Method for training feature extraction model, feature extraction method, and thermal runaway prediction method and apparatus

Abstract

Provided in the present application are a method for training a feature extraction model, a feature extraction method, and a thermal runaway prediction method and apparatus. The method for training a feature extraction model comprises: acquiring historical battery data before a thermal runaway occurs in a battery; extracting training samples from the historical battery data on the basis of a dynamic sliding window, wherein the step size of the dynamic sliding window is determined on the basis of the similarity between training samples respectively corresponding to the current sliding window and the previous sliding window; and using the training samples to train a residual network, so as to obtain a trained feature extraction model. In the present application, training samples are extracted from historical battery data by means of a dynamic sliding window, wherein when the historical battery data varies slightly, training samples are extracted using a sliding window having a large step size; and when the historical battery data varies greatly, training samples are extracted using a sliding window having a small step size, which reduces the number of similar extracted training samples, thereby reducing the computation amount in model training and the consumption of hardware resources, and improving the training efficiency.

Description

Training method of feature extraction model, feature extraction method, thermal runaway prediction method and device TECHNICAL FIELD

[0001] The present application relates to the technical field of computer models, in particular to a training method of a feature extraction model, a feature extraction method, a thermal runaway prediction method and device. BACKGROUND

[0002] Lithium-ion batteries are widely used in energy storage systems and electric vehicles due to their high power density and energy density, long cycle life, and low self-discharge rate. However, with the widespread use and popularity, a series of accidents have occurred. In particular, accidents such as spontaneous combustion and explosion caused by thermal runaway can have serious consequences.

[0003] Existing thermal runaway prediction methods use feature extraction models to extract useful feature information from relevant parameters of the battery before thermal runaway, and make thermal runaway predictions based on the feature information. However, before thermal runaway, the battery is in a normal working state for a long time, and its battery-related parameters are basically the same. When training the feature extraction model, a large number of similar features will be collected, increasing the computational load during model training and reducing the efficiency of model training. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a training method of a feature extraction model, a feature extraction method, and a thermal runaway prediction method and device to improve the efficiency of model training and the consumption of hardware resources.

[0005] In a first aspect, the embodiments of the present application provide a training method of a feature extraction model, comprising:

[0006] Obtaining battery historical data of the battery before thermal runaway occurs;

[0007] Extracting training samples from the battery historical data based on a dynamic sliding window; wherein the window size of the dynamic sliding window is a preset size, and the step size is determined according to the similarity between the training samples corresponding to the current sliding window and the previous sliding window;

[0008] Training a residual network using the training samples to obtain a trained feature extraction model.

[0009] The embodiments of the present application extract training samples from battery historical data using a dynamic sliding window. When the battery historical data changes little, a large step size sliding window is used to extract training samples; when the battery historical data changes greatly, a smaller step size sliding window is used to extract training samples, thereby reducing the number of similar training samples extracted, reducing the computational load during model training, reducing the consumption of hardware resources, and improving the efficiency of training.

[0010] In any embodiment, the training sample is extracted from the battery historical data based on a dynamic sliding window, comprising:

[0011] obtaining a first training sample extracted by a current sliding window and a second training sample extracted by a previous sliding window;

[0012] calculating a similarity between the first training sample and the second training sample;

[0013] determining a step length of a next sliding window according to the similarity and a step length corresponding to the current sliding window;

[0014] extracting a training sample corresponding to the next sliding window from the battery historical data based on a window size and the step length.

[0015] The embodiment of the present application adjusts the step length of the sliding window based on the similarity by calculating the similarity between the training samples corresponding to the adjacent two sliding windows, so that the interval of collecting data can be adjusted to be slightly larger when the battery historical data changes slightly, and the data is collected more densely when the battery historical data changes greatly, so that sufficient useful features can be collected, and the calculation amount of model training is reduced.

[0016] In any embodiment, the step length of the current sliding window is determined according to the similarity and the step length corresponding to the current sliding window, comprising:

[0017] if the similarity is greater than a first threshold value, a first preset step length is added to the step length corresponding to the current sliding window;

[0018] if the similarity is less than a second threshold value, a second preset step length is reduced from the step length corresponding to the current sliding window.

[0019] The embodiment of the present application adjusts the step length of the sliding window based on the similarity, by calculating the similarity between the training samples corresponding to the adjacent two sliding windows, so that the data collection interval can be adjusted to be slightly larger when the battery historical data changes slightly, and the data is more densely collected when the battery historical data changes greatly, so that sufficient useful features can be collected, and the computation amount of model training is reduced.

[0020] In any embodiment, the first training sample comprises a plurality of first training data, and the second training sample comprises a plurality of second training data, and the similarity between the first training sample and the second training sample is calculated, comprising:

[0021] a distance matrix is calculated according to the first training data and the second training data;

[0022] a shortest path from a top-left element to a bottom-right element of the distance matrix is determined;

[0023] Similarity is determined based on the shortest path.

[0024] This application embodiment calculates the distance matrix between the first training data and the second training data, and calculates the shortest path from the top left element to the bottom right element of the distance matrix. A larger shortest path indicates a smaller similarity between the two training data, and a smaller shortest path indicates a larger similarity between the two training data. This calculation method can be used to calculate the similarity of data with time-series characteristics.

[0025] In any embodiment, the residual network includes an input module, a first residual module, and a second residual module; training the residual network using training samples includes:

[0026] Obtain the amount of data for the training samples;

[0027] The number of the first residual module and the second residual module is determined based on the amount of data.

[0028] A residual network is generated based on the number of the first residual module and the second residual module;

[0029] The residual network is trained by using training samples as input to the model and the features of the training samples as labels.

[0030] In this embodiment of the application, during model training, the number of first residual modules and second residual modules in the residual network is determined according to the number of samples. When the number of samples is large, more first residual modules and second residual modules are used to fully extract features; when the number of samples is small, fewer first residual modules and second residual modules are used to reduce the resource consumption of network training.

[0031] Secondly, embodiments of this application provide a feature extraction method, including:

[0032] Collect battery data from batteries within a preset time period prior to the current time;

[0033] Battery data is input into a feature extraction model to obtain battery features; wherein the feature extraction model is trained using the training method described in the first aspect.

[0034] Thirdly, embodiments of this application provide a method for predicting thermal runaway, including:

[0035] Obtain the battery features of the battery to be predicted; the battery features are obtained by extracting them using the feature extraction method described in the second aspect.

[0036] Battery characteristics are input into the thermal runaway prediction model to obtain the prediction results output by the thermal runaway prediction model; the prediction results characterize whether the battery has the possibility of thermal runaway.

[0037] Fourthly, embodiments of this application provide a training apparatus for a feature extraction model, comprising:

[0038] The historical data acquisition module is used to acquire historical battery data before thermal runaway occurs.

[0039] The sample extraction module is used to extract training samples from the battery history data based on a dynamic sliding window; wherein the window size of the dynamic sliding window is a preset size, and the step size is determined based on the similarity between the training samples corresponding to the first two sliding windows.

[0040] The training module is used to train the residual network using training samples to obtain a trained feature extraction model.

[0041] Fifthly, embodiments of this application provide a feature extraction apparatus, comprising:

[0042] The battery data acquisition module is used to collect battery data for a preset time period up to the current time.

[0043] The feature extraction module is used to input battery data into the feature extraction model to obtain battery features; wherein the feature extraction model is trained using the training method described in the first aspect.

[0044] Sixthly, embodiments of this application provide a thermal runaway prediction device, comprising:

[0045] The feature acquisition module is used to acquire the battery features of the battery to be predicted; the battery features are obtained by extracting the features using the feature extraction method described in the second aspect.

[0046] The prediction module is used to input battery characteristics into the thermal runaway prediction model and obtain the prediction results output by the thermal runaway prediction model; the prediction results characterize whether the battery has the possibility of thermal runaway.

[0047] In a seventh aspect, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus, wherein,

[0048] The processor and the memory communicate with each other via the bus;

[0049] The memory stores program instructions that can be executed by the processor, and the processor can execute the first, second, or third aspect of the method by calling the program instructions.

[0050] Eighthly, embodiments of this application provide a non-transitory computer-readable storage medium, comprising:

[0051] The non-transitory computer-readable storage medium stores computer instructions that cause the computer to perform the methods of the first aspect, the second aspect, or the third aspect.

[0052] Ninthly, embodiments of this application provide a computer program product, including computer program instructions, which, when read and executed by a processor, perform the methods of the first aspect, the second aspect, or the third aspect.

[0053] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 is a schematic diagram of a training method for a feature extraction model provided in an embodiment of this application;

[0056] Figure 2 is a schematic diagram of a residual block provided in an embodiment of this application;

[0057] Figure 3 is a schematic diagram of the residual network model structure provided in the embodiment of this application;

[0058] Figure 4 is a schematic flowchart of a feature extraction method provided in an embodiment of this application;

[0059] Figure 5 is a schematic diagram of a method for collecting training samples using a sliding time window, as provided in an embodiment of this application.

[0060] Figure 6 is a schematic flowchart of a thermal runaway prediction method provided in an embodiment of this application;

[0061] Figure 7 is a schematic diagram of a training device for a feature extraction model provided in an embodiment of this application;

[0062] Figure 8 is a schematic diagram of a feature extraction device provided in an embodiment of this application;

[0063] Figure 9 is a schematic diagram of a thermal runaway prediction device provided in an embodiment of this application;

[0064] Figure 10 is a schematic diagram of the physical structure of the electronic device provided in the embodiment of this application. Detailed Implementation

[0065] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0067] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0068] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0069] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0070] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0071] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0072] The safety of power batteries and energy storage batteries is one of the most important performance indicators. During battery use, thermal runaway can occur. Thermal runaway refers to the phenomenon where the internal separator of a single lithium-ion battery melts due to factors such as puncture or overheating, causing a rapid electrochemical reaction that generates heat and leads to a sharp rise in the battery's temperature. Typically, thermal runaway first occurs in a single lithium-ion battery. The large amount of heat released during thermal runaway causes a rapid temperature increase in that battery. The heat spreads to nearby lithium-ion batteries, causing them to overheat and continue to experience thermal runaway. Ultimately, this can lead to the entire lithium-ion battery pack igniting, resulting in serious consequences.

[0073] Therefore, predicting thermal runaway is of great significance. Specifically, a model can be used to extract features from the parameters of the battery during use, and then thermal runaway can be predicted based on these extracted features. While training the model can utilize data from before thermal runaway occurs, in actual battery use, the data before thermal runaway is often similar. Using this data for feature extraction would result in many similar features being extracted, increasing the computational cost of model training and reducing its efficiency.

[0074] To address the aforementioned technical problems, this application provides a dynamic sliding window for extracting training samples. The step size of the sliding window is determined based on the training samples corresponding to two adjacent sliding windows. This minimizes the extraction of similar training samples and maximizes the extraction of different training samples, thereby reducing the computational load of training.

[0075] It is understood that the training method of the feature extraction model provided in this application embodiment can be used not only in battery thermal runaway scenarios, but also for training models using data with time-series features in other scenarios as training samples.

[0076] The model training method, feature extraction method, and thermal runaway prediction method provided in this application can be applied to electronic devices, including terminals and servers. Specifically, the terminal can be a smartphone, tablet computer, computer, personal digital assistant (PDA), etc.; the server can be an application server or a web server. Furthermore, the model training method, feature extraction method, and thermal runaway prediction method can all be executed on the same terminal device or on different terminal devices.

[0077] To facilitate understanding, the technical solutions provided in the embodiments of this application will be described below using a terminal device as the execution subject, illustrating the application scenarios of the model training method, feature extraction method, and thermal runaway prediction method provided in the embodiments of this application.

[0078] Figure 1 is a schematic flowchart of a training method for a feature extraction model provided in an embodiment of this application. As shown in Figure 1, the method includes:

[0079] Step 101: Obtain historical battery data prior to thermal runaway;

[0080] Step 102: Extract training samples from battery history data based on a dynamic sliding window; wherein, the window size of the dynamic sliding window is a preset size, and the step size is determined based on the similarity between the training samples corresponding to the current sliding window and the previous sliding window respectively;

[0081] Step 103: Train the residual network using training samples to obtain a trained feature extraction model.

[0082] In specific implementation, the battery can be a power battery or an energy storage battery. Battery history data refers to relevant parameters collected by acquisition devices during battery use, such as temperature, voltage, and current data during charging and discharging. It is understood that battery history data are parameters that characterize whether the battery has experienced thermal runaway. Since the main purpose of this application's embodiments is to predict whether the battery will experience thermal runaway, the battery history data obtained is the data prior to thermal runaway.

[0083] A dynamic sliding window refers to a window with a preset size, whose step size is dynamically adjusted according to the changing trend of battery history data. A window of the preset size is used to extract a period of battery history data as a training sample. The step size is determined based on the training sample corresponding to the current sliding window and the training sample corresponding to the previous sliding window. For example, if the similarity between the training sample corresponding to the current sliding window and the training sample corresponding to the previous sliding window is poor, it is considered that the battery history data has not changed much, and the step size can be increased to reduce the collection of too many similar training samples; conversely, if the similarity between the training sample corresponding to the current sliding window and the training sample corresponding to the previous sliding window is low, it is considered that the battery history data has changed significantly, and the step size can be decreased to collect as many features as possible.

[0084] After collecting training samples using a dynamic sliding window, these multiple training samples, possessing temporal sequence properties, are sorted chronologically and used as a single training sample to train the residual network. That is, a single training sample serves as the input to the residual network, with the occurrence of thermal runaway serving as a label to optimize the network's internal parameters. It should be noted that this training process is completed through multiple iterations; that is, historical data from multiple batteries prior to thermal runaway can be collected, and then this historical data is used to iteratively train the residual network.

[0085] A residual network is composed of many parts with similar structures, each typically called a "residual block." Figure 2 shows a schematic diagram of a residual block structure provided in an embodiment of this application. Assume H(x) is the bottom-level mapping of the neural network, and x represents the first-layer input of these layers. We expect the bottom-level mapping of the neural network to approach H(x)-x, i.e., to approach F(x) = H(x)-x. The original H(x) = F(x)+x. The advantage of a residual network is that it reduces the problem of learning the complete output to the problem of learning the residual. If only the output of the shallow layers is mapped using an identity (i.e., F(x) = 0) and input into the deeper layers, network degradation will not occur even as the network deepens. Therefore, a "short-circuit" mechanism is added to the network to solve the gradient vanishing problem.

[0086] Residual networks have become a mainstream model in the field of image processing, and are mostly used for feature extraction of two-dimensional images. However, battery thermal runaway data is one-dimensional time-series data. When designing neural networks for one-dimensional time-series data, the problem of network degradation is also faced. The embodiments of this application integrate the idea of ​​residual blocks to design a residual network for thermal runaway scenarios.

[0087] This application embodiment extracts training samples from battery history data using a dynamic sliding window. When the battery history data changes little, a sliding window with a large step size is used to extract training samples; when the battery history data changes significantly, a sliding window with a smaller step size is used to extract training samples. This reduces the number of similar training samples extracted, thereby reducing the computational load during model training, reducing hardware resource consumption, and improving training efficiency.

[0088] Based on the above embodiments, training samples are extracted from battery history data using a dynamic sliding window, including:

[0089] Get the first training sample extracted by the current sliding window and the second training sample extracted by the previous sliding window;

[0090] Calculate the similarity between the first training sample and the second training sample;

[0091] The step size of the next sliding window is determined based on the similarity and the step size corresponding to the current sliding window.

[0092] The training samples corresponding to the next sliding window are extracted from the battery history data based on the window size and step size.

[0093] In practical implementation, if the battery is in normal operating condition, its related parameters are relatively stable. However, when thermal runaway is imminent, these parameters often undergo significant changes and exhibit certain characteristics. Since the battery operates normally most of the time, a large number of training samples may be collected during normal operating conditions, undoubtedly increasing the subsequent computational load. Furthermore, these training samples do not offer significant help in predicting thermal runaway. The characteristics of battery parameters when thermal runaway is imminent are of great reference value for prediction; therefore, it is necessary to collect more features during this period. Therefore, this embodiment employs a dynamic sliding window approach for training sample collection.

[0094] The current sliding window refers to the sliding window corresponding to the first training sample last extracted from the battery history data at the current moment. The previous sliding window is the sliding window corresponding to the second training sample extracted from the battery history data the time before the current sliding window. After obtaining the first and second training samples, the similarity between the first and second training samples can be calculated.

[0095] The step size of the current sliding window refers to the step size between the previous sliding window and the current sliding window. The step size of the next sliding window refers to the step size between the current sliding window and the next sliding window. If the similarity is high, it means that the first training sample and the second training sample are quite similar, and the battery history data has not changed much. In this case, the step size of the current sliding window can be increased again as the step size of the next sliding window. If the similarity is low, it means that the first training sample and the second training sample are quite different, and the battery history data has changed significantly. In this case, the step size of the current sliding window can be decreased as the step size of the next sliding window. After obtaining the step size of the next sliding window, the training sample corresponding to the next sliding window can be extracted from the battery history data based on the window size and the step size. Therefore, after each training sample is collected using the sliding window, the similarity between the extracted training sample and the previously extracted training sample is calculated. Then, the next step size is determined based on the calculated similarity. The window is then moved using this step size to continue collecting training samples until the collection is complete.

[0096] This application embodiment calculates the similarity between training samples corresponding to two adjacent sliding windows and adjusts the step size of the sliding window based on the similarity. This allows for a slightly larger data collection interval when the battery history data changes little and a more intensive data collection when the battery history data changes significantly. This approach can collect a sufficient number of useful features while reducing the computational load of model training.

[0097] Based on the above embodiments, the step size of the current sliding window is determined according to the similarity and the step size corresponding to the current sliding window, including:

[0098] If the similarity is greater than the first threshold, then the first preset step size is increased based on the step size corresponding to the current sliding window.

[0099] If the similarity is less than the second threshold, then the second preset step size is reduced based on the step size corresponding to the current sliding window.

[0100] The first and second thresholds can be set according to actual conditions. For example, the first threshold may include, but is not limited to, 80%, and the second threshold may include, but is not limited to, 50%. When the similarity is greater than 80%, it indicates that the battery's historical data has not changed much, and the length of the first preset step can be increased based on the current step size corresponding to the sliding window. For example, if the current step size corresponding to the sliding window is 3 seconds, it can be increased to 5 seconds. The specific increase in step size can be adjusted according to actual conditions, and this application embodiment does not impose specific limitations on it.

[0101] When the similarity is less than 50%, it indicates that the battery's historical data has changed significantly. In this case, the length of the second preset step size can be reduced from the current step size corresponding to the sliding window. For example, if the current step size corresponding to the sliding window is 5 seconds, it can be reduced to 4 seconds. The specific increase in step size can be adjusted according to the actual situation, and this application embodiment does not impose a specific limitation on it.

[0102] It should be noted that the first preset step size can be the same as or different from the second preset step size. Furthermore, the first preset step size can be adjusted based on the specific similarity level. For example, the first preset step size can be determined based on the magnitude by which the similarity level exceeds the first threshold; the larger the magnitude, the larger the first preset step size; the smaller the magnitude, the smaller the first preset step size. The second preset step size is similar to the first preset step size and will not be elaborated further here.

[0103] In addition, if the similarity is between the first and second thresholds, the step size does not need to be adjusted and can be equal to the step size corresponding to the current sliding window.

[0104] This application embodiment calculates the similarity between training samples corresponding to two adjacent sliding windows and adjusts the step size of the sliding window based on the similarity. This allows for a slightly larger data collection interval when the battery history data changes little and a more intensive data collection when the battery history data changes significantly. This approach can collect a sufficient number of useful features while reducing the computational load of model training.

[0105] Based on the above embodiments, the first training sample includes multiple first training data, the second training sample includes multiple second training data, and calculating the similarity between the first training sample and the second training sample includes:

[0106] The distance matrix is ​​calculated based on the first training data and the second training data.

[0107] Determine the shortest path from the top-left element to the bottom-right element of the distance matrix;

[0108] Similarity is determined based on the shortest path.

[0109] In the specific implementation process, the similarity between the first training sample and the second training sample can be calculated in the following way:

[0110] 1. Fill the distance matrix:

[0111] Given two time series A (the first training data) and B (the second training data), with lengths m and n respectively.

[0112] Construct an m×n matrix grid, where each element (i,j) represents the distance between the i-th point in A and the j-th point in B. Euclidean distance is typically used as the metric.

[0113] When filling the matrix, start from the bottom left corner (0,0) and move upwards or to the right step by step, calculating and accumulating the distance of each point.

[0114] 2. Find the shortest path:

[0115] After filling the distance matrix, we need to find a path from the bottom left corner to the top right corner such that the sum of the distances of all points on the path is minimized.

[0116] This path is called a warping path, which represents the best alignment between A and B, or similarity.

[0117] The selection of a path needs to meet certain constraints, such as continuity (it can only match adjacent points) and monotonicity (points on the path move monotonically over time and cannot be reversed).

[0118] This application embodiment calculates the distance matrix between the first training data and the second training data, and calculates the shortest path from the top left element to the bottom right element of the distance matrix. A larger shortest path indicates a smaller similarity between the two training data, and a smaller shortest path indicates a larger similarity between the two training data. This calculation method can be used to calculate the similarity of data with time-series characteristics.

[0119] Based on the above embodiments, the residual network includes an input module, a first residual module, and a second residual module; training the residual network using training samples includes:

[0120] Obtain the amount of data for the training samples;

[0121] The number of the first residual module and the second residual module is determined based on the amount of data.

[0122] A residual network is generated based on the number of the first residual module and the second residual module;

[0123] The residual network is trained by using training samples as input to the model and the features of the training samples as labels.

[0124] In practical implementation, traditional residual networks, such as ResNet18, ResNet 34, ResNet 50, ResNet101, and ResNet 152, include 64-dimensional convolutional layers, 64-dimensional residual blocks, 128-dimensional residual blocks, 256-dimensional residual blocks, and 512-dimensional residual blocks. Thermal runaway scenarios differ significantly in data volume from image processing, with relatively less data in thermal runaway. Therefore, this embodiment can use 64-dimensional convolutional layers, 64-dimensional residual blocks, and 128-dimensional residual blocks. Furthermore, compared to traditional residual networks, this embodiment configures the convolutional kernels and strides within the convolutional layers and residual blocks to accommodate the temporal characteristics of thermal runaway data.

[0125] Figure 3 is a schematic diagram of the residual network model structure provided in the embodiment of this application. As shown in Figure 3, the network structure is divided into three parts: the input part (top dashed box in Figure 3), the 64-dimensional residual block (leftmost dashed box in Figure 3), and the 128-dimensional residual block (rightmost dashed box in Figure 3). The input part consists of a convolutional layer Convld-64 (3*1 convolutional kernel, stride 1), a normalization layer BachNormld, an activation layer ReLU, and a pooling layer MaxPoolld (3*1 convolutional kernel, stride 2). This input layer simplifies the input data, greatly reducing the storage space occupied. Five 64-dimensional residual blocks and eight 128-dimensional residual blocks are designed. It should be noted that the number of 64-dimensional and 128-dimensional residual blocks can be determined according to the amount of data in the training samples. The amount of data can refer to the number of data points contained in the training samples or the data length. The larger the data volume, the larger the number of first and second residual modules; conversely, the smaller the data volume, the smaller the number of first and second residual modules. Specifically, when the data volume is 900, the number of 64-dimensional residual blocks can be set to 5 by default, and the number of 128-dimensional residual blocks can be set to 8 by default. When the data length increases by 20%, the number of residual blocks dynamically increases by 5%; similarly, when the data length decreases by 20%, the number of residual blocks dynamically decreases by 5%.

[0126] In this embodiment of the application, during model training, the number of first residual modules and second residual modules in the residual network is determined according to the number of samples. When the number of samples is large, more first residual modules and second residual modules are used to fully extract features; when the number of samples is small, fewer first residual modules and second residual modules are used to reduce the resource consumption of network training.

[0127] Figure 4 is a schematic flowchart of a feature extraction method provided in an embodiment of this application. As shown in Figure 4, the method includes:

[0128] Step 401: Collect battery data for a preset time period prior to the current time;

[0129] Step 402: Input the battery data into the feature extraction model to obtain battery features; wherein, the feature extraction model is trained using the training methods described in the above embodiments.

[0130] Figure 5 is a schematic diagram of a sliding time window for collecting training samples according to an embodiment of this application. As shown in Figure 5, battery data mainly includes voltage and temperature. The window size can be 64. Assuming one data sampling point per second, a 64-second time window represents the battery voltage and temperature data within 64 seconds as a training sample, and simple statistical features (maximum and minimum values, mean, and variance, etc.) are added to this sample. As can be seen from Figure 5, the voltage and temperature changes are relatively small in the first half, and the data from the two consecutive 64-second intervals are basically the same. If a fixed step size is used, the adjacent samples obtained after sliding will be very similar. Therefore, this embodiment of the application uses a dynamic sliding step size, and the method for adjusting the step size is as follows:

[0131] 1. Set the current step size variable S = 1 (default value).

[0132] 2. Calculate the Dynamic Time Warping (DTW) distance between the current window and the time series data in the previous window. DTW(i,j)=d(i,j)+min{DTW(i-1,j),DTW(i,j-1)}

[0133] As shown in the formula above, DTW(i,j) represents the shortest path distance from the starting point to the (i,j)-th element, and d(i,j) represents the distance to the (i,j)-th element (for more detailed information on dynamic time warping, please refer to Baidu, which will not be elaborated here). DTW's distance formula measures the similarity between two sequences by calculating the shortest path between them. Specifically, DTW treats the two sequences as a two-dimensional matrix, where each element represents the distance between the data points at two different times. Then, using dynamic programming, it calculates the shortest path from the starting point to the ending point; this path represents the optimal match between the two sequences.

[0134] 3. When the DTW distance is less than the set threshold (the specific value is set according to the characteristics of the thermal runaway data), it indicates that the data within the preceding and following windows are highly similar. Therefore, the step size variable S needs to be doubled. This way, the step size of the sliding window increases with each subsequent slide, reducing the number of similar samples. Similarly, when the DTW distance is greater than the set threshold, it indicates that the data within the preceding and following windows are significantly different. In this case, the step size variable S needs to be halved to ensure that more samples are collected in areas where battery voltage and temperature changes are significant, which is more beneficial for subsequent feature extraction by the neural network.

[0135] 4. Input the extracted training samples into the feature extraction model to obtain battery features.

[0136] Figure 6 is a schematic flowchart of a thermal runaway prediction method provided in an embodiment of this application. As shown in Figure 6, the method includes:

[0137] Step 601: Obtain the battery features of the battery to be predicted;

[0138] Step 602: Input the battery features into the thermal runaway prediction model and obtain the prediction results output by the thermal runaway prediction model; the prediction results characterize whether the battery has the possibility of thermal runaway.

[0139] In the specific implementation process, the battery features are extracted using the feature extraction methods described in the above embodiments. After obtaining the battery features, they are input into the thermal runaway prediction model. The thermal runaway prediction model analyzes the battery features to obtain prediction results. The thermal runaway prediction model is pre-trained using training samples, which include training battery features and labels indicating whether thermal runaway has occurred. The training battery features are extracted using the feature extraction models provided in the above embodiments.

[0140] This application embodiment extracts training samples from battery historical data with time-series characteristics using a dynamic sliding window, solving the problem of a large number of similar training samples. Furthermore, it proposes a feature extraction model based on a residual network, reducing the possibility of gradient vanishing. Simultaneously, during the training of the feature extraction model, this application embodiment adjusts the number of residual modules according to the amount of training data. This achieves a deeper network structure when the data volume is large, fully extracting the features of thermal runaway; and reduces the number of network layers when the data volume is small, lowering the resource consumption for network training, thus improving the overall feature extraction performance for thermal runaway.

[0141] Figure 7 is a schematic diagram of a training device for a feature extraction model provided in an embodiment of this application. This device can be a module, program segment, or code on an electronic device. It should be understood that this device corresponds to the method embodiment in Figure 1 above and is capable of executing the various steps involved in the method embodiment in Figure 1. The specific functions of this device can be found in the description above; to avoid repetition, detailed descriptions are appropriately omitted here. The device includes: a historical data acquisition module 701, a sample extraction module 702, and a training module 703, wherein:

[0142] The historical data acquisition module 701 is used to acquire historical battery data before thermal runaway occurs;

[0143] The sample extraction module 702 is used to extract training samples from the battery history data based on a dynamic sliding window; wherein, the window size of the dynamic sliding window is a preset size, and the step size is determined based on the similarity between the training samples corresponding to the first two sliding windows;

[0144] The training module 703 is used to train the residual network using training samples to obtain a trained feature extraction model.

[0145] Based on the above embodiments, the sample extraction module 702 is specifically used for:

[0146] Get the first training sample extracted by the current sliding window and the second training sample extracted by the previous sliding window;

[0147] Calculate the similarity between the first training sample and the second training sample;

[0148] The step size of the next sliding window is determined based on the similarity and the step size corresponding to the current sliding window.

[0149] Based on the window size and the step size, the training sample corresponding to the next sliding window is extracted from the battery history data.

[0150] Based on the above embodiments, the sample extraction module 702 is specifically used for:

[0151] If the similarity is greater than the first threshold, then the first preset step length is added to the step length corresponding to the current sliding window;

[0152] If the similarity is less than the second threshold, then the second preset step length is reduced based on the step length corresponding to the current sliding window.

[0153] Based on the above embodiments, the first training sample includes multiple first training data, the second training sample includes multiple second training data, and the sample extraction module 702 is specifically used for:

[0154] The distance matrix is ​​calculated based on the first training data and the second training data;

[0155] Determine the shortest path from the top-left element to the bottom-right element of the distance matrix;

[0156] The similarity is determined based on the shortest path.

[0157] Based on the above embodiments, the residual network includes an input module, a first residual module, and a second residual module; the training module 703 is specifically used for:

[0158] Obtain the amount of data from the training samples;

[0159] The number of the first residual module and the second residual module is determined based on the amount of data.

[0160] The residual network is generated based on the number of the first residual module and the second residual module;

[0161] The training samples are used as input to the model, and the features of the training samples are used as labels to train the residual network.

[0162] Figure 8 is a schematic diagram of a feature extraction device provided in an embodiment of this application. This device can be a module, program segment, or code on an electronic device. It should be understood that this device corresponds to the method embodiment in Figure 4 above and is capable of executing the various steps involved in the method embodiment in Figure 4. The specific functions of this device can be found in the description above; to avoid repetition, detailed descriptions are appropriately omitted here. The device includes: a battery data acquisition module 801 and a feature extraction module 802, wherein:

[0163] The battery data acquisition module 801 is used to collect battery data for a preset time period prior to the current time.

[0164] The feature extraction module 802 is used to input battery data into the feature extraction model to obtain battery features; wherein, the feature extraction model is trained using the training methods described in the above embodiments.

[0165] Figure 9 is a schematic diagram of a thermal runaway prediction device provided in an embodiment of this application. This device can be a module, program segment, or code on an electronic device. It should be understood that this device corresponds to the method embodiment in Figure 6 above and is capable of executing the various steps involved in the method embodiment in Figure 6. The specific functions of this device can be found in the description above; to avoid repetition, detailed descriptions are appropriately omitted here. The device includes: a feature acquisition module 901 and a prediction module 902, wherein:

[0166] The feature acquisition module is used to acquire the battery features of the battery to be predicted; the battery features are extracted using the feature extraction methods described in the above embodiments.

[0167] The prediction module is used to input battery characteristics into the thermal runaway prediction model and obtain the prediction results output by the thermal runaway prediction model; the prediction results characterize whether the battery has the possibility of thermal runaway.

[0168] Figure 10 is a schematic diagram of the physical structure of an electronic device provided in an embodiment of this application. As shown in Figure 10, the electronic device includes: a processor 1001, a memory 1002, and a bus 1003; wherein,

[0169] The processor 1001 and the memory 1002 communicate with each other through the bus 1003;

[0170] The processor 1001 is used to call program instructions in the memory 1002 to execute the methods provided in the above-described method embodiments, such as: acquiring battery historical data before thermal runaway; extracting training samples from the battery historical data based on a dynamic sliding window; wherein the window size of the dynamic sliding window is a preset size, and the step size is determined based on the similarity between the training samples corresponding to the current sliding window and the previous sliding window; and training the residual network using the training samples to obtain a trained feature extraction model.

[0171] The processor 1001 can be an integrated circuit chip with signal processing capabilities. The processor 1001 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor.

[0172] The memory 1002 may include, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0173] This embodiment discloses a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by the computer, the computer can execute the methods provided in the above-described method embodiments, such as: acquiring battery historical data before thermal runaway; extracting training samples from the battery historical data based on a dynamic sliding window; wherein the window size of the dynamic sliding window is a preset size, and the step size is determined based on the similarity between the training samples corresponding to the current sliding window and the previous sliding window; and training a residual network using the training samples to obtain a trained feature extraction model.

[0174] This embodiment provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the methods provided in the above-described method embodiments. These instructions include, for example: acquiring historical battery data prior to thermal runaway; extracting training samples from the historical battery data based on a dynamic sliding window; wherein the size of the dynamic sliding window is a preset size, and the step size is determined based on the similarity between the training samples corresponding to the current sliding window and the previous sliding window; and training a residual network using the training samples to obtain a trained feature extraction model.

[0175] It should be noted that the aforementioned electronic devices, computer program products, and non-transitory computer-readable storage media can also be used to execute feature extraction methods and thermal runaway prediction methods.

[0176] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0177] Furthermore, 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.

[0178] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0179] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0180] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A training method for a feature extraction model, characterized in that, include: Acquire historical battery data prior to thermal runaway; the historical battery data consists of relevant parameters that characterize whether thermal runaway has occurred in the battery. Training samples are extracted from the battery history data based on a dynamic sliding window; wherein the window size of the dynamic sliding window is a preset size, and the step size is determined based on the similarity between the training samples corresponding to the current sliding window and the previous sliding window respectively. The residual network is trained using the training samples to obtain a trained feature extraction model; The residual network includes an input module, a first residual module, and a second residual module; training the residual network using the training samples includes: Obtain the amount of data from the training samples; The number of the first residual module and the second residual module is determined based on the amount of data. The residual network is generated based on the number of the first residual module and the second residual module; The training samples are used as input to the model, and the features of the training samples are used as labels to train the residual network.

2. The method according to claim 1, characterized in that, The step of extracting training samples from the battery history data based on a dynamic sliding window includes: Get the first training sample extracted by the current sliding window and the second training sample extracted by the previous sliding window; Calculate the similarity between the first training sample and the second training sample; The step size of the next sliding window is determined based on the similarity and the step size corresponding to the current sliding window. Based on the window size and the step size, the training sample corresponding to the next sliding window is extracted from the battery history data.

3. The method according to claim 2, characterized in that, Determining the step size of the current sliding window based on the similarity and the step size corresponding to the current sliding window includes: If the similarity is greater than the first threshold, then the first preset step length is added to the step length corresponding to the current sliding window; If the similarity is less than the second threshold, then the second preset step length is reduced based on the step length corresponding to the current sliding window.

4. The method according to claim 2, characterized in that, The first training sample includes multiple first training data sets, and the second training sample includes multiple second training data sets. Calculating the similarity between the first training sample and the second training sample includes: The distance matrix is ​​calculated based on the first training data and the second training data; Determine the shortest path from the top-left element to the bottom-right element of the distance matrix; The similarity is determined based on the shortest path.

5. A feature extraction method, characterized in that, include: Collect battery data from batteries within a preset time period prior to the current time; The battery data is input into a feature extraction model to obtain battery features; wherein the feature extraction model is trained using the training method described in any one of claims 1-4.

6. A method for predicting thermal runaway, characterized in that, include: Obtain the battery features of the battery to be predicted; the battery features are obtained by extracting them using the feature extraction method as described in claim 5. The battery characteristics are input into the thermal runaway prediction model to obtain the prediction results output by the thermal runaway prediction model; the prediction results characterize whether the battery has the possibility of thermal runaway.

7. A training device for a feature extraction model, characterized in that, include: The historical data acquisition module is used to acquire battery historical data before thermal runaway occurs; the battery historical data are relevant parameters that can characterize whether the battery has experienced thermal runaway. The sample extraction module is used to extract training samples from the battery history data based on a dynamic sliding window; wherein the window size of the dynamic sliding window is a preset size, and the step size is determined based on the similarity between the training samples corresponding to the first two sliding windows. The training module is used to train the residual network using the training samples to obtain a trained feature extraction model; the residual network includes an input module, a first residual module, and a second residual module; the training module is specifically used for: Obtain the amount of data from the training samples; The number of the first residual module and the second residual module is determined based on the amount of data. The residual network is generated based on the number of the first residual module and the second residual module; The training samples are used as input to the model, and the features of the training samples are used as labels to train the residual network.

8. A feature extraction device, characterized in that, include: The battery data acquisition module is used to collect battery data for a preset time period up to the current time. The feature extraction module is used to input the battery data into the feature extraction model to obtain battery features; wherein the feature extraction model is trained using the training method described in any one of claims 1-4.

9. A thermal runaway prediction device, characterized in that, include: The feature acquisition module is used to acquire the battery features of the battery to be predicted; the battery features are obtained by extracting the features using the feature extraction method as described in claim 5. The prediction module is used to input the battery characteristics into the thermal runaway prediction model and obtain the prediction results output by the thermal runaway prediction model; the prediction results characterize whether the battery has the possibility of thermal runaway.

10. An electronic device, characterized in that, include: Processor, memory, and bus, among which, The processor and the memory communicate with each other via the bus; The memory stores program instructions that can be executed by the processor, and the processor can execute the method as described in any one of claims 1-6 by calling the program instructions.

11. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions, which, when executed by a computer, cause the computer to perform the method as described in any one of claims 1-6.

12. A computer program product, characterized in that, It includes computer program instructions, which, when read and executed by a processor, perform the method as described in any one of claims 1-6.

Images