 3: Multiply the tenth sub-target perception feature map and the attention weight map to obtain the thirteenth sub-target perception feature map.

[0230] For example, multiplying the tenth small target perception feature map A4 with the attention weight map S7 yields the thirteenth small target perception feature map A7.

[0231] For example, let's take the input of the tenth small target perception feature map A4 (with 16 channels) as an example to illustrate step S125 above:

[0232] 1) Input the tenth small target perception feature map A4 into the horizontal average pooling unit to obtain the horizontal feature vector S1 (size 320×1 pixel, 16 channels), and input the tenth small target perception feature map A4 into the vertical average pooling unit to obtain the vertical feature vector S2 (size 1×320 pixels, 16 channels). Use the Concat operation to concatenate the horizontal feature vector S1 and the vertical feature vector S2 to obtain the sixteenth small target perception feature map S3 (size 640×1 pixel, 16 channels).

[0233] 2) Input the sixteenth small target perception feature map S3 into the Conv1*1 convolution to obtain the small target perception feature map S4 (size 640×1 pixel, 8 channels). Then input the small target perception feature map S4 into the ReLU activation function to obtain the small target perception feature map S5 (size 640×1 pixel, 8 channels).

[0234] 3) Next, the small target perception feature map S5 is input into the Conv1*1 convolution to perform channel upscaling to restore the original number of channels, resulting in the small target perception feature map S6 (size 640×1 pixel, 16 channels). Then, the small target perception feature map S6 is input into the Sigmoid activation function to obtain the attention weight map S7 (size 320×320 pixels, 16 channels). Finally, the tenth small target perception feature map A4 and the attention weight map S7 are multiplied to obtain the thirteenth small target perception feature map A7.

[0235] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0236] S2, input multiple small target perception feature maps into the dynamic environment adaptation feature fusion network unit to obtain target environment adaptation fusion feature maps;

[0237] For example, multiple small target perception feature maps are input into a dynamic environment adaptive feature fusion network unit for multi-scale feature fusion, and the features are adaptively reconstructed and calibrated to effectively eliminate the interference of reflection, shadow and dust on feature extraction.

[0238] In one possible implementation, Figure 8 This is a schematic diagram of the structure of the dynamic environment adaptation feature fusion network unit provided in an embodiment of this application. Figure 8As shown, the Dynamic Environment-Adaptive Feature Fusion Neck (DEAFFN) unit includes: multiple depthwise separable convolutional units, multiple dynamic environment perception-adaptive feature correction (DEP-AFCM) units, a first average pooling unit, a fifth convolutional unit, a downsampling unit, a normalization and activation unit, a third dilated convolutional unit, and an upsampling unit;

[0239] The multiple depthwise separable convolutional units include: a first depthwise separable convolutional unit and a second depthwise separable convolutional unit;

[0240] The multiple dynamic environment perception feature adaptive correction units include: a first dynamic environment perception feature adaptive correction unit, a second dynamic environment perception feature adaptive correction unit, and a third dynamic environment perception feature adaptive correction unit.

[0241] Accordingly, one possible implementation of step S2 above includes the following steps:

[0242] S21. The second small target perception feature map is sequentially input into the first depthwise separable convolutional unit, the first dynamic environment perception feature adaptive correction unit, the first average pooling unit, the fifth convolutional unit, and the downsampling unit to obtain the first environment adaptive fusion feature map.

[0243] For example, if the first depthwise separable convolutional unit is a DWConv3*3 depthwise separable convolution and the fifth convolutional unit is a Conv1*1 convolution, then the second small target perception feature map F4 is input into the DWConv3*3 depthwise separable convolution to extract shallow spatial details and obtain the environmental adaptation fusion feature map F11. Then, the environmental adaptation fusion feature map F11 is input into the first DEP-AFCM unit to obtain the environmental adaptation fusion feature map F12. Then, the environmental adaptation fusion feature map F12 is input into the first average pooling unit for average pooling to obtain the environmental adaptation fusion feature map F13. Then, it is input into the Conv1*1 convolution for channel adjustment to obtain the environmental adaptation fusion feature map F14. Finally, the environmental adaptation fusion feature map F14 is input into the downsampling unit to obtain the first environmental adaptation fusion feature map F15.

[0244] S22. The fourth small target perception feature map is sequentially input into the second depthwise separable convolutional unit, the second dynamic environment perception feature adaptive correction unit, and the normalization and activation unit to obtain the second environment adaptive fusion feature map.

[0245] For example, the normalization and activation unit can be a BN+SiLU module, and the second depthwise separable convolutional unit is a DWConv1*1 depthwise separable convolution. Then, the fourth small target perception feature map F7 is input into the DWConv1*1 depthwise separable convolution to obtain the environment adaptation fusion feature map F16. Then, the environment adaptation fusion feature map F16 is input into the second DEP-AFCM unit for environment adaptation enhancement to obtain the environment adaptation fusion feature map F17. After that, the environment adaptation fusion feature map F17 is input into the BN+SiLU module for feature distribution adjustment to obtain the second environment adaptation fusion feature map F18.

[0246] S23. The sixth small target perception feature map is sequentially input into the third dynamic environment perception feature adaptive correction unit, the third dilated convolution unit, and the upsampling unit to obtain the third environment adaptive fusion feature map.

[0247] For example, if the third dilated convolutional unit is a DilatedConv3*3 dilated convolution with an inflation rate of 5, then the sixth small target perception feature map F10 is input into the third DEP-AFCM unit for deep semantic feature calibration to obtain the environment adaptation fusion feature map F19. Then, the environment adaptation fusion feature map F19 is input into the DilatedConv3*3 dilated convolution with an inflation rate of 5 to capture long-range contextual information to obtain the environment adaptation fusion feature map F20. After that, the environment adaptation fusion feature map F20 is input into the upsampling unit to enlarge the feature map size to obtain the third environment adaptation fusion feature map F21.

[0248] S24. The first environmental adaptation fusion feature map, the second environmental adaptation fusion feature map, and the third environmental adaptation fusion feature map are spliced ​​together to obtain the target environmental adaptation fusion feature map.

[0249] For example, the first environment adaptation fusion feature map F15, the second environment adaptation fusion feature map F18, and the third environment adaptation fusion feature map F21 are spliced ​​together using the Concat operation to obtain the target environment adaptation fusion feature map F22.

[0250] For example, regarding step S2 above, let's take the second small target perception feature map F4 (320×320 pixels, 64 channels), the fourth small target perception feature map F7 (160×160 pixels, 128 channels), and the sixth small target perception feature map F10 (80×80 pixels, 256 channels) as examples for illustration:

[0251] 1) The second small target perception feature map F4 (320×320 pixels, 64 channels) is input into the DWConv3*3 depthwise separable convolution to obtain the environment adaptation fusion feature map F11 (320×320 pixels, 64 channels). Then, the environment adaptation fusion feature map F11 is input into the first DEP-AFCM unit to obtain the environment adaptation fusion feature map F12 (320×320 pixels, 64 channels). The environment adaptation fusion feature map F12 is then input into the first average pooling unit for average pooling to obtain the environment adaptation fusion feature map F13 (320×320 pixels, 64 channels). Then, it is input into the Conv1*1 convolution to obtain the environment adaptation fusion feature map F14 (320×320 pixels, 128 channels). Finally, the environment adaptation fusion feature map F14 is input into the downsampling unit to obtain the first environment adaptation fusion feature map F15 (160×160 pixels, 128 channels).

[0252] 2) Input the fourth small target perception feature map F7 into the DWConv1*1 depthwise separable convolution to obtain the environment adaptation fusion feature map F16 (size 160×160 pixels, 128 channels). Then, input the environment adaptation fusion feature map F16 into the second DEP-AFCM unit for environment adaptation enhancement to obtain the environment adaptation fusion feature map F17 (size 160×160 pixels, 128 channels). After that, input the environment adaptation fusion feature map F17 into the BN+SiLU module for feature distribution adjustment to obtain the second environment adaptation fusion feature map F18 (size 160×160 pixels, 128 channels).

[0253] 3) Input the sixth small target perception feature map F10 into the third DEP-AFCM unit to obtain the environment adaptation fusion feature map F19 (size 80×80 pixels, 256 channels). Then, input the environment adaptation fusion feature map F19 into a dilatedConv3*3 dilated convolution with a dilation rate of 5 to obtain the environment adaptation fusion feature map F20 (size 80×80 pixels, 256 channels). Then, input the environment adaptation fusion feature map F20 into the upsampling unit to obtain the third environment adaptation fusion feature map F21 (size 160×160 pixels, 256 channels).

[0254] 4) The first environmental adaptation fusion feature map F15, the second environmental adaptation fusion feature map F18, and the third environmental adaptation fusion feature map F21 are stitched together using the Concat operation to obtain the target environmental adaptation fusion feature map F22 (size 160×160 pixels, 512 channels).

[0255] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0256] In one possible implementation, Figure 9 This is a schematic diagram of the structure of the dynamic environment perception feature adaptive correction unit provided in an embodiment of this application. Figure 9 As shown, each dynamic environment-aware feature adaptive correction unit includes: a sixth convolutional unit, a third depthwise separable convolutional unit, a batch normalization unit, a second average pooling unit, a third activation unit, a dynamic convolutional unit, and an efficient multi-scale attention unit (EMA).

[0257] For example, the dynamic environment perception feature adaptive correction unit adopts an environmental parameter decoupling and dynamic weight generation architecture. It uses parallel environment perception branches to extract global illumination intensity and background noise priors in real time. Combined with dynamic convolution and efficient multi-scale attention mechanism, it adaptively fuses feature responses for complex and variable illumination conditions at the energy storage power station site, thereby achieving active suppression of environmental interference and ensuring the stability and accuracy of the feature fusion process.

[0258] Accordingly, one possible implementation of step S21 above includes the following steps:

[0259] S211. Input the second small target perception feature map into the first depthwise separable convolutional unit to obtain the fourth environment adaptation fusion feature map;

[0260] For example, if the first depthwise separable convolutional unit is a DWConv3*3 depthwise separable convolution, then the second small target perception feature map F4 is input into the DWConv3*3 depthwise separable convolution to extract shallow spatial details and obtain the fourth environmental adaptation fusion feature map E1 (i.e., environmental adaptation fusion feature map F11).

[0261] S212. Input the fourth environmental adaptation fusion feature map into the sixth convolutional unit and the batch normalization unit in sequence to obtain the fifth environmental adaptation fusion feature map.

[0262] For example, if the sixth convolutional unit is a Conv3*3 convolution and the batch normalization unit is a BN module, then the fourth environmental adaptation fusion feature map E1 is input into the Conv3*3 convolution to extract local features, resulting in the environmental adaptation fusion feature map E2. Then, the environmental adaptation fusion feature map E2 is input into the BN module for distribution adjustment, resulting in the fifth environmental adaptation fusion feature map E4.

[0263] S213. Input the fourth environmental adaptation fusion feature map into the third depthwise separable convolutional unit and the second average pooling unit in sequence to obtain the sixth environmental adaptation fusion feature map.

[0264] For example, if the third depthwise separable convolutional unit is a DWConv1*1 depthwise separable convolution, then the fourth environment adaptation fusion feature map E1 is input into the DWConv1*1 depthwise separable convolution for lightweight feature transformation to obtain the environment adaptation fusion feature map E3. Then, the environment adaptation fusion feature map E3 is input into the second average pooling unit to extract global environment information to obtain the sixth environment adaptation fusion feature map E5.

[0265] S214. Perform feature overlay processing on the fifth and sixth environmental adaptation fusion feature maps and input them into the third activation unit to obtain the environmental adaptation fusion weight map.

[0266] For example, if the third activation unit is the Softmax activation function, then the fifth environmental adaptation fusion feature map E4 and the sixth environmental adaptation fusion feature map E5 are superimposed by the Add operation to obtain the environmental adaptation fusion feature map E6, which is then input into the Softmax activation function to obtain the environmental adaptation fusion weight map E8.

[0267] S215. Input the sixth environmental adaptation fusion feature map into the dynamic convolution unit and the efficient multi-scale attention unit in sequence to obtain the seventh environmental adaptation fusion feature map.

[0268] For example, the sixth environmental adaptation fusion feature map E5 is input into the dynamic convolution unit to obtain the environmental adaptation fusion feature map E7. Then, the environmental adaptation fusion feature map E7 is input into the efficient multi-scale attention unit for enhancement processing to obtain the seventh environmental adaptation fusion feature map E9.

[0269] S216. Multiply the seventh environmental adaptation fusion feature map and the environmental adaptation fusion weight map to obtain the eighth environmental adaptation fusion feature map;

[0270] For example, multiplying the seventh environmental adaptation fusion feature map E9 and the environmental adaptation fusion weight map E8 yields the eighth environmental adaptation fusion feature map E10 (i.e., environmental adaptation fusion feature map F12).

[0271] S217. The eighth environmental adaptation fusion feature map is sequentially input into the first average pooling unit, the fifth convolution unit, and the downsampling unit to obtain the first environmental adaptation fusion feature map.

[0272] For example, the eighth environmental adaptation fusion feature map E10 (i.e., environmental adaptation fusion feature map F12) is sequentially input into the first average pooling unit, the fifth convolution unit, and the downsampling unit to obtain the first environmental adaptation fusion feature map F15.

[0273] For example, in step S21 above, the fourth environment adaptation fusion feature map E1 is illustrated using the environment adaptation fusion feature map F11 (320×320 pixels, 64 channels):

[0274] The fourth environmental adaptation fusion feature map E1 is input into the Conv3*3 convolution to obtain the environmental adaptation fusion feature map E2 (size is 320×320 pixels, 64 channels). Then, the environmental adaptation fusion feature map E2 is input into the BN module to obtain the fifth environmental adaptation fusion feature map E4 (size is 320×320 pixels, 64 channels).

[0275] The fourth environmental adaptation fusion feature map E1 is input into the DWConv1*1 depthwise separable convolution to obtain the environmental adaptation fusion feature map E3 (size 320×320 pixels, 64 channels). Then, the environmental adaptation fusion feature map E3 is input into the second average pooling unit to obtain the sixth environmental adaptation fusion feature map E5 (size 1×1 pixel, 64 channels).

[0276] The fifth environmental adaptation fusion feature map E4 and the sixth environmental adaptation fusion feature map E5 are added together by an Add operation to obtain the environmental adaptation fusion feature map E6 (320×320 pixels, 64 channels). Then, it is input into the Softmax activation function to obtain the environmental adaptation fusion weight map E8 (320×320 pixels, 64 channels).

[0277] The sixth environmental adaptation fusion feature map E5 is input into the dynamic convolution unit to obtain the environmental adaptation fusion feature map E7 (size 1×1 pixel, 64 channels). Then, the environmental adaptation fusion feature map E7 is input into the efficient multi-scale attention unit for enhancement processing to obtain the seventh environmental adaptation fusion feature map E9 (size 1×1 pixel, 64 channels).

[0278] Multiplying the seventh environmental adaptation fusion feature map E9 and the environmental adaptation fusion weight map E8 yields the eighth environmental adaptation fusion feature map E10 (320×320 pixels, 64 channels), which is the environmental adaptation fusion feature map F12.

[0279] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0280] S3, input the target environment adaptation fusion feature map into the asymmetric spatial refinement unit to obtain the target spatial refinement feature map;

[0281] For example, the target environment adaptation fusion feature map is input into the asymmetric spatial refinement unit. Using orthogonal asymmetric convolution decomposition and cascaded pooling mechanism, low-rank reconstruction is performed on the regular geometric boundary features of the energy storage device. This significantly reduces computational redundancy while generating high-precision spatial attention weights, achieving pixel-level refinement correction of the target boundary.

[0282] In one possible implementation, Figure 10 This is a schematic diagram of the structure of the asymmetric spatial refinement unit provided in an embodiment of this application. Figure 10 As shown, the Asymmetric Spatial Refinement Module (ASR) unit includes: a first asymmetric convolutional layer, a second asymmetric convolutional layer, a third average pooling unit, a fourth average pooling unit, a max pooling unit, a seventh convolutional unit, and an eighth convolutional unit.

[0283] Accordingly, one possible implementation of step S3 above includes the following steps:

[0284] S31. Input the target environment adaptation fusion feature map into the first asymmetric convolutional layer and the second asymmetric convolutional layer in sequence to extract features in the horizontal and vertical directions, and obtain the first spatial refinement feature map.

[0285] For example, if the first asymmetric convolutional layer is a Conv1*3 asymmetric convolution and the second asymmetric convolutional layer is a Conv3*1 asymmetric convolution, then the target environment adaptation fusion feature map F22 is input into the Conv1*3 asymmetric convolution to extract the linear spatial features in the horizontal direction, resulting in the spatial refinement feature map F23. Then, the spatial refinement feature map F23 is input into the Conv3*1 asymmetric convolution to extract the linear spatial features in the vertical direction, resulting in the first spatial refinement feature map F24.

[0286] S32. Input the first spatial refinement feature map into the third average pooling unit, the max pooling unit, the fourth average pooling unit, and the seventh convolution unit in sequence to obtain the second spatial refinement feature map.

[0287] For example, if the seventh convolutional unit is a Conv1*1 convolution, the first spatial refinement feature map F24 is input into the third average pooling unit for feature smoothing to obtain the spatial refinement feature map F25. Then, the spatial refinement feature map F25 is input into the max pooling unit for feature filtering to obtain the spatial refinement feature map F26. After that, the spatial refinement feature map F26 is input into the fourth average pooling unit for secondary smoothing to obtain the spatial refinement feature map F27. Then, the spatial refinement feature maps F25, F26, and F27 are concatenated through the Concat operation to obtain the spatial refinement feature map F28. Finally, the spatial refinement feature map F28 is input into the Conv1*1 convolution to obtain the second spatial refinement feature map F29.

[0288] S33. Perform feature overlay processing on the first space refinement feature map and the second space refinement feature map to obtain the third space refinement feature map;

[0289] For example, the first spatial refinement feature map F24 and the second spatial refinement feature map F29 are input into the Add operation to perform feature superposition, resulting in a spatial refinement feature map F30. Then, the spatial refinement feature map F30 is input into the Sigmoid activation function to obtain a third spatial refinement feature map F31.

[0290] S34. Input the third space refinement feature map into the eighth convolutional unit to obtain the target space refinement feature map.

[0291] For example, if the eighth convolutional unit is a Conv3*3 convolution, then the third space refinement feature map F31 is input into the Conv3*3 convolution to obtain the target space refinement feature map F32.

[0292] For example, the above step S3 is illustrated using the target environment adaptation fusion feature map F22 (160×160 pixels, 512 channels) as an example:

[0293] The target environment adaptation fusion feature map F22 is input into a Conv1*3 asymmetric convolution to obtain a spatial thinning feature map F23 (size 160×160 pixels, 512 channels). Then, the spatial thinning feature map F23 is input into a Conv3*1 asymmetric convolution to obtain the first spatial thinning feature map F24 (size 160×160 pixels, 512 channels).

[0294] The first spatial refinement feature map F24 is input into the third average pooling unit to obtain the spatial refinement feature map F25 (size 160×160 pixels, 512 channels). Then, the spatial refinement feature map F25 is input into the max pooling unit to obtain the spatial refinement feature map F26 (size 160×160 pixels, 512 channels). Finally, the spatial refinement feature map F26 is input into the fourth average pooling unit for secondary smoothing to obtain the spatial refinement feature map F27 (size...). The spatial thinning feature maps F25, F26, and F27 are then concatenated using a Concat operation to obtain spatial thinning feature map F28 (160×160 pixels, 1536 channels). F28 is then input into a Conv1*1 convolution to obtain the second spatial thinning feature map F29 (160×160 pixels, 512 channels).

[0295] The first spatial refinement feature map F24 and the second spatial refinement feature map F29 are input into the Add operation for feature superposition to obtain the spatial refinement feature map F30 (size 160×160 pixels, 512 channels). Then, the spatial refinement feature map F30 is input into the Sigmoid activation function to obtain the third spatial refinement feature map F31 (size 160×160 pixels, 512 channels). Finally, the third spatial refinement feature map F31 is input into the Conv3*3 convolution to obtain the target spatial refinement feature map F32 (size 160×160 pixels, 512 channels).

[0296] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0297] S4, input the refined feature map of the target space into the detection head unit to obtain the estimated defect information;

[0298] For example, the detection head unit can be a standard object detection head, which is a general deep learning object detection network terminal component. Its core principle is to use convolutional layers to map the upstream feature map to a specific output channel dimension, decoupling the high-dimensional features into class probabilities, bounding box regression parameters, and confidence scores.

[0299] The detection head unit is used to perform the final prediction and inference task, and outputs the detection image, including: specific category labels and geometric coordinates of defects in energy storage devices, realizing the final transformation from features to visualized detection results.

[0300] For example, the target space refined feature map F32 is input into the detection head unit to obtain the estimated defect information, which includes: defect bounding box coordinates, confidence level, and defect category. The defect level can be determined by comparing the confidence level with a preset threshold.

[0301] S5. Update the parameters of the initial defect detection model based on the estimated defect information and the labeled defect information corresponding to each image. Repeat steps S1-S5 until the preset training stopping condition is met to obtain the energy storage device defect detection model.

[0302] For example, the initial defect detection model is trained and optimized by supervised learning based on the estimated defect information and the labeled defect information corresponding to each image. The parameters are continuously updated by the backpropagation algorithm until the performance index of the initial defect detection model on the validation set converges, thus obtaining the energy storage device defect detection model.

[0303] In one possible implementation, the initial defect detection model can be trained and optimized by supervised learning based on the training set and validation set of the energy storage device image dataset randomly divided according to a preset ratio in the aforementioned step S201. The training environment can be based on a deep learning framework, with batch size and training period set, and Stochastic Gradient Descent (SGD) is used as the optimizer, with its initial learning rate and momentum factor set, and a cosine annealing strategy is used to periodically and dynamically adjust the learning rate to avoid getting trapped in local optima.

[0304] In addition, a composite loss function can be constructed, which mainly consists of three weighted parts: CIoU regression loss, which measures the overlap between the predicted bounding box in the estimated defect information and the true bounding box in the labeled defect information; Variable Focal Loss (VFL), which measures the accuracy of defect category classification; and Distribution Focal Loss (DFL), which optimizes the distribution of target bounding boxes. By calculating the composite loss function and backpropagating, end-to-end parameter optimization of the initial defect detection model can be achieved.

[0305] During the training process, the mean mAP (average accuracy) of the model is evaluated on the validation set after each training cycle until the performance of the initial defect detection model on the validation set no longer improves significantly. This results in a trained defect detection model for energy storage devices that can accurately identify energy storage devices and their subtle defects.

[0306] In one possible implementation, Figure 11 This is a schematic diagram of the structure of the energy storage device defect detection model provided in an embodiment of this application. Figure 11As shown, the Energy Storage Device Inspection Detection Network (ESDIN) is the initial defect detection model after training stops. It specifically includes: a small target perception multi-scale feature extraction network unit, a dynamic environment adaptation feature fusion network unit, an asymmetric spatial refinement unit, and a detection head unit.

[0307] The small target perception multi-scale feature extraction network unit includes: multiple adaptive small target enhancement detection units, fast spatial pyramid pooling units, multiple initial feature extraction units, and multiple deep feature extraction units. Each adaptive small target enhancement detection unit includes: a cutting unit, a first convolutional unit, a second convolutional unit, a spatial structure perception feature calibration unit, a first dilated convolutional unit, and a second dilated convolutional unit. The spatial structure perception feature calibration unit includes: a horizontal average pooling unit, a vertical average pooling unit, a third convolutional unit, a fourth convolutional unit, a first activation unit, and a second activation unit.

[0308] The dynamic environment adaptive feature fusion network unit includes: multiple depthwise separable convolutional units, multiple dynamic environment-aware feature adaptive correction units, a first average pooling unit, a fifth convolutional unit, a downsampling unit, a normalization and activation unit, a third dilated convolutional unit, and an upsampling unit. Each dynamic environment-aware feature adaptive correction unit includes: a sixth convolutional unit, a third depthwise separable convolutional unit, a batch normalization unit, a second average pooling unit, a third activation unit, a dynamic convolutional unit, and an efficient multi-scale attention unit.

[0309] The asymmetric spatial refinement unit includes: a first asymmetric convolutional layer, a second asymmetric convolutional layer, a third average pooling unit, a fourth average pooling unit, a max pooling unit, a seventh convolutional unit, and an eighth convolutional unit.

[0310] It is understandable that the energy storage device defect detection model and the initial defect detection model have the same model structure, the difference being that the model parameters are different. The parameters of the initial defect detection model are updated until the preset training stopping condition is met, thus obtaining the energy storage device defect detection model.

[0311] This application provides a defect detection method for energy storage devices. First, it acquires images of multiple energy storage devices and corresponding labeled defect information for each image. Then, it trains an initial defect detection model based on the images and labeled defect information to obtain a defect detection model for the energy storage devices. In this embodiment, training the initial defect detection model based on images of multiple energy storage devices and corresponding labeled defect information allows it to learn the unique appearance features and minor defect patterns of the energy storage devices, enabling the defect detection model obtained from the initial defect detection model to accurately output defect information.

[0312] Based on the above embodiments, the following are embodiments of the apparatus involved in this application:

[0313] Figure 12 A schematic diagram of the defect detection device for energy storage equipment provided in the embodiments of this application is shown below. Figure 12 As shown, the defect detection device 1200 for the energy storage device includes:

[0314] The acquisition module 1201 is used to acquire the target image of the energy storage device to be tested;

[0315] The processing module 1202 is used to input the target image into the energy storage device defect detection model to obtain the defect information of the energy storage device to be detected. The energy storage device defect detection model is determined based on images of multiple energy storage devices and the labeled defect information corresponding to each image. The defect information includes: defect type and defect level.

[0316] In one or more embodiments, before inputting the target image into the energy storage device defect detection model to obtain the defect information of the energy storage device to be detected, the processing module 1202 is further configured to:

[0317] Acquire images of multiple energy storage devices and corresponding defect annotations for each image;

[0318] The initial defect detection model is trained based on images of multiple energy storage devices and the corresponding labeled defect information for each image, thus obtaining the energy storage device defect detection model.

[0319] In one or more embodiments, the initial defect detection model includes: a small target perception multi-scale feature extraction network unit, a dynamic environment adaptation feature fusion network unit, an asymmetric spatial refinement unit, and a detection head unit;

[0320] Accordingly, the processing module 1202 trains the initial defect detection model based on images of multiple energy storage devices and the labeled defect information corresponding to each image, to obtain an energy storage device defect detection model, specifically used for:

[0321] S1, For each energy storage device image, input the energy storage device image into the small target perception multi-scale feature extraction network unit to obtain multiple small target perception feature maps;

[0322] S2, input multiple small target perception feature maps into the dynamic environment adaptation feature fusion network unit to obtain target environment adaptation fusion feature maps;

[0323] S3, input the target environment adaptation fusion feature map into the asymmetric spatial refinement unit to obtain the target spatial refinement feature map;

[0324] S4, input the refined feature map of the target space into the detection head unit to obtain the estimated defect information;

[0325] S5. Update the parameters of the initial defect detection model based on the estimated defect information and the labeled defect information corresponding to each image. Repeat steps S1-S5 until the preset training stopping condition is met to obtain the energy storage device defect detection model.

[0326] In one or more embodiments, the small target perception multi-scale feature extraction network unit includes: multiple adaptive small target enhancement detection units, fast spatial pyramid pooling units, multiple initial feature extraction units, and multiple deep feature extraction units;

[0327] Correspondingly, the processing module 1202, for each energy storage device's image, inputs the image of the energy storage device into the small target perception multi-scale feature extraction network unit to obtain multiple small target perception feature maps, specifically used for:

[0328] 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 first deep feature extraction unit for feature extraction to obtain the first small target perception feature map;

[0329] The first small target perception feature map and the image of the energy storage device are input into the first adaptive small target enhancement and detection unit for feature enhancement to obtain the second small target perception feature map.

[0330] The first small target perception feature map is sequentially input into the second initial feature extraction unit and the second deep feature extraction unit for feature extraction to obtain the third small target perception feature map.

[0331] The second and third small target perception feature maps are input into the second adaptive small target enhancement and detection unit for feature enhancement to obtain the fourth small target perception feature map.

[0332] The third small target perception feature map is sequentially input into the third initial feature extraction unit and the fast spatial pyramid pooling unit for feature extraction and feature aggregation to obtain the fifth small target perception feature map.

[0333] The fourth and fifth small target perception feature maps are input into the third adaptive small target enhancement detection unit for feature enhancement to obtain the sixth small target perception feature map.

[0334] In one or more embodiments, each adaptive small target enhancement detection unit includes: a cutting unit, a first convolutional unit, a second convolutional unit, a spatial structure perception feature calibration unit, a first dilated convolutional unit, and a second dilated convolutional unit;

[0335] Correspondingly, the processing module 1202 inputs the first small target perception feature map and the image of the energy storage device into the first adaptive small target enhancement and detection unit for feature enhancement to obtain the second small target perception feature map, specifically used for:

[0336] The first small target perception feature map and the image of the energy storage device are superimposed and fused to obtain the seventh small target perception feature map.

[0337] The seventh small target perception feature map is input into the segmentation unit to obtain the eighth, ninth, and tenth small target perception feature maps;

[0338] The feature map of the eighth small target is input into the first convolutional unit to obtain the feature map of the eleventh small target.

[0339] The ninth small target perception feature map is input into the second convolutional unit to obtain the twelfth small target perception feature map;

[0340] The tenth small target perception feature map is input into the spatial structure perception feature calibration unit to obtain the thirteenth small target perception feature map;

[0341] The eleventh, twelfth, and thirteenth small target perception feature maps are stitched together to obtain the fourteenth small target perception feature map.

[0342] The fourteenth small target perception feature map is input into the first dilated convolution unit and the second dilated convolution unit respectively to obtain the second small target perception feature map.

[0343] In one or more embodiments, the spatial structure-aware feature calibration unit includes: a horizontal average pooling unit, a vertical average pooling unit, a third convolution unit, a fourth convolution unit, a first activation unit, and a second activation unit.

[0344] Correspondingly, processing module 1202 inputs the tenth small target perception feature map to the spatial structure perception feature calibration unit to obtain the thirteenth small target perception feature map, which is specifically used for:

[0345] The tenth small target perception feature map is input into the horizontal average pooling unit and the vertical average pooling unit respectively to obtain the sixteenth small target perception feature map;

[0346] The sixteenth small target perception feature map is sequentially input into the third convolutional unit, the first activation unit, the fourth convolutional unit, and the second activation unit to obtain the attention weight map;

[0347] Multiplying the tenth sub-target perception feature map and the attention weight map yields the thirteenth sub-target perception feature map.

[0348] In one or more embodiments, the dynamic environment adaptation feature fusion network unit includes: multiple depthwise separable convolutional units, multiple dynamic environment-aware feature adaptive correction units, a first average pooling unit, a fifth convolutional unit, a downsampling unit, a normalization and activation unit, a third dilated convolutional unit, and an upsampling unit.

[0349] Correspondingly, the processing module 1202 inputs multiple small target perception feature maps into the dynamic environment adaptation feature fusion network unit to obtain a target environment adaptation fusion feature map, specifically used for:

[0350] The second small target perception feature map is sequentially input into the first depthwise separable convolutional unit, the first dynamic environment perception feature adaptive correction unit, the first average pooling unit, the fifth convolutional unit, and the downsampling unit to obtain the first environment adaptive fusion feature map;

[0351] The fourth small target perception feature map is sequentially input into the second depthwise separable convolutional unit, the second dynamic environment perception feature adaptive correction unit, and the normalization and activation unit to obtain the second environment adaptive fusion feature map.

[0352] The sixth small target perception feature map is sequentially input into the third dynamic environment perception feature adaptive correction unit, the third dilated convolution unit, and the upsampling unit to obtain the third environment adaptive fusion feature map;

[0353] The first environmental adaptation fusion feature map, the second environmental adaptation fusion feature map, and the third environmental adaptation fusion feature map are stitched together to obtain the target environmental adaptation fusion feature map.

[0354] In one or more embodiments, each dynamic environment-aware feature adaptive correction unit includes: a sixth convolutional unit, a third depthwise separable convolutional unit, a batch normalization unit, a second average pooling unit, a third activation unit, a dynamic convolutional unit, and an efficient multi-scale attention unit.

[0355] Correspondingly, the processing module 1202 sequentially inputs the second small target perception feature map into the first depthwise separable convolutional unit, the first dynamic environment perception feature adaptive correction unit, the first average pooling unit, the fifth convolutional unit, and the downsampling unit to obtain the first environment-adaptive fusion feature map, specifically used for:

[0356] The second small target perception feature map is input into the first depthwise separable convolutional unit to obtain the fourth environment-adaptive fusion feature map.

[0357] The fourth environmental adaptation fusion feature map is sequentially input into the sixth convolutional unit and the batch normalization unit to obtain the fifth environmental adaptation fusion feature map.

[0358] The fourth environmental adaptation fusion feature map is sequentially input into the third depthwise separable convolutional unit and the second average pooling unit to obtain the sixth environmental adaptation fusion feature map.

[0359] The fifth and sixth environmental adaptation fusion feature maps are overlaid and then input into the third activation unit to obtain the environmental adaptation fusion weight map.

[0360] The sixth environment adaptation fusion feature map is sequentially input into the dynamic convolution unit and the efficient multi-scale attention unit to obtain the seventh environment adaptation fusion feature map.

[0361] The seventh environmental adaptation fusion feature map and the environmental adaptation fusion weight map are multiplied to obtain the eighth environmental adaptation fusion feature map;

[0362] The eighth environmental adaptation fusion feature map is sequentially input into the first average pooling unit, the fifth convolution unit, and the downsampling unit to obtain the first environmental adaptation fusion feature map.

[0363] In one or more embodiments, the asymmetric spatial refinement unit includes: a first asymmetric convolutional layer, a second asymmetric convolutional layer, a third average pooling unit, a fourth average pooling unit, a max pooling unit, a seventh convolutional unit, and an eighth convolutional unit.

[0364] Correspondingly, the processing module 1202 inputs the target environment adaptation fusion feature map into the asymmetric spatial refinement unit to obtain the target spatial refinement feature map, specifically used for:

[0365] The target environment adaptation fusion feature map is sequentially input into the first asymmetric convolutional layer and the second asymmetric convolutional layer to extract features in the horizontal and vertical directions, thereby obtaining the first spatial refinement feature map;

[0366] The first spatially refined feature map is sequentially input into the third average pooling unit, the max pooling unit, the fourth average pooling unit, and the seventh convolutional unit to obtain the second spatially refined feature map.

[0367] The first space refined feature map and the second space refined feature map are superimposed to obtain the third space refined feature map;

[0368] The refined feature map of the third space is input into the eighth convolutional unit to obtain the refined feature map of the target space.

[0369] The defect 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.

[0370] Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 13 As shown, the electronic device 1300 includes: a processor 1301, a memory 1302, and a bus 1303;

[0371] The memory 1302 is used to store computer-executed instructions from the processor 1301;

[0372] The processor 1301 is configured to execute the technical solutions of any of the foregoing method embodiments by executing computer execution instructions.

[0373] The specific implementation process of processor 1301 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0374] 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.

[0375] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0376] 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.

[0377] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0378] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0379] 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.

[0380] 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.

[0381] 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.

[0382] 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.

[0383] 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.

[0384] 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.

[0385] 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.

[0386] 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 defect detection method for energy storage devices, characterized in that, include: Acquire a target image of the energy storage device to be tested; The target image is input into the energy storage device defect detection model to obtain the defect information of the energy storage device to be detected. The energy storage device defect detection model is determined based on images of multiple energy storage devices and the labeled defect information corresponding to each image. The defect information includes: defect type and defect level.

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

3. The method according to claim 2, characterized in that, The initial defect detection model includes: a small target perception multi-scale feature extraction network unit, a dynamic environment adaptation feature fusion network unit, an asymmetric spatial refinement unit, and a detection head unit; Accordingly, the step of training the initial defect detection model based on the images of the multiple energy storage devices and the labeled defect information corresponding to each image to obtain the energy storage device defect detection model includes: S1, For each energy storage device image, the image of the energy storage device is input into the small target perception multi-scale feature extraction network unit to obtain multiple small target perception feature maps; S2, input the multiple small target perception feature maps into the dynamic environment adaptation feature fusion network unit to obtain the target environment adaptation fusion feature map; S3, input the target environment adaptation fusion feature map into the asymmetric spatial refinement unit to obtain the target spatial refinement feature map; S4, input the refined feature map of the target space into the detection head unit to obtain the estimated defect information; S5, update the parameters of the initial defect detection model according to the estimated defect information and the labeled defect information corresponding to each image, repeat steps S1-S5 until the preset training stopping condition is met, and obtain the energy storage device defect detection model.

4. The method according to claim 3, characterized in that, The small target perception multi-scale feature extraction network unit includes: multiple adaptive small target enhancement detection units, fast spatial pyramid pooling units, multiple initial feature extraction units, and multiple deep feature extraction units; Accordingly, for each energy storage device, the image of the energy storage device is input into the small target perception multi-scale feature extraction network unit to obtain multiple small target perception 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 first depth feature extraction unit for feature extraction to obtain the first small target perception feature map; The first small target perception feature map and the image of the energy storage device are input into the first adaptive small target enhancement and detection unit for feature enhancement to obtain the second small target perception feature map. The first small target perception feature map is sequentially input into the second initial feature extraction unit and the second deep feature extraction unit for feature extraction to obtain the third small target perception feature map. The second small target perception feature map and the third small target perception feature map are input into the second adaptive small target enhancement and detection unit for feature enhancement to obtain the fourth small target perception feature map. The third small target perception feature map is sequentially input into the third initial feature extraction unit and the fast spatial pyramid pooling unit for feature extraction and feature aggregation to obtain the fifth small target perception feature map. The fourth and fifth small target perception feature maps are input into the third adaptive small target enhancement detection unit for feature enhancement to obtain the sixth small target perception feature map.

5. The method according to claim 4, characterized in that, Each adaptive small target enhancement detection unit includes: a cutting unit, a first convolutional unit, a second convolutional unit, a spatial structure perception feature calibration unit, a first dilated convolutional unit, and a second dilated convolutional unit; Accordingly, the step of inputting the first small target perception feature map and the image of the energy storage device into the first adaptive small target enhancement and detection unit for feature enhancement to obtain the second small target perception feature map includes: The first small target perception feature map and the image of the energy storage device are superimposed and fused to obtain the seventh small target perception feature map. The seventh small target perception feature map is input into the cutting unit to obtain the eighth small target perception feature map, the ninth small target perception feature map, and the tenth small target perception feature map; The eighth small target perception feature map is input into the first convolutional unit to obtain the eleventh small target perception feature map; The ninth small target perception feature map is input into the second convolutional unit to obtain the twelfth small target perception feature map; The tenth small target perception feature map is input into the spatial structure perception feature calibration unit to obtain the thirteenth small target perception feature map; The eleventh small target perception feature map, the twelfth small target perception feature map, and the thirteenth small target perception feature map are stitched together to obtain the fourteenth small target perception feature map. The fourteenth small target perception feature map is input into the first dilated convolution unit and the second dilated convolution unit respectively to obtain the second small target perception feature map.

6. The method according to claim 5, characterized in that, The spatial structure perception feature calibration unit includes: a horizontal average pooling unit, a vertical average pooling unit, a third convolution unit, a fourth convolution unit, a first activation unit, and a second activation unit. Accordingly, the step of inputting the tenth small target perception feature map into the spatial structure perception feature calibration unit to obtain the thirteenth small target perception feature map includes: The tenth small target perception feature map is input into the horizontal average pooling unit and the vertical average pooling unit respectively to obtain the sixteenth small target perception feature map; The sixteenth small target perception feature map is sequentially input into the third convolutional unit, the first activation unit, the fourth convolutional unit, and the second activation unit to obtain the attention weight map; The thirteenth small target perception feature map is obtained by multiplying the tenth small target perception feature map and the attention weight map.

7. The method according to any one of claims 4-6, characterized in that, The dynamic environment adaptation feature fusion network unit includes: multiple depthwise separable convolutional units, multiple dynamic environment-aware feature adaptive correction units, a first average pooling unit, a fifth convolutional unit, a downsampling unit, a normalization and activation unit, a third dilated convolutional unit, and an upsampling unit. Accordingly, the step of inputting the plurality of small target perception feature maps into the dynamic environment adaptation feature fusion network unit to obtain the target environment adaptation fusion feature map includes: The second small target perception feature map is sequentially input into the first depthwise separable convolutional unit, the first dynamic environment perception feature adaptive correction unit, the first average pooling unit, the fifth convolutional unit, and the downsampling unit to obtain the first environment adaptive fusion feature map; The fourth small target perception feature map is sequentially input into the second depthwise separable convolutional unit, the second dynamic environment perception feature adaptive correction unit, and the normalization and activation unit to obtain the second environment adaptive fusion feature map; The sixth small target perception feature map is sequentially input into the third dynamic environment perception feature adaptive correction unit, the third dilated convolution unit, and the upsampling unit to obtain the third environment adaptive fusion feature map; The first environmental adaptation fusion feature map, the second environmental adaptation fusion feature map, and the third environmental adaptation fusion feature map are stitched together to obtain the target environmental adaptation fusion feature map.

8. The method according to claim 7, characterized in that, Each dynamic environment-aware feature adaptive correction unit includes: a sixth convolutional unit, a third depthwise separable convolutional unit, a batch normalization unit, a second average pooling unit, a third activation unit, a dynamic convolutional unit, and an efficient multi-scale attention unit. Accordingly, the step of sequentially inputting the second small target perception feature map into the first depthwise separable convolutional unit, the first dynamic environment perception feature adaptive correction unit, the first average pooling unit, the fifth convolutional unit, and the downsampling unit to obtain the first environment-adaptive fusion feature map includes: The second small target perception feature map is input into the first depthwise separable convolutional unit to obtain the fourth environment-adaptive fusion feature map. The fourth environmental adaptation fusion feature map is sequentially input into the sixth convolutional unit and the batch normalization unit to obtain the fifth environmental adaptation fusion feature map. The fourth environment-adaptive fusion feature map is sequentially input into the third depthwise separable convolutional unit and the second average pooling unit to obtain the sixth environment-adaptive fusion feature map. The fifth and sixth environmental adaptation fusion feature maps are overlaid and then input into the third activation unit to obtain the environmental adaptation fusion weight map. The sixth environment adaptation fusion feature map is sequentially input into the dynamic convolutional unit and the efficient multi-scale attention unit to obtain the seventh environment adaptation fusion feature map; The seventh environmental adaptation fusion feature map and the environmental adaptation fusion weight map are multiplied to obtain the eighth environmental adaptation fusion feature map; The eighth environment adaptation fusion feature map is sequentially input into the first average pooling unit, the fifth convolution unit, and the downsampling unit to obtain the first environment adaptation fusion feature map.

9. The method according to claim 5, characterized in that, The asymmetric spatial refinement unit includes: a first asymmetric convolutional layer, a second asymmetric convolutional layer, a third average pooling unit, a fourth average pooling unit, a max pooling unit, a seventh convolutional unit, and an eighth convolutional unit. Accordingly, the step of inputting the target environment adaptation fusion feature map into the asymmetric spatial refinement unit to obtain the target spatial refinement feature map includes: The target environment adaptation fusion feature map is sequentially input into the first asymmetric convolutional layer and the second asymmetric convolutional layer to extract features in the horizontal and vertical directions, thereby obtaining the first spatial refinement feature map; The first spatially refined feature map is sequentially input into the third average pooling unit, the max pooling unit, the fourth average pooling unit, and the seventh convolutional unit to obtain the second spatially refined feature map. The first spatial refinement feature map and the second spatial refinement feature map are superimposed to obtain the third spatial refinement feature map; The third spatial refinement feature map is input into the eighth convolutional unit to obtain the target spatial refinement feature map.

10. A defect detection device for energy storage equipment, characterized in that, include: The acquisition module is used to acquire the target image of the energy storage device to be tested; The processing module is used to input the target image into the energy storage device defect detection model to obtain the defect information of the energy storage device to be detected. The energy storage device defect detection model is determined based on images of multiple energy storage devices and the labeled defect information corresponding to each image. The defect information includes: defect type and defect level.

Images