Method and apparatus for detecting battery anomalies

Abstract

The present application relates to battery technology, in particular to a method and device for training a target detection model and a method and device for detecting battery abnormalities using the target detection model. The method for training a target detection model according to an aspect of the present application comprises the following steps: generating a plurality of corresponding feature maps from a plurality of battery data sets as training samples, wherein each of the battery data sets contains a time series of state parameters of a plurality of battery cells belonging to the same battery, and each of the plurality of feature maps contains a plurality of regions corresponding to the plurality of battery cells, each of the regions having a time series feature associated with the corresponding battery cell; dividing the plurality of feature maps into a training set and a validation set, and labeling the feature maps assigned to the training set; and training the target detection model using the plurality of feature maps.

Description

Technical Field

[0001] This application relates to battery technology, and more particularly to methods and apparatus for training object detection models, and methods and apparatus for detecting battery anomalies using the aforementioned object detection models. Background Technology

[0002] A power battery pack consists of multiple cells. When a cell experiences excessive self-discharge or a short circuit, it can lead to the risk of thermal runaway. Therefore, safety monitoring of vehicle power batteries is urgently needed.

[0003] Current methods for identifying anomalies in vehicle power batteries include expert-based rule-based methods using parameter calibration and data-driven methods. The former requires extensive experimental parameter calibration combined with human experience; due to its sensitivity to certain monitoring parameters, it is prone to numerous false alarms, and it lacks generalizability. The latter method extracts features from data collected by battery sensors and performs binary classification of abnormal batteries; however, this method has a high false negative rate and can only determine whether a battery is abnormal overall, failing to pinpoint the specific cell. Summary of the Invention

[0004] The purpose of this application includes providing a method and apparatus for training a target detection model, wherein the trained target detection model has good generalization ability, can locate anomalies at the cell level, and has high prediction accuracy.

[0005] The purpose of this application also includes providing methods and apparatus for detecting battery anomalies, which can locate the anomalies at the cell level and have a high predictive accuracy.

[0006] According to one aspect of this application, a method for training an object detection model is provided, the object detection model being used for battery anomaly detection, the method comprising the following steps: Multiple feature maps are generated from multiple battery datasets as training samples. Each battery dataset contains a time series of state parameters of multiple cells belonging to the same battery. Each of the multiple feature maps contains multiple regions corresponding to the multiple cells, and each region has temporal features associated with the corresponding cell. The multiple feature maps are divided into a training set and a validation set, and the feature maps assigned to the training set are labeled. The target detection model is trained using the multiple feature maps.

[0007] Optionally, in the above method, the state parameter is one or more of the following: voltage, current, and temperature.

[0008] In addition to one or more of the features mentioned above, each of the plurality of feature maps in the above method is generated in the following manner: Each cell's state parameters are used to generate its own sequence features; Using the same encoder, the coding features of each battery cell are generated from their respective sequence features as the temporal features; A projection map representing the projection of each cell onto the cross-section of the battery is generated, and the coded features of each cell are associated with the corresponding regions in the projection map to obtain the feature map of the battery. The position of each cell on the feature map is adjusted according to the set rules.

[0009] Optionally, in the above method, before training the target detection model using the training samples of the plurality of feature maps, the following steps are further included: The scales of the multiple feature maps are adjusted to a set value.

[0010] Optionally, in the above method, the sequence features are one or more of the following: the mean, standard deviation, minimum, maximum, median, minimum difference, and maximum difference of the state parameters within a time window of a set length.

[0011] Optionally, in the above method, the encoder is selected from one of the following: recurrent neural network, bidirectional recurrent neural network, long short-term memory network, bidirectional long short-term memory network, and gated recurrent unit neural network.

[0012] Optionally, in the above method, the set rules include the following rules with decreasing priority: Make adjacent cells in the current loop adjacent on the feature diagram; Cells that are spatially adjacent are adjacent on the feature map.

[0013] Optionally, in the above method, the feature maps in the training set are labeled in the following manner: If a battery cell is an abnormal cell, then the area corresponding to it on the feature map, as well as the area adjacent to the corresponding area, are marked as abnormal areas.

[0014] Optionally, in the above method, the target detection model is one of the following: YOLO model, SSD model, and CenterNet model.

[0015] According to another aspect of this application, an apparatus for training an object detection model is provided, the object detection model being used for battery anomaly detection, the apparatus comprising: At least one processor; At least one memory; A computer program stored in the memory, when executed on the processor, causes the following operations: Multiple feature maps are generated from multiple battery datasets as training samples. Each battery dataset contains a time series of state parameters of multiple cells belonging to the same battery. Each of the multiple feature maps contains multiple regions corresponding to the multiple cells, and each region has temporal features associated with the corresponding cell. The multiple feature maps are divided into a training set and a validation set, and the feature maps assigned to the training set are labeled. The target detection model is trained using the multiple feature maps.

[0016] Optionally, in the above-described device, the state parameter is one or more of the following: voltage, current, and temperature.

[0017] In addition to one or more of the features described above, each of the plurality of feature maps in the above-described apparatus is generated in the following manner: Each cell's state parameters are used to generate its own sequence features; Using the same encoder, the coding features of each battery cell are generated from their respective sequence features as the temporal features; A projection map representing the projection of each cell onto the cross-section of the battery is generated, and the coded features of each cell are associated with the corresponding regions in the projection map to obtain the feature map of the battery. The position of each cell on the feature map is adjusted according to the set rules.

[0018] Optionally, in the above apparatus, before training the target detection model using the plurality of feature map training samples, the following operations are further included: The scales of the multiple feature maps are adjusted to a set value.

[0019] Optionally, in the above-described apparatus, the sequence features are one or more of the following: the average, standard deviation, minimum, maximum, median, minimum difference, and maximum difference of the state parameters within a time window of a set length.

[0020] Optionally, in the above-described device, the encoder is selected from one of the following: recurrent neural network, bidirectional recurrent neural network, long short-term memory network, bidirectional long short-term memory network, and gated recurrent unit neural network.

[0021] Optionally, in the above-described apparatus, the set rules include the following rules with decreasing priority: Make adjacent cells in the current loop adjacent on the feature diagram; Cells that are spatially adjacent are adjacent on the feature map.

[0022] Optionally, in the above-described apparatus, the feature maps in the training set are labeled in the following manner: If a battery cell is an abnormal cell, then the area corresponding to it on the feature map, as well as the area adjacent to the corresponding area, are marked as abnormal areas.

[0023] Optionally, in the above-described apparatus, the target detection model is one of the following: YOLO model, SSD model, and CenterNet model.

[0024] According to another aspect of this application, a method for detecting battery malfunctions is provided, comprising the following steps: A feature map to be detected is generated from the dataset of the battery to be detected, wherein the dataset of the battery to be detected contains a time series of state parameters of multiple cells belonging to the battery to be detected, and the feature map to be detected contains multiple regions corresponding to the multiple cells of the battery to be detected, each region having temporal features associated with the corresponding cell. The feature map to be detected is input into the target detection model to perform battery anomaly detection, wherein the target detection model is trained by the method described above for training target detection models.

[0025] Optionally, the above method may further include the following steps: A dataset of the battery under test is generated from the sampled signals acquired by the sensor of the battery under test.

[0026] In addition to one or more of the features mentioned above, the method also includes the following steps: Output the detection results of the target detection model, including the location of the abnormal battery cell.

[0027] According to another aspect of this application, an apparatus for detecting battery malfunctions is provided, comprising: A feature map generation unit is configured to generate a feature map to be detected from a dataset of the battery to be detected, wherein the dataset of the battery to be detected contains a time series of state parameters of multiple cells belonging to the battery to be detected, and the feature map to be detected contains multiple regions corresponding to the multiple cells of the battery to be detected, each region having temporal features associated with the corresponding cell. An anomaly detection engine is configured to input the feature map to be detected into a target detection model to perform battery anomaly detection, wherein the target detection model is trained by the aforementioned apparatus for training the target detection model.

[0028] Optionally, the above-described apparatus further includes: A preprocessing unit is configured to generate a dataset of the battery under test from sampled signals acquired by the sensors of the battery under test.

[0029] In addition to one or more of the features described above, the device also includes: An output unit is configured to output the detection results of the target detection model, the detection results including the location of the abnormal battery cell.

[0030] In the embodiments of this application, the method for detecting battery anomalies is based on data-driven deep learning algorithms to extract and mine features contained in vehicle battery data. This avoids the shortcomings of parameter calibration methods and improves the accuracy of prediction. Furthermore, in some embodiments, combining recurrent neural networks and object detection algorithms to detect battery anomalies provides model stability. Moreover, in some embodiments of this application, scale unification processing enables the object detection model to be compatible with batteries of different topologies, thus exhibiting good generalization ability. Finally, in some embodiments of this application, anomaly detection can be performed at the cell level by associating regions in the feature map with battery cells and identifying anomalies in those regions. Attached Figure Description

[0031] The above and / or other aspects and advantages of this application will become clearer and more readily understood from the following description taken in conjunction with the accompanying drawings, in which the same or similar elements are denoted by the same reference numerals. The drawings include: Figure 1 This is a flowchart of a method for training an object detection model according to some embodiments of this application.

[0032] Figure 2 This is a flowchart of a method for generating feature maps for each battery dataset according to some other embodiments of this application.

[0033] Figure 3 This is a schematic diagram illustrating the annotation of a feature map according to some embodiments of this application.

[0034] Figure 4 This is a schematic block diagram of an apparatus for training an object detection model according to other embodiments of this application.

[0035] Figure 5 This is a flowchart of a method for detecting battery malfunctions according to some other embodiments of this application.

[0036] Figure 6 This is a schematic block diagram of an apparatus for detecting battery malfunctions according to other embodiments of this application. Detailed Implementation

[0037] The present application will now be described more fully with reference to the accompanying drawings, which illustrate exemplary embodiments thereof. However, the present application may be implemented in various forms and should not be construed as being limited to the embodiments given herein. The foregoing embodiments are intended to make the disclosure herein complete and thorough, so as to more fully convey the scope of protection of the present application to those skilled in the art.

[0038] In this specification, terms such as “comprising” and “including” indicate that, in addition to having the units and steps that are directly and explicitly stated in the specification and claims, the technical solution of this application does not exclude the presence of other units and steps that are not directly or explicitly stated.

[0039] Figure 1 This is a flowchart illustrating a method for training an object detection model according to some embodiments of this application, the object detection model being used for battery anomaly detection. The object detection models described herein include various convolutional neural network models that employ object detection algorithms to identify objects of interest (e.g., abnormal battery cells) in an image. Examples of object detection models include, but are not limited to, the YOLO model, the SSD model, and the CenterNet model.

[0040] See Figure 1 The method shown includes the following steps: Step 110: Generation of feature maps as training samples In step 110, multiple feature maps are generated from multiple battery datasets as training samples. As mentioned above, a power battery pack may contain multiple cells. In some embodiments, each battery dataset corresponds to a battery or power battery pack and contains a time series of state parameters belonging to multiple cells of that battery. The state parameters mentioned here should be understood as various parameters that can reflect the working state of the cell, including but not limited to the cell's voltage, current, and temperature. Each feature map is used to characterize the characteristics of the state parameters of each cell in a battery changing over time (hereinafter also referred to as temporal features). In order to enable the target detection model to be used to locate anomalies in the cells, multiple cells belonging to the same battery are represented in the form of feature maps containing multiple regions, with each cell having a corresponding region in the feature map; in addition, the temporal features of each cell are assigned to or given their respective corresponding regions. The content of the feature maps will be further described below.

[0041] In some embodiments, it may be used Figure 2 The method shown generates corresponding feature maps for each battery dataset as training samples.

[0042] Specifically, Figure 2The method flow shown begins at step 210. In this step, a time series of state parameters for each cell of a battery is generated as a dataset for that battery. Exemplarily, the time series of state parameters can be generated in the cloud or on-vehicle as follows: First, sampling signals from different cell sensors within the battery are acquired (e.g., one or more of voltage, current, and temperature sampling signals). For example, the cloud receives sampling signals from the vehicle's battery management system, or the battery management system receives sampling signals from the cell sensors.

[0043] Since the sampling frequencies or sampling times of multiple battery cell sensors may not be consistent, time alignment processing needs to be performed at the vehicle end or the cloud. Specifically, for time points with corresponding sampling signals, these signals can be directly used as the state parameter values ​​at that time point. For time points without corresponding sampling signals, interpolation algorithms can be used to calculate the state parameter values ​​at that time point. Through time alignment processing, time series of state parameters for multiple battery cells will be obtained, and these time series will have the same timing.

[0044] After step 210, Figure 2 The method shown proceeds to step 220. In this step, a sequence feature is generated from the time series of the state parameters of each cell. Optionally, the sequence feature may be one or more of the following: the mean, standard deviation, minimum, maximum, median, minimum difference, and maximum difference of the state parameters within a time window of a set length (e.g., the length of the time series).

[0045] The process then proceeds to step 230. In this step, an encoder is used to generate coded features from the sequence features of each cell, which serve as the aforementioned temporal features. Encoding the sequence features further highlights the time-varying characteristics of the cell's state parameters.

[0046] In some embodiments, a recurrent neural network can be used to implement the encoder used in step 230. Specifically, the recurrent neural network includes an input layer, a hidden layer, and an output layer. The sequence features of each cell are fed into the input nodes of the input layer, and the corresponding encoded features are output in the output layer. Various encoders can be used to generate the encoded features, including but not limited to recurrent neural networks, bidirectional recurrent neural networks, long short-term memory networks, bidirectional long short-term memory networks, gated recurrent unit neural networks, etc.

[0047] In step 230, the same encoder (i.e., an encoder with the same structure and configuration parameters) can be used for different battery cells. Furthermore, the same encoder can also be used for different battery datasets.

[0048] Next, proceed to step 240. In this step, a projection map of each cell on the cross-section of the battery is generated, and the encoded features of each cell generated in step 230 are associated with the corresponding regions in the projection map to obtain the feature map of the battery. For example, the bottom cross-section of the battery can be used as the projection plane, and the physical location of the cells inside the battery is projected onto the projection map, thus forming multiple regions on the projection map, each region corresponding to one of the cells. Therefore, the encoded features of each cell can be "embedded" into the corresponding region in the projection map, that is, the encoded features are associated with the corresponding region in the projection map to obtain the feature map.

[0049] Optionally, after step 240, Figure 2 The method then proceeds to step 250. In this step, the positions of the individual cells on the feature map are adjusted. In some embodiments, the adjustment may be based on one or more predefined rules to bring cells with electrical property dependencies and / or spatial dependencies as close as possible. Adjustment rules with different priority levels may be used; for example, in one example, the adjustment rules include: Rule A: Make adjacent cells in the current loop (e.g., two cells directly connected) adjacent on the projection diagram.

[0050] Rule B: Make cells that are spatially adjacent (e.g., two cells that are physically adjacent) adjacent on the projection map.

[0051] In the example above, the priority of rule A can be set to be higher than that of rule B, or the priority of rule B can be set to be higher than that of rule A.

[0052] Step 120: Annotation of feature maps See Figure 1 After performing step 110, Figure 1 The method flow shown then proceeds to step 120. In this step, the multiple feature maps generated in step 110 are divided into a training set and a validation set, and the feature maps assigned to the training set are labeled.

[0053] Figure 3 This is a schematic diagram illustrating the annotation of a feature map according to some embodiments of this application. Figure 3 The feature map shown depicts nine small rectangles or regions, each representing one of the battery cells. The gray-filled areas represent cells identified as being in an abnormal state. In an exemplary annotation method, the abnormal cell represented by the gray area, along with its surrounding area, is also labeled as an abnormal region. That is, as... Figure 3 As shown, the large rectangular box formed by the gray area and its surrounding areas (a total of 9 areas) corresponds to the outlier area (hereinafter also referred to as the target box or prediction box).

[0054] Step 130: Adjusting the feature map scale In order to Figure 1 The training method shown is applicable to anomaly detection for various types of batteries. The scale of the feature map generated in step 120 can be adjusted to a set value to obtain a feature map with a uniform scale. For this purpose, after performing step 120, Figure 1 The illustrated method flow then proceeds to step 130. In this step, a bilinear interpolation algorithm can be used to adjust the scale of the feature map to a set value. Examples of bilinear interpolation algorithms include ROI Align and ROIPooling.

[0055] It should be noted that step 130 is optional. For example, when the trained object detection model is only for a specific type of battery, scale adjustment is unnecessary.

[0056] Step 140: Training the object detection model See Figure 1 After performing step 130, Figure 1 The method flow shown then proceeds to step 140. In this step, feature maps with a uniform scale are used as training samples to train the object detection model.

[0057] In some embodiments, the object detection model may include convolutional layers, ReLU layers, detection branch groups, and a post-processing module. The convolutional layers use convolutional algorithms to process the encoded features of each region in the feature map, thereby extracting the features contained within the feature map. The convolutional algorithms used include, but are not limited to, grouped convolution, separable convolution, dilated convolution, and depthwise convolution. The ReLU layers are used to activate the features extracted by the convolutional layers to obtain a deep feature map, which is then output to the detection branch group. The detection branch group includes an object classification branch, a center point offset prediction branch, and a size prediction branch. The object classification branch determines the score for the presence of an abnormal cell near the center point of each grid cell in the deep feature map; the center point offset prediction branch determines the predicted offset of the center point of the prediction box relative to the grid center point; and the size prediction branch determines the predicted size of the prediction box. The processing results of the detection branch group are output to the post-processing module to obtain the final processing result. For example, the post-processing module can be configured to select cells whose scores exceed an alarm threshold, then offset the center point of the cells exceeding the alarm threshold according to the offset prediction value to obtain the center point of the prediction box, and determine the range of the prediction box based on the size prediction value with the center point of the prediction box as the center.

[0058] In some embodiments, the following form of function can be used as the loss function Ldet during the training process of the object detection model: (1) in The loss component that contributes to the classification branch of the target category can be described, for example, by the binary cross-entropy function. The loss component representing the contribution of the center point offset prediction branch can be described, for example, by the root mean square error function. The loss component contributing to the size prediction branch can be described, for example, by the root mean square error function. and These are the weighting factors for the contribution components of the size prediction branch and the center point offset prediction branch, respectively.

[0059] As described above, the feature maps are divided into training and validation sets. In an exemplary training process, a target prediction model can be trained using the feature maps in the training set based on a loss function. The trained model is then validated on the validation set, and the weights of the model with the highest accuracy are saved. This process is repeated until the model converges or the maximum number of iterations is reached.

[0060] It should be noted that the above uses Figure 1 The described method can be implemented collaboratively by the cloud and the vehicle. In this scenario, the vehicle (e.g., the battery management system) periodically or irregularly reports sampling signals collected by sensors in each battery cell within the battery pack to the cloud. The cloud then executes the above-mentioned functions based on the received sampling signals. Figure 1 The described battery anomaly detection method can also be optionally implemented at the vehicle end (e.g., by the battery management system). In this case, the battery management system can periodically or irregularly receive sampling signals from the cell sensors and perform the above-described method based on the received sampling signals. Figure 1 The described battery anomaly detection method.

[0061] Figure 4 This is a schematic block diagram of an apparatus for training an object detection model according to other embodiments of this application.

[0062] Figure 4 The illustrated device 40 includes one or more memories 410 (e.g., non-volatile memory such as flash memory, ROM, hard disk drive, magnetic disk, optical disk, etc.), one or more processors 420, and a computer program 430 stored on the memories 410.

[0063] Computer program 430 can be run by one processor 420, or by multiple processors 420 (which may be located in the same physical location or in different physical locations) working together to perform the above-mentioned tasks. Figure 1 and 2 The described method and steps.

[0064] Optionally, Figure 4The device 40 shown can be implemented as a cloud computing platform or cloud platform, in which computing, networking, and storage services are provided in the form of resource pools, where most computing resources are hosted on the Internet and allocated to users on demand. Alternatively, the device 40 can also be implemented as a standalone system, in which computing resources are basically limited to a single local system, and computer applications, data storage, and processing capabilities are close to or relatively close to the user.

[0065] Figure 5 This is a flowchart of a method for detecting battery malfunctions according to some other embodiments of this application.

[0066] Figure 5 The method shown includes the following steps: Step 510: Generation of the dataset for the batteries to be tested In step 510, a dataset of the battery under test is generated from sampling signals (e.g., one or more of voltage sampling signals, current sampling signals, and temperature sampling signals) obtained by cell sensors within the battery under test. This dataset contains time series of state parameters belonging to multiple cells of the battery under test.

[0067] In some embodiments, the cloud receives sampling signals from the vehicle's battery management system and then performs time alignment processing to obtain time series of state parameters for multiple cells, these time series having the same timing. In other embodiments, the battery management system receives sampling signals from cell sensors and then performs time alignment processing to obtain multiple time series of state parameters with the same timing.

[0068] Step 520: Generation of the feature map to be detected In step 520, a corresponding feature map or a feature map to be detected is generated from the dataset of the battery to be detected. The feature map to be detected may contain multiple regions corresponding to multiple cells of the battery to be detected, and each region has temporal features associated with its corresponding cell. The method for generating the feature map has been referred to above. Figure 1 and 2 The above has been described in detail, so I will not repeat it here.

[0069] Step 530: Adjusting the feature map scale Optionally, in step 530, the scale of the feature map can be adjusted to a set value so that the feature map of the battery to be detected has the same scale as the training samples used to train the target detection model. The scale adjustment can be achieved using a bilinear interpolation algorithm, such as ROI Align and ROI Pooling.

[0070] Step 540: Battery Malfunction Detection In step 540, the scale-adjusted feature map is input into the object detection model to perform battery anomaly detection. The object detection model used for anomaly detection could, for example, utilize the above-mentioned reference... Figure 1-3 The target detection model obtained by the training method described above.

[0071] Step 550: Output of detection results In step 550, the detection results of the target detection model are output. Optionally, the detection results include an indication of whether the battery is abnormal and the location of the abnormal cell when the abnormality occurs (or an identifier of the abnormal cell, which is equivalent to the cell's location). For example, when the battery abnormality detection is completed in the cloud, the cloud can return the detection results to the vehicle.

[0072] Figure 6 This is a schematic block diagram of an apparatus for detecting battery malfunctions according to other embodiments of this application.

[0073] Figure 6 The apparatus 60 shown includes a preprocessing unit 610, a feature map generation unit 620, an anomaly detection engine 630, and an output unit 640.

[0074] The preprocessing unit 610 is configured to generate a dataset of the battery to be tested from the sampling signal acquired by the sensor of the battery to be tested. The feature map generation unit 620 is coupled to the preprocessing unit 610 and the anomaly detection engine 630, and is configured to generate a feature map to be detected from the dataset of the battery to be tested and output the generated feature map to the anomaly detection engine 630. The method for generating the feature map has been described above. Figure 1 and 2 The above description has been fully provided and will not be repeated here. Optionally, the feature map generation unit 620 is also configured to adjust the scale of the feature map to be the same as that of the training samples used to train the object detection model. The anomaly detection engine 630 includes an object detection model configured to input the feature map to be detected into the object detection model to perform battery anomaly detection and output the detection result to the output unit 640, wherein the object detection model used for anomaly detection may, for example, utilize the above-mentioned reference Figure 1-3 The target detection model obtained by the training method described above. The output unit 640 is configured to output the detection results of the target detection model.

[0075] Optionally, Figure 6The device 60 shown can be implemented as a cloud computing platform or cloud platform, in which computing, networking, and storage services are provided in the form of resource pools, where most computing resources are hosted on the Internet and allocated to users on demand. Alternatively, the device 60 can also be implemented as a standalone system, in which computing resources are basically limited to a single local system, and computer applications, data storage, and processing capabilities are close to or relatively close to the user.

[0076] The embodiments and examples presented herein are provided to best illustrate embodiments according to the present technology and its particular applications, thereby enabling those skilled in the art to implement and use the application. However, those skilled in the art will understand that the above description and examples are provided for ease of illustration and example only. The descriptions presented are not intended to cover all aspects of the application or to limit the application to the precise forms disclosed.

Claims

1. A method for training an object detection model, the object detection model being used for battery anomaly detection, the method comprising the following steps: Multiple feature maps are generated from multiple battery datasets as training samples, where, Each of the battery datasets contains a time series of state parameters of multiple cells belonging to the same battery, and each of the multiple feature maps contains multiple regions corresponding to the multiple cells, and each region has temporal features associated with the corresponding cell; The multiple feature maps are divided into a training set and a validation set, and the feature maps assigned to the training set are labeled. The feature maps in the training set are labeled in the following way: if a battery cell is an abnormal battery cell, the region corresponding to it on the feature map and the region adjacent to the corresponding region are labeled as abnormal regions. The target detection model is trained using the multiple feature maps. Each of the plurality of feature maps is generated in the following manner: Each cell's state parameters are used to generate its own sequence features; Using the same encoder, the coding features of each battery cell are generated from their respective sequence features as the temporal features; A projection map representing the projection of each cell onto the cross-section of the battery is generated, and the coded features of each cell are associated with the corresponding regions in the projection map to obtain the feature map of the battery. The positions of each battery cell on the feature map are adjusted according to the set rules. The rules set include the following rules with decreasing priority: Make adjacent cells in the current loop adjacent on the feature diagram; Cells that are spatially adjacent are adjacent on the feature map.

2. The method as described in claim 1, wherein, The state parameters are one or more of the following: voltage, current, and temperature.

3. The method as described in claim 1, wherein, Before training the object detection model using the multiple feature map training samples, the following steps are also included: The scales of the multiple feature maps are adjusted to a set value.

4. The method of claim 1, wherein, The sequence features are one or more of the following: the mean, standard deviation, minimum, maximum, median, minimum difference, and maximum difference of the state parameters within a time window of a set length.

5. The method of claim 1, wherein, The encoder is selected from one of the following: recurrent neural network, bidirectional recurrent neural network, long short-term memory network, bidirectional long short-term memory network, and gated recurrent unit neural network.

6. The method of claim 1, wherein, The target detection model is one of the following: YOLO model, SSD model, and CenterNet model.

7. An apparatus for training an object detection model, the object detection model being used for battery anomaly detection, the apparatus comprising: At least one processor; At least one memory; A computer program stored in the memory, when executed on the processor, causes the following operations: Multiple feature maps are generated from multiple battery datasets as training samples. Each battery dataset contains a time series of state parameters of multiple cells belonging to the same battery. Each of the multiple feature maps contains multiple regions corresponding to the multiple cells, and each region has temporal features associated with the corresponding cell. The multiple feature maps are divided into a training set and a validation set, and the feature maps assigned to the training set are labeled. The feature maps in the training set are labeled in the following way: if a battery cell is an abnormal battery cell, the region corresponding to it on the feature map and the region adjacent to the corresponding region are labeled as abnormal regions. The target detection model is trained using the multiple feature maps. Each of the plurality of feature maps is generated in the following manner: Each cell's state parameters are used to generate its own sequence features; Using the same encoder, the coding features of each battery cell are generated from their respective sequence features as the temporal features; A projection map representing the projection of each cell onto the cross-section of the battery is generated, and the coded features of each cell are associated with the corresponding regions in the projection map to obtain the feature map of the battery. The positions of each battery cell on the feature map are adjusted according to the set rules. The rules set include the following rules with decreasing priority: Make adjacent cells in the current loop adjacent on the feature diagram; Cells that are spatially adjacent are adjacent on the feature map.

8. The apparatus of claim 7, wherein, The state parameters are one or more of the following: voltage, current, and temperature.

9. The apparatus of claim 7, wherein, Before training the object detection model using the multiple feature map training samples, the following operations are also included: The scales of the multiple feature maps are adjusted to a set value.

10. The apparatus of claim 7, wherein, The sequence features are one or more of the following: the mean, standard deviation, minimum, maximum, median, minimum difference, and maximum difference of the state parameters within a time window of a set length.

11. The apparatus of claim 7, wherein, The encoder is selected from one of the following: recurrent neural network, bidirectional recurrent neural network, long short-term memory network, bidirectional long short-term memory network, and gated recurrent unit neural network.

12. The apparatus of claim 7, wherein, The target detection model is one of the following: YOLO model, SSD model, and CenterNet model.

13. A method for detecting battery malfunctions, comprising the following steps: A feature map to be detected is generated from the dataset of batteries to be detected, where, The dataset of the battery to be tested contains time series of state parameters of multiple cells belonging to the battery to be tested, and the feature map to be tested contains multiple regions corresponding to the multiple cells of the battery to be tested, each region having temporal features associated with the corresponding cell. The feature map to be detected is input into the target detection model to perform battery anomaly detection, wherein the target detection model is trained by the method described in any one of claims 1-6.

14. The method of claim 13, wherein, It also includes the following steps: A dataset of the battery under test is generated from the sampled signals acquired by the sensor of the battery under test.

15. The method of claim 13 or 14, wherein, It also includes the following steps: Output the detection results of the target detection model, including the location of the abnormal battery cell.

16. An apparatus for detecting battery malfunctions, comprising: A feature map generation unit is configured to generate a feature map to be detected from a dataset of the battery to be detected, wherein the dataset of the battery to be detected contains a time series of state parameters of multiple cells belonging to the battery to be detected, and the feature map to be detected contains multiple regions corresponding to the multiple cells of the battery to be detected, each region having temporal features associated with the corresponding cell. An anomaly detection engine is configured to input the feature map to be detected into a target detection model to perform battery anomaly detection, wherein the target detection model is trained by the apparatus as described in any one of claims 7-12.

17. The apparatus of claim 16, wherein, Also includes: A preprocessing unit is configured to generate a dataset of the battery under test from sampled signals acquired by the sensors of the battery under test.

18. The apparatus of claim 16 or 17, wherein, Also includes: An output unit is configured to output the detection results of the target detection model, the detection results including the location of the abnormal battery cell.

Images