Battery failure prediction method, device, and storage medium

Abstract

The application discloses a battery fault prediction method, device and storage medium, the method comprises the following steps: obtaining time series data about a plurality of battery parameters, wherein the time series data is collected during the use of the battery, and the time series data comprises first parameter values of the plurality of battery parameters at different collection time points; mapping the time series data to a charge domain to obtain a charge domain sequence, wherein the charge domain sequence comprises second parameter values corresponding to different charge amounts of the plurality of battery parameters; extracting features from the charge domain sequence to obtain fault-related features of the battery; and performing fault prediction by using the fault-related features of the battery to obtain a fault prediction result of the battery. In the foregoing manner, the application can accurately predict the battery fault.

Description

Technical Field

[0001] This application relates to the field of fault detection technology, and in particular to a battery fault prediction method, device and storage medium. Background Technology

[0002] Batteries are widely used in energy storage, new energy vehicles, and other fields. Battery failures are often insidious and can rapidly escalate into thermal runaway, posing significant safety risks. Furthermore, in practical applications, battery charging and discharging conditions are complex and variable, and battery operating data is fragmented and lacks a fixed cyclical pattern, severely interfering with fault feature extraction. Existing prediction methods have poor accuracy, with significant issues of misjudgment and missed detection. Therefore, achieving accurate prediction of battery failures is of great significance. Summary of the Invention

[0003] The main technical problem addressed by this application is to provide a battery fault prediction method, device, and storage medium that can accurately predict battery faults.

[0004] To address the aforementioned technical problems, this application provides a battery fault prediction method. This method includes: acquiring time-series data on multiple battery parameters, wherein the time-series data is collected during battery use and includes first parameter values ​​of the multiple battery parameters at different acquisition times; mapping the time-series data to a charge domain to obtain a charge domain sequence, wherein the charge domain sequence includes second parameter values ​​of the multiple battery parameters corresponding to different charge amounts; extracting features from the charge domain sequence to obtain fault-related features of the battery; and using the fault-related features of the battery to perform fault prediction to obtain a battery fault prediction result.

[0005] The process involves several battery parameters, including current. The time-series data is mapped to the charge domain to obtain a charge domain sequence, which includes: obtaining the current value at each acquisition time from the time-series data; integrating the current value at each acquisition time to obtain the charge amount of the battery at each acquisition time; and generating multiple sets of parameters corresponding to different charge amounts using the charge amount at each acquisition time and the first parameter value of the battery parameters at each acquisition time. Each parameter set includes a second parameter value mapping each battery parameter to the charge amount.

[0006] Specifically, by utilizing the charge amount at each acquisition time and the first parameter value of the battery parameters at each acquisition time, multiple sets of parameter groups corresponding to different charge amounts are generated. This includes: using the charge amount at each acquisition time and the first parameter value of the battery parameters at each acquisition time to form multiple sets of mapping groups, each mapping group including the charge amount at the same acquisition time and the first parameter value of each battery parameter; and using the mapping relationship between the charge amount represented by the multiple mapping groups and each battery parameter to generate multiple sets of parameter groups corresponding to different charge amounts.

[0007] The charge domain sequence includes multiple sets of parameters corresponding to each charge node during battery use. Each charge node takes values ​​at preset charge step intervals, and the parameter set includes the second parameter value of each battery parameter corresponding to the charge node.

[0008] Before extracting features from the charge domain sequence to obtain the battery's fault-related features, the method further includes: using the charge domain sequence to generate at least one electrochemical feature sequence of the battery in the charge domain, each electrochemical feature sequence including a second parameter value corresponding to different charge amounts of an electrochemical feature parameter; and adding at least one electrochemical feature sequence to the charge domain sequence.

[0009] Among them, multiple battery parameters include voltage; using the charge domain sequence, at least one electrochemical feature sequence of the battery in the charge domain is generated, including: using the voltage values ​​corresponding to different charge amounts in the charge domain sequence to obtain the incremental capacity sequence and differential voltage sequence in the charge domain; and using the incremental capacity sequence and differential voltage sequence as electrochemical feature sequences.

[0010] The process includes: extracting features from the charge domain sequence to obtain battery fault-related features, including: identifying charge domain pseudo-periods in the charge domain sequence; performing first feature extraction on the first charge domain sequence based on the charge domain pseudo-periods to obtain time-series fault-related features, wherein the time-series fault-related features include at least one of local features within a period and global features across periods; the first charge domain sequence includes second parameter values ​​corresponding to different charge amounts for the first battery parameters; and the first battery parameters are at least one of the battery parameters included in the charge domain sequence; and / or, performing second feature extraction on the second charge domain sequence based on the physical coupling relationship of each cell within the battery to obtain spatial fault-related features, wherein the second charge domain sequence includes second parameter values ​​corresponding to different charge amounts for the second battery parameters; and the second battery parameters are at least one of the battery parameters included in the charge domain sequence.

[0011] The process of identifying the charge domain pseudo-period of the charge domain sequence includes: performing a fast Fourier transform on the charge domain sequence to obtain the charge domain spectrum; and using a preset number of frequency points with the largest amplitude in the charge domain spectrum to calculate the charge domain pseudo-period.

[0012] The time-series fault-related features include intra-period local features and cross-period global features. Based on the charge domain pseudo-period, the first charge domain sequence is subjected to first feature extraction to obtain the time-series fault-related features. This includes: dividing the first charge domain sequence into several periodic segments according to the charge domain pseudo-period to obtain a segment sequence; extracting intra-period local features from the segment sequence using the first convolutional submodule of the time-series feature extraction module, and extracting cross-period global features from the segment sequence using the second convolutional submodule of the time-series feature extraction module, wherein the convolutional kernel of the first convolutional submodule is smaller than the convolutional kernel of the second convolutional submodule; and fusing the intra-period local features and the cross-period global features to obtain the time-series fault-related features.

[0013] The process involves extracting intra-period local features from the segment sequence using the first convolutional submodule of the temporal feature extraction module, and extracting cross-period global features from the segment sequence using the second convolutional submodule of the temporal feature extraction module. This includes: treating each period segment as the current segment; extracting features from the current segment using the first convolutional submodule to obtain intra-period local features; extracting features from the current segment using the second convolutional submodule based on the current segment and a reference segment to obtain cross-period global features corresponding to the current segment, where the reference segment is at least one period segment other than the current segment; and fusing the intra-period local features and cross-period global features to obtain temporal fault-related features, including: fusing the intra-period local features and cross-period global features of the current segment to obtain the temporal fault-related features corresponding to the current segment.

[0014] The segment sequence is a two-dimensional tensor, which is obtained by folding several periodic segments.

[0015] Among them, the local features within the cycle are the fluctuation features within the cycle, and the global features across the cycle are the global degradation features across the cycle.

[0016] The first battery parameters include voltage, temperature, incremental capacity, and differential voltage.

[0017] Specifically, based on the physical coupling relationship of each cell within the battery, a second feature is extracted from the second charge domain sequence to obtain spatial fault-related features. This includes: taking the data corresponding to the same moment in the second charge domain sequence as a data frame, obtaining the second battery parameter values ​​of each cell in the data frame, and using them as the initial frame features of the cells; taking each cell as a target cell, determining the influence weight of adjacent nodes on the target cell based on the initial frame features of the target cell and the initial frame features of the target cell's adjacent cells, and using the initial frame features of each adjacent cell of the target cell and the corresponding influence weights to obtain the updated frame features of the target cell; and fusing the updated frame features of each cell to obtain the spatial fault-related features of the battery corresponding to the same moment.

[0018] The process involves obtaining the second battery parameter values ​​of each cell in the data frame, which are used as the initial frame features of the cells. This includes: constructing a physical topology graph of the battery based on the physical coupling relationship of each cell within the battery, where each node in the physical topology graph represents a cell in the battery, and the edges between nodes represent the physical adjacency relationship between the corresponding cells; using the second battery parameter values ​​of each cell in the data frame as the initial features of the corresponding nodes; and taking each cell as a target cell, determining the influence weights of adjacent nodes on the target cell based on the initial frame features of the target cell and the initial frame features of its adjacent cells, and obtaining the updated frame features of the target cell using the initial frame features and corresponding influence weights of its adjacent cells. This includes: performing self-attention processing on the initial frame features of each node in the physical topology graph using a graph attention network to obtain the influence weights between adjacent nodes in the physical topology graph, where adjacent nodes are nodes connected by edges; and for each node, weighted aggregation of the initial frame features of its adjacent nodes using the influence weights of the nodes to obtain the updated frame features of the node.

[0019] The second battery parameters include voltage, temperature, internal resistance, and state of charge.

[0020] The steps of feature extraction from the charge domain sequence and fault prediction using battery fault-related features are executed by the fault prediction model. The fault prediction model sequentially performs the following processing on each charge domain pseudo-period of the charge domain sequence: extracting fault-related features corresponding to the charge domain pseudo-period, and performing fault prediction based on the fault-related features of the charge domain pseudo-period to obtain the fault prediction result corresponding to the charge domain pseudo-period. The fault-related features corresponding to the charge domain pseudo-period include temporal fault-related features and spatial fault-related features. The temporal fault-related features corresponding to the charge domain pseudo-period include intra-period local features and cross-period global features. The intra-period local features corresponding to the charge domain pseudo-period are extracted from the current segment corresponding to the charge domain pseudo-period. The cross-period global features corresponding to the charge domain pseudo-period are extracted based on the current segment corresponding to the charge domain pseudo-period and reference segments corresponding to other charge domain periods. The spatial fault-related features corresponding to the charge domain pseudo-period are obtained by fusing the spatial fault-related features corresponding to each time step in the charge domain pseudo-period.

[0021] Among them, the fault-related features include time-series fault-related features and spatial fault-related features; using the fault-related features of the battery to perform fault prediction, the battery fault prediction results are obtained, including: fusing time-series fault-related features and spatial fault-related features to obtain fused features; and making predictions based on the fused features to obtain fault prediction results.

[0022] The steps of feature extraction from the charge domain sequence and fault prediction using battery fault-related features are performed by the fault prediction model. The method also includes: acquiring the sample charge domain sequence of a sample battery; processing the sample charge domain sequence using the fault prediction model to obtain sample fault prediction results; determining the prediction loss of the fault prediction model based on the difference between the sample fault prediction results and the fault labeling results of the sample charge domain sequence; determining the actual electrochemical degradation state of the battery based on the sample fault prediction results, and determining the physical penalty loss of the battery electrochemical degradation mechanism based on the difference between the actual electrochemical degradation state and the theoretical electrochemical degradation state of the battery; combining the prediction loss and the physical penalty loss to obtain the total loss; and adjusting the parameters of the fault prediction model based on the total loss.

[0023] The fault prediction model sequentially outputs the sample fault prediction results corresponding to each sample period segment in the sample charge domain sequence. Each sample period segment corresponds to a sample charge domain pseudo-period in the sample charge domain sequence. The sample fault prediction results include the sample fault warning probability of the sample battery. Based on the sample fault prediction results, the true electrochemical degradation state of the battery is determined, including: obtaining the true electrochemical degradation state based on the rate of change of the current sample fault warning probability relative to the historical sample fault warning probability. Here, the current sample fault warning probability is the sample fault warning probability output for the current sample period segment, and the historical sample fault warning probability is the sample fault warning probability output for the previous sample period segments.

[0024] Among them, the theoretical electrochemical degradation state of the battery is the Arrhenius reaction rate.

[0025] The total loss is obtained by combining the predicted loss and the physical penalty loss. The total loss is the sum of the predicted loss and the physical weighted result. The physical weighted result is the product of the physical penalty loss and the physical penalty weight. The magnitude of the physical penalty weight is related to the environment in which the battery is located.

[0026] The fault prediction results include the battery fault type and fault warning probability. After obtaining the battery fault prediction results by using the battery's fault-related features, the system further includes: triggering a potential anomaly warning in response to the fault warning probability exceeding a first threshold; triggering a high-risk fault warning in response to the fault warning probability exceeding a second threshold, and generating fault prompt information, wherein the second threshold is greater than the first threshold, and the fault prompt information includes at least one of fault location, fault type, and fault cause analysis.

[0027] The battery in question is the vehicle battery, and the time-series data is extracted from the vehicle's operating data.

[0028] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a fault prediction device, including a memory and a processor coupled to each other, wherein the memory stores program instructions; and the processor is used to execute the program instructions stored in the memory to implement the above method.

[0029] To solve the above-mentioned technical problems, another technical solution adopted in this application is to provide a computer-readable storage medium for storing program instructions that can be executed to implement the above-mentioned method.

[0030] The above scheme, which maps time-series data to the charge domain, can unify the fragmented battery usage process with varying durations using charge as the measurement benchmark. This eliminates data scale differences caused by different battery usage conditions, such as fast / slow charging and differences in initial SOC. It enables horizontal comparison of battery data under different usage conditions, making it easier to extract effective fault features and improve the accuracy of battery fault prediction. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating an embodiment of the battery fault prediction method provided in this application; Figure 2 This is a flowchart illustrating an embodiment of extracting timing fault-related features using charge domain pseudo-periods provided in this application; Figure 3 This is a flowchart illustrating an embodiment of extracting spatial fault-related features based on the physical coupling relationship of each battery cell, as provided in this application. Figure 4 This is a flowchart illustrating an embodiment of the fault prediction model training method provided in this application; Figure 5 This is a schematic diagram of the framework of an embodiment of the fault prediction device provided in this application; Figure 6 This is a schematic diagram of the framework of the computer-readable storage medium provided in this application. Detailed Implementation

[0032] To make the purpose, technical solution and effects of this application clearer and more explicit, the following describes this application in further detail with reference to the accompanying drawings and embodiments.

[0033] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0034] It should be noted that the inventors found in their research that traditional battery fault prediction methods mostly use fault diagnosis models based on absolute time series for fault identification and prediction. However, battery charging and discharging conditions are complex and variable, and battery operating data is fragmented and has no fixed cyclical pattern. For example, the length of charging and discharging data segments changes dynamically, the initial SOC (state of charge) is randomly distributed, and there is no fixed period for charging and discharging. These operating condition characteristics will seriously destroy the feature validity of absolute time series battery data, causing traditional models to be unable to accurately extract effective fault features, ultimately resulting in low fault diagnosis and prediction accuracy and a high rate of false positives and false negatives.

[0035] To address this issue, this application proposes a novel battery fault prediction method. This method maps time-series data to the charge domain and uses the battery parameter sequences (such as voltage sequences and temperature sequences) mapped to the charge domain to extract fault features and predict faults.

[0036] Understandably, there is no charge flow during the battery's resting period, and the corresponding charge domain coordinates do not increase. Therefore, mapping to the charge domain can filter out invalid blank periods. At the same time, this mapping method is not affected by the difference in the battery's initial SOC or the different charging currents corresponding to fast and slow charging. As long as the cumulative charge and discharge capacity is the same, it will be mapped to the same coordinate point in the charge domain. This can eliminate the data length scale deviation caused by different charging rates, achieve standardized alignment of multi-operating condition data, and thus help extract effective fault features and improve the accuracy of battery fault prediction.

[0037] To facilitate understanding of this application, the battery failure prediction device and storage medium provided in this application are described below with reference to specific embodiments: Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the battery fault prediction method provided in this application. It should be noted that if substantially the same result is obtained, this embodiment does not necessarily reflect that outcome. Figure 1 The illustrated process sequence is limited. For example... Figure 1 As shown, this embodiment includes: S11: Obtain time-series data on multiple battery parameters. The time-series data is collected during battery use and includes the first parameter values ​​of multiple battery parameters at different collection times.

[0038] This embodiment is used to convert time-based time series data into charge domain sequences based on charge quantity, and to use the charge domain sequences for feature extraction and fault prediction.

[0039] The batteries described in this article can be, but are not limited to, vehicle batteries, as well as batteries in mobile phones, tablets, laptops, wearable devices, or related energy storage batteries. For ease of description, the following description will use vehicle batteries as an example.

[0040] Time-series data of multiple battery parameters, such as state change data of each battery parameter during actual battery use, obtained through sensors or battery management systems.

[0041] The first parameter value is the value of the battery parameter at a specific acquisition time. For example, multiple battery parameters include battery parameters such as current, voltage, and temperature. The corresponding time series data includes the current value, voltage value, and temperature value acquired at each acquisition time. That is, the time series data of multiple battery parameters includes current time series data, voltage time series data, and temperature time series data.

[0042] In a specific application scenario, the battery to be predicted for fault is a vehicle battery, and the time-series data of multiple battery parameters is extracted from the vehicle's operating data. This vehicle operating data includes, for example, real-time operating messages continuously uploaded by the vehicle or historical operating messages. After obtaining the operating messages, the parameter values ​​of each battery parameter during battery use are extracted to obtain multi-dimensional battery data.

[0043] In one implementation scenario, the acquired operation message fully records the operation data corresponding to at least one charge-discharge cycle of the battery. Therefore, the operation data from one or more charge-discharge cycles can be used to predict battery faults.

[0044] In one embodiment, to improve the accuracy of battery fault prediction, after acquiring time-series data on multiple battery parameters, the acquired time-series data can be preprocessed, and the preprocessed time-series data can be used for subsequent charge domain mapping. Data preprocessing may include cleaning and calibrating outlier data in the time-series data; for example, a local outlier factor can be used to remove outlier spurious values, cubic spline interpolation can be used to complete missing frames, and filters (such as Savitzky-Golay filters) can be used to smooth the time-series data (voltage sequences, temperature sequences).

[0045] S12: Map the time series data to the charge domain to obtain the charge domain sequence, wherein the charge domain sequence includes the second parameter values ​​of multiple battery parameters corresponding to different charge amounts.

[0046] In this embodiment, mapping time-series data to the charge domain to obtain the charge domain sequence can be understood as follows: using the cumulative charge of the battery at each moment (corresponding to the charge amount in this article) as the horizontal axis (or anchor point), and using the second parameter value of each battery parameter at the corresponding charge amount as the vertical axis (or the associated data of the anchor point), the time domain is converted to the charge domain to obtain the charge domain sequence.

[0047] In one embodiment, each charge node takes values ​​at preset charge step intervals. The charge domain sequence includes multiple sets of parameters corresponding to each charge node during battery use. Each parameter set includes the second parameter value of the charge node corresponding to each battery parameter. For example, if the battery parameters include voltage and temperature, then the parameter set corresponding to the charge node includes the voltage and temperature values ​​corresponding to that charge node.

[0048] The preset charge step size can be set to a fixed value (e.g., 1 Ah) or dynamically adjusted according to changes in charge; this is not limited here. A charge node can be understood as a discrete point on the horizontal axis corresponding to the charge domain. For example, the battery's full state of charge range can be divided into several equally spaced intervals, with each interval's boundary point corresponding to a charge node. The parameter set of a charge node refers to the set of all monitored battery parameters (such as voltage, temperature, internal resistance, etc.) and their values ​​at that charge node.

[0049] In one embodiment, multiple battery parameters include current, and the corresponding acquired time-series data includes current values ​​at each acquisition time. In this embodiment, the time-series data can be mapped to the charge domain using the current values ​​at each acquisition time to obtain a charge domain sequence. Specifically, the steps include: Step 1: Obtain the current value at each acquisition time from the time series data.

[0050] Step 2: Integrate the current values ​​at each sampling time to obtain the charge of the battery at each sampling time.

[0051] In one embodiment, the real-time integration method shown in the following formula can be used to integrate the current values ​​at each sampling time to obtain the charge of the battery at each sampling time:

[0052] in, express The current value at any given moment; the charging current and discharging current have opposite signs; t represents the current cumulative duration. The charge at time t represents the cumulative charge flowing through the battery from time 0 to time t.

[0053] Step 3: Using the charge amount at each acquisition time and the first parameter value of the battery parameters at each acquisition time, generate multiple sets of parameters corresponding to different charge amounts. The parameter sets include the second parameter value of each battery parameter mapped to the charge amount.

[0054] First, multiple mapping groups can be formed using the charge amount at each acquisition time (corresponding to the charge amount nodes mentioned above) and the first parameter values ​​of the battery parameters at each acquisition time. Then, using the mapping relationship between the charge amount represented by the multiple mapping groups and each battery parameter, multiple parameter groups corresponding to different charge amounts can be generated. Each mapping group includes the charge amount at the same acquisition time and the first parameter value of each battery parameter.

[0055] Each mapping group consists of the charge quantity at the same acquisition time and the first parameter value of each battery parameter. For example, if the charge quantity at time t is Qt, the voltage is Vt, and the temperature is Tt, then a mapping group (Qt, Vt, Tt) is formed. This embodiment essentially packages and combines the acquired data of the charge quantity and each battery parameter at the same time to form mapping groups, thereby establishing the correspondence between the charge quantity and each battery parameter at each time. Then, using the mapping relationship between the charge quantity and each battery parameter represented by multiple mapping groups, multiple parameter groups corresponding to different charge quantities are generated.

[0056] In one embodiment, each mapping group can be directly used as a set of parameters with different charge amounts.

[0057] In another implementation, considering the differences in the range of charge quantity across multiple charge-discharge cycles of the same battery, which makes it impossible to align the charge quantity ranges of different cycles, and the potential uneven distribution of the original sampling points for multiple battery parameters, to achieve uniformity in the charge dimension value range and standardized sampling intervals across all charge-discharge cycles, interpolation and data completion operations can be performed based on the mapping relationship between charge quantity and each battery parameter. This will normalize and map the parameter sequences of different local intervals to the same complete charge interval. For example, when only the 30%-100% SOC interval is collected during a single charge, the mapping relationship (charge quantity and parameter mapping relationship) established by the battery's other complete charge-discharge segments or its existing charge-discharge segments can be combined to complete the charge domain parameters corresponding to the 0%-30% SOC interval through interpolation and curve fitting, resulting in a unified sequence covering the standardized charge interval of 0%-100%. During a single charge or discharge cycle, each SOC value uniquely corresponds to a charge quantity.

[0058] The above mapping of time series data to the charge domain results in a charge domain sequence that can be understood as follows: multi-dimensional time series data based on absolute time sampling is reconstructed into a multi-dimensional data sequence based on equal charge step sampling through time-aware interpolation and resampling algorithms. This charge domain sequence uses the charge amount (the cumulative charge amount of the battery at each time step) as the horizontal axis.

[0059] S13: Extract features from the charge domain sequence to obtain fault-related features of the battery.

[0060] In this embodiment, the charge domain sequence can be used to extract temporal features to obtain temporal fault-related features, and / or the charge domain sequence can be used to extract spatial features to obtain spatial fault-related features.

[0061] In one implementation scenario, time-series fault-related features can be understood as the changes in battery parameters of a single cell or battery during the charging and discharging process, used to characterize the time-series gradual anomalies of the cell or battery; spatial fault-related features can be understood as the linkage differences between different cells at the same time, used to characterize spatial faults such as abnormal cell consistency and local thermal diffusion.

[0062] In some embodiments, extracting time-series fault-related features using charge domain sequences includes: identifying charge domain pseudo-periods of the charge domain sequence; performing a first feature extraction (time-series feature extraction) on the first charge domain sequence based on the charge domain pseudo-periods to obtain time-series fault-related features. The time-series fault-related features include at least one of intra-period local features and cross-period global features.

[0063] The charge domain pseudo-period refers to the repetitive or periodic fluctuation pattern of battery parameters (voltage, temperature, etc.) as the amount of charge changes under charging and discharging conditions. The charge domain pseudo-period is the charge domain period with the amount of charge as the horizontal axis. Because battery data contains noise and operating condition disturbances, it does not have a strict mathematical period, so it is defined as a pseudo-period.

[0064] For example, a charge domain pseudo-period corresponds to a charging segment in a charge-discharge cycle. The first charge node of this pseudo-period corresponds to the initial charge of 0mAh (corresponding to 0% SOC), and the last charge node corresponds to the full charge at the end of the charge cycle (corresponding to 100% SOC). Alternatively, a charge domain pseudo-period can also correspond to a charging segment and a discharging segment in a charge-discharge cycle. The first charge node of this pseudo-period corresponds to the initial charge of 0mAh (corresponding to 0% SOC), the middle charge node corresponds to the full charge at the end of the charge cycle (corresponding to 100% SOC), and the last charge node corresponds to the full charge at the end of the charge cycle (corresponding to 0% SOC).

[0065] The first charge domain sequence includes second parameter values ​​corresponding to different amounts of charge for the first battery parameters. The first battery parameter is at least one of the battery parameters included in the charge domain sequence. That is, the first battery parameter is one or more battery parameters selected from the battery parameters included in the charge domain sequence. For example, the battery parameters included in the charge domain sequence include current, voltage, temperature, incremental capacity, differential voltage, internal resistance, and state of charge. The first battery parameters include voltage, temperature, incremental capacity, and differential voltage. The corresponding first charge domain sequence includes voltage sequence, temperature sequence, incremental capacity sequence, and differential voltage sequence constructed with charge amount as the horizontal axis, respectively.

[0066] Among them, the local features within the period are the fluctuation features within the period. The fluctuation features within the period can be understood as the short-term changes in the first battery parameters (voltage, temperature, etc.) as the amount of charge changes within a charge domain pseudo-cycle (such as a single charge-discharge cycle), such as instantaneous fluctuations and abnormal offsets. It can reflect instantaneous abnormal operating conditions and local minor disturbance faults within a single charge domain pseudo-cycle.

[0067] Cross-cycle global characteristics are cross-cycle global degradation characteristics. Cross-cycle global degradation characteristics are long-term aging / degradation characteristics such as capacity decay and increased internal resistance that accumulate gradually with the number of cycles, spanning multiple charge domain pseudo-cycles (such as spanning multiple charge-discharge cycles). They are used to characterize the overall slow degradation trend of the battery.

[0068] It is understandable that local features within a cycle are used to characterize short-term battery anomalies, while global features across cycles are used to characterize long-term aging faults in the battery. The specific features to be extracted can be determined by combining the condition of the battery to be predicted and the detection requirements.

[0069] In a specific implementation scenario, to accurately predict faults by comprehensively considering both short-term battery anomalies and long-term aging fault characteristics, the extracted time-series fault-related features include intra-cycle local features and cross-cycle global features. In this embodiment, during the extraction of intra-cycle local features and cross-cycle global features, the first charge domain sequence is first segmented based on the charge domain pseudo-period, resulting in multiple independent periodic segments and forming a segment sequence. Then, a time-series feature extraction module performs dual-scale feature extraction to obtain intra-cycle local features and cross-cycle global features. Specifically, the time-series feature extraction module includes a first convolutional submodule and a second convolutional submodule. The first convolutional submodule, with a smaller kernel size, extracts intra-cycle local features within a single periodic segment from the segment sequence. The second convolutional submodule, with a larger kernel size and a receptive field covering multiple periodic segments, extracts cross-cycle global features from the segment sequence. Finally, the two types of features are fused to output complete time-series fault-related features. Further implementation details are provided below. Figure 2 The relevant description of the illustrated embodiment.

[0070] Among them, the method of determining the pseudo-period of the charge domain makes the originally irregular cycle in the time domain present a periodic consistency, which significantly enhances the ability of the time-series feature extraction module to perceive long-term battery degradation and gradual anomalies.

[0071] In a specific implementation scenario, the temporal feature extraction module is the feature extraction module in the temporal network model TimesNet, the charge domain pseudo-period is extracted using the period extraction module in the temporal network model TimesNet, and the steps of using the extracted charge domain pseudo-period to segment periodic segments and construct segment sequences are also performed by the temporal network model TimesNet.

[0072] In this embodiment, the charge domain sequence used for feature extraction can be, but is not limited to, the charge domain sequence obtained by the above mapping, or it can be the charge domain sequence obtained by adding at least one electrochemical feature sequence of the battery in the charge domain based on the charge domain sequence obtained by the above mapping.

[0073] In one embodiment, each electrochemical feature sequence is generated using the charge domain sequence obtained by the above mapping, and each electrochemical feature sequence includes a second parameter value corresponding to different charge amounts. Simply put, the charge domain sequence obtained by the above mapping is first used to generate at least one electrochemical feature sequence of the battery in the charge domain. Then, the at least one electrochemical feature sequence is added to the charge domain sequence to obtain a new charge domain sequence. Subsequently, relevant features are extracted from the new charge domain sequence.

[0074] Electrochemical characteristic parameters can be understood as newly added battery parameters in the original charge domain sequence obtained by mapping, used to reflect the electrochemical reaction state inside the cell or battery. Electrochemical characteristic parameters can be, but are not limited to, incremental capacity, differential voltage, internal resistance, polarization-related electrochemical characteristics, or temperature rise rate, etc.

[0075] In one implementation scenario, electrochemical characteristic parameters include incremental capacity and differential voltage, corresponding to an incremental capacity sequence and a differential voltage sequence. Specifically, the incremental capacity sequence and differential voltage sequence in the charge domain can be obtained first by utilizing the voltage values ​​corresponding to different charge amounts in the charge domain sequence; then, these sequences are used as the electrochemical characteristic sequences. Here, incremental capacity is the differential value dQ / dv of the charge amount and the corresponding voltage value, and differential voltage is the differential value dv / dQ of the voltage value and the corresponding charge amount. Furthermore, the peak position shift and peak height decay characteristics of the incremental capacity sequence and differential voltage sequence directly reflect the internal characteristics of the electrochemical reaction stages within the battery.

[0076] While the curves corresponding to the incremental capacity sequence and the differential voltage sequence can intuitively reflect the quality of the battery's internal reaction, the original data is noisy and cluttered. Therefore, in another implementation scenario, the incremental capacity sequence and the differential voltage sequence can be filtered first to obtain the filtered sequence (curve), and then key values ​​representing aging, such as peak value and area, can be extracted to further improve the accuracy of battery fault prediction.

[0077] It is understandable that the newly added incremental capacity sequence and differential voltage sequence are determined using charge and voltage. Therefore, the method of extracting time-series fault-related features using the charge domain sequence containing original battery parameters and electrochemical characteristic parameters can retain the variation characteristics of each original battery parameter (voltage, temperature) and enhance the coupling relationship between the linkage changes of multiple battery parameters. This method is more in line with the actual operating conditions of the battery and can extract more effective features that fit the actual operating conditions.

[0078] In some embodiments, extracting space fault-related features using a charge domain sequence includes: performing a second feature extraction on a second charge domain sequence based on the physical coupling relationship of each cell within the battery to obtain space fault-related features, wherein the second charge domain sequence includes second parameter values ​​corresponding to different charge amounts of the second battery parameters, and the second battery parameters are at least one of the battery parameters included in the charge domain sequence.

[0079] For example, the battery parameters included in the charge domain sequence include current, voltage, temperature, incremental capacity, differential voltage, internal resistance, and state of charge. The second battery parameters include voltage, temperature, internal resistance, and state of charge. The corresponding second charge domain sequence includes voltage sequence, temperature sequence, internal resistance sequence, and state of charge sequence constructed with charge quantity as the horizontal axis, respectively.

[0080] Within a battery, there are inherent physical connections and couplings (such as thermal and electrical coupling) between the individual cells. Battery faults exhibit significant spatial propagation characteristics; localized cell anomalies can easily spread to surrounding cells through thermal and electrical coupling links. Therefore, it is necessary to rely on the inherent physical coupling relationships between cells to extract fault features related to spatial propagation characteristics from the second charge domain sequence. A specific scheme for extracting spatial fault-related features based on the physical coupling relationships between the cells within the battery can be found below. Figure 3 The relevant description of the illustrated embodiment.

[0081] S14: Utilize the fault-related characteristics of the battery to perform fault prediction and obtain the battery fault prediction results.

[0082] In this embodiment, the step of using battery fault-related features to predict faults and obtain battery fault prediction results is performed by a fault prediction model. The fault-related features include at least one of temporal fault-related features and spatial fault-related features.

[0083] In one embodiment, the fault-related features include temporal fault-related features and spatial fault-related features. When performing fault prediction, the temporal fault-related features and spatial fault-related features can be fused first to obtain fused features; then, prediction is performed based on the fused features to obtain the fault prediction result.

[0084] For example, temporal fault-related features and spatial fault-related features can be fused through a gating mechanism, and the weights of temporal fault-related features and spatial fault-related features can be adaptively allocated to perform feature fusion, obtain fused features, and then use the fused features to feed into the fully connected layer of the fault prediction model to output the final fault prediction result.

[0085] In some embodiments, the fault prediction result includes fault type and fault warning probability. The fault type is, for example, thermal runaway, internal short circuit, or inconsistency anomaly; the fault warning probability is the likelihood of a fault occurring in the battery during the current charge domain pseudo-cycle; for example, predicting the fault probability of the current charge domain pseudo-cycle. Of course, in other implementation scenarios, the fault probability of the Nth future charge domain pseudo-cycle can also be predicted using the current charge domain sequence.

[0086] In another embodiment, the fault prediction result includes, in addition to the fault type and fault warning probability mentioned above, at least one of the fault location, state of health (SOH) of the battery, and remaining life (RUL).

[0087] The above scheme, which maps time-series data to the charge domain, can unify the fragmented battery usage process with varying durations using charge as the measurement benchmark. This eliminates data scale differences caused by different battery usage conditions, such as fast / slow charging and differences in initial SOC. It enables horizontal comparison of battery data under different usage conditions, making it easier to extract effective fault features and improve the accuracy of battery fault prediction.

[0088] In one embodiment, after obtaining the battery fault prediction result by using the fault-related features of the battery, the method further includes: if the fault warning probability exceeds a first threshold, triggering a potential anomaly warning; if the fault warning probability exceeds a second threshold, triggering a high-risk fault warning and generating fault prompt information, wherein the second threshold is greater than the first threshold, and the fault prompt information includes at least one of fault location, fault type, and fault cause analysis.

[0089] In simple terms, this embodiment sets tiered warning thresholds. When the probability of a fault warning exceeds a first threshold (low threshold), a potential anomaly warning is triggered to indicate a potential battery malfunction. When the probability of a fault warning exceeds a second threshold (high threshold), a high-risk fault warning is triggered, and a fault indication message is generated containing at least one of the following: fault location, fault type, and fault cause analysis. The fault cause analysis can be determined based on information such as fault type and fault location output by the fault prediction model.

[0090] In some embodiments, a warning of the corresponding level may be triggered when the fault prediction model predicts that the probability of a fault warning corresponding to a first number of consecutive charge domain pseudo-periods exceeds a first threshold or a second threshold.

[0091] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of extracting timing fault-related features using charge domain pseudo-periods provided in this application. In this embodiment, a first feature extraction is performed on a first charge domain sequence based on charge domain pseudo-periods to obtain timing fault-related features, including: S21: According to the pseudo-period of the charge domain, the first charge domain sequence is divided into several periodic segments to obtain the segment sequence.

[0092] In this embodiment, the operation of extracting the first feature from the first charge domain sequence to obtain the time-series fault-related features is performed by the time-series feature extraction module in the fault prediction model. Specifically, after obtaining the segment sequence corresponding to several periodic segments, it is input into the time-series feature extraction module for time-series fault-related feature extraction.

[0093] It should be noted that the number of periodic segments contained in the charge domain sequence is determined based on the period length corresponding to the charge domain pseudo-period. In some embodiments, the periodicity of the charge domain sequence can be analyzed by autocorrelation function, or the pseudo-period length can be determined by finding the interval of local extrema, or the pseudo-period length (charge domain pseudo-period) can be determined based on the dominant frequency of the charge domain spectrum obtained by performing a fast Fourier transform on the charge domain sequence.

[0094] In one specific embodiment, a fast Fourier transform can be performed on the charge domain sequence to obtain the charge domain spectrum; then, a preset number of frequency points with the largest amplitude in the charge domain spectrum can be used to calculate the charge domain pseudo-period.

[0095] The Fast Fourier Transform (FFT) is a computational algorithm that converts time-domain or frequency-domain signals into frequency-domain signals. Here, it is used to analyze the periodic patterns implicit in charge-domain sequences. The charge-domain spectrum refers to the frequency distribution of the signal after the FFT, with the horizontal axis representing frequency and the vertical axis representing amplitude, reflecting the intensity of different frequency components in the original signal. The preset number of frequency points refers to selecting the K frequency points with the largest amplitudes in the spectrum to capture the dominant periodic characteristics of the signal; K is a positive integer greater than or equal to 1. The calculated charge-domain pseudo-period refers to calculating the corresponding period length based on the selected frequency values. For example, the K frequency points with the largest amplitudes (e.g., K=3), f1, f2, and f3, correspond to periods T1, T2, and T3, respectively. T1 can be taken as the charge-domain pseudo-period, or T1, T2, and T3 can be taken separately, or a combined charge-domain pseudo-period can be obtained from T1, T2, and T3.

[0096] In one implementation scenario, the charge domain pseudo-periods are extracted using the period extraction module in the Time Series Network model. Specifically, the period extraction module performs a Fast Fourier Transform on the charge domain sequence and performs spectral analysis to extract a preset number of charge domain pseudo-periods with the largest amplitude.

[0097] In one embodiment, to facilitate processing by the temporal feature extraction module, several periodic segments can be folded to obtain a sequence of segments in the form of a two-dimensional tensor; that is, the sequence of segments obtained using several periodic segments is a sequence of segments in the form of a two-dimensional tensor. In one implementation scenario, the steps of segmenting periodic segments using charge domain pseudo-periods and constructing a sequence of segments in the form of a two-dimensional tensor are also performed by the temporal network model TimesNet, and the temporal feature extraction module is a network module in the temporal network model TimesNet.

[0098] Each row of the two-dimensional tensor corresponding to the segment sequence represents a periodic segment (one charge-discharge cycle), and each column represents the position of the same relative charge (i.e., the same charge). Among them, the same charge domain sequence may correspond to K charge domain pseudo-periods, and each charge domain pseudo-period corresponds to a segment sequence in the form of a two-dimensional tensor. Therefore, K charge domain pseudo-periods will correspond to K segment sequences in the form of two-dimensional tensors.

[0099] For example, if the length of the first charge domain sequence is 90 and the length of the charge domain pseudo-period is 30, then the first charge domain sequence can be divided into 3 periodic segments. Then, with the periodic segments as the row dimension and the charge quantity nodes in a single period as the column dimension, the 3 periodic segments are stacked and folded row by row in the order of charging and discharging to form a segment sequence in the form of a two-dimensional tensor.

[0100] In some implementation scenarios, if the first charge domain sequence of the input fault prediction model already contains multiple complete periodic segments, it can be directly segmented according to the pseudo-period of the charge domain; if the first charge domain sequence is streaming real-time time-series input data, it can be continuously segmented using a sliding window method to form the corresponding segment sequence.

[0101] S22: The first convolutional submodule of the temporal feature extraction module is used to extract local features within a period from the fragment sequence, and the second convolutional submodule of the temporal feature extraction module is used to extract global features across periods from the fragment sequence.

[0102] The kernel of the first convolutional submodule is smaller than the kernel of the second convolutional submodule.

[0103] It should be noted that step S21 may generate K segment sequences, i.e., K two-dimensional tensors. For each segment sequence, the corresponding feature extraction method is the same. The feature extraction process is briefly introduced below using a segment sequence as an example: When performing time-series fault-related features, multiple periodic segments can be input at once or continuously. That is, the current segment sequence to be processed may include one or more periodic segments. The fault prediction model processes each periodic segment in the segment sequence to obtain the intra-period local features and cross-period global features of the current periodic segment.

[0104] Specifically, each periodic segment can be taken as the current segment. Then, the first convolutional submodule is used to extract features from the current segment to obtain the local features within the period of the current segment. The second convolutional submodule is used to extract features based on the current segment and the reference segment to obtain the cross-period global features corresponding to the current segment. The reference segment is at least one periodic segment other than the current segment. The reference segment can be the periodic segments before the current segment, or the periodic segments before and after the current segment.

[0105] In one implementation scenario, the reference segment is the periodic segments preceding the current segment. For example, in a scenario where the first charge domain sequence is streaming real-time time-series input data, there are no subsequent periodic segments after the current segment. Therefore, in this implementation scenario, cross-period global feature extraction can be performed based on the current segment and the periodic segments preceding the current segment (historical segments).

[0106] In another implementation scenario, the reference segment includes all period segments before and after the current segment. For example, in a scenario where the first charge domain sequence already contains multiple complete period segments, cross-period global feature extraction of the current segment can be directly extracted based on the multiple period segments contained in the entire first charge domain sequence. Of course, in a scenario where the first charge domain sequence is streaming real-time time-series input data, a delay buffer can be set to wait for a small number of subsequent period segments to be collected before feature extraction. In this case, cross-period global feature extraction can be performed simultaneously using the period segments before and after the current segment.

[0107] S23: Integrate local features within a cycle and global features across cycles to obtain time-series fault-related features.

[0108] In one embodiment, for the current segment, the intra-period local features and cross-period global features of the current segment can be fused to obtain the temporal fault-related features corresponding to the current segment, so as to facilitate subsequent fault prediction using the temporal fault-related features corresponding to the current segment. The fusion operation can be implemented through splicing, weighted summation, or attention mechanisms, so that the obtained temporal fault-related features contain both fine-grained local anomaly features and stable global degradation features.

[0109] It should be noted that the above-mentioned segment sequence can be a segment sequence corresponding to the entire battery, or a segment sequence corresponding to each individual cell in the battery. In the scenario where it is a segment sequence for each individual cell, the above method can be used to extract the temporal fault-related features of each individual cell in each cycle segment, as well as the spatial fault-related features of each individual cell in each cycle segment.

[0110] Furthermore, it should be noted that when there are K segment sequences, each segment sequence is independently input into the timing feature extraction module. For each individual cell, the above processing yields K initial timing fault-related features. These K initial timing fault-related features are then fused to obtain the final timing fault-related features for that individual cell. During fusion, the K initial timing fault-related features are concatenated to generate a K-dimensional composite feature. This composite feature is then subjected to dimensionality reduction processing to eliminate multidimensional differences caused by different segment sequences and inconsistent period lengths, compressing it into a one-dimensional global feature. Alternatively, feature fusion can be achieved by directly summing the K timing fault-related features element by element.

[0111] Please see Figure 3 , Figure 3 This is a flowchart illustrating an embodiment of extracting space fault-related features based on the physical coupling relationship of each cell provided in this application. In this embodiment, based on the physical coupling relationship of each cell within the battery, a second feature is extracted from the second charge domain sequence to obtain space fault-related features, including: S31: Take the data corresponding to the same moment in the second charge domain sequence as a data frame, obtain the second battery parameter values ​​of each cell in the data frame, and use them as the initial frame features of the cells respectively.

[0112] In this embodiment, each moment corresponds to a charge quantity (the charge quantity node mentioned above), meaning that a data frame is actually data corresponding to a charge quantity node. For example, voltage data, internal resistance data, and temperature data at charge quantity node A are considered as one data frame. The data corresponding to charge quantity nodes are linked together by time. Furthermore, the same charge quantity nodes exist in different periodic segments; therefore, to distinguish them, time is used here to represent the charge quantity node at the corresponding time.

[0113] In one embodiment, a physical topology graph of the battery can be constructed based on the physical coupling relationship of each cell in the battery. Each node in the physical topology graph represents each cell in the battery, and the edges between nodes represent the physical adjacency relationship between the cells corresponding to the nodes. Then, the second battery parameter values ​​of each cell in the data frame are used as the initial features of the corresponding nodes. The specific second battery parameter values ​​of each cell can be referred to the description of the second battery parameter values ​​above.

[0114] For example, a battery is abstracted as a physical topology graph G=(V,E), where the number of nodes is the same as the total number of cells in the battery; each node V represents a cell in the battery, and each edge E represents the physical adjacency relationship between cells (such as physical adjacency, electrical series relationship and thermal coupling relationship). Only cells that are physically adjacent are connected to fit the real physical coupling relationship in the battery, such as fitting the real heat spread path in the battery.

[0115] S32: Take each cell as the target cell, determine the influence weight of the adjacent nodes on the target cell based on the initial frame features of the target cell and the initial frame features of the adjacent cells of the target cell, and obtain the updated frame features of the target cell by using the initial frame features of each adjacent cell of the target cell and the corresponding influence weights.

[0116] In one embodiment, the graph attention network in the fault prediction model can be used to perform self-attention processing on the initial frame features of each node in the physical topology graph to obtain the influence weights between adjacent nodes in the physical topology graph, wherein adjacent nodes are nodes connected by edges; for each node, the initial frame features of each adjacent node are weighted and aggregated using the influence weights corresponding to each adjacent node of the node to obtain the updated frame features of the node.

[0117] For example, the influence weights between adjacent nodes in a physical topology graph can be calculated using the following formula:

[0118] in, This represents the influence weight between node i and its neighboring node j. , Let i and j represent the initial frame features of nodes i and j respectively, and W represent the learnable weight matrix used to transform the initial frame features (such as linear transformation). Let i represent the set of neighboring nodes of node i, and k represent any neighboring node in the set of neighboring nodes. This represents a vector concatenation operation. This represents the transpose of the learnable attention parameter vector. Represents a non-linear activation function. This indicates exponentiation.

[0119] Understandably, when a battery cell experiences a micro-short circuit or localized heat buildup, the characteristic differences between the faulty node and its neighboring nodes will increase dramatically, and the corresponding attention weights (influence weights) will be significantly reconstructed. This method can effectively capture the spatial correlation anomalies caused by the fault.

[0120] For each node, the initial frame features of each of its neighboring nodes are weighted and aggregated using the influence weights of those neighboring nodes. The updated frame features of the node are obtained as follows:

[0121] in, This represents the update frame characteristics of node i. Represents a non-linear activation function. , , , The meaning of "etc." can be found in the description of the previous formula.

[0122] It should be noted that the initial frame features Already involved in the above-mentioned influence weighting The calculation process is such that the updated frame features obtained after weighted aggregation naturally integrate the initial frame features of the node itself and the features of neighboring nodes.

[0123] S33: By fusing the update frame features of each cell, spatial fault-related features of the battery at the same time are obtained.

[0124] This embodiment fuses the update frame features of all cells in the battery (such as splicing or weighted fusion) to obtain the spatial fault-related features of the battery at the same time.

[0125] Among them, the spatial fault-related features of the battery at the same time represent the physical coupling (e.g., electrothermal coupling) correlation state between the cells in the battery. It can quantify the differences in the distribution of battery parameters (voltage, temperature) between cells and reflect cross-cell spatial coupling faults such as cell consistency degradation, local heat conduction diffusion, and electrical matching abnormalities.

[0126] In other words, the above-mentioned method of extracting spatial fault-related features based on the physical coupling relationship of each cell is beneficial for capturing spatially related faults such as inconsistency anomalies and local thermal propagation between cells, thereby achieving accurate location and spatial perception of battery faults.

[0127] In one specific embodiment, the steps of feature extraction from the charge domain sequence and fault prediction using battery fault-related features are performed by a fault prediction model. The charge domain sequence includes a data sequence containing at least one charge domain pseudo-period. Specific charge domain pseudo-periods are described above.

[0128] In this embodiment, the fault prediction model sequentially performs the following processing on each charge domain pseudo-period of the charge domain sequence: extracting fault-related features corresponding to the charge domain pseudo-period, and performing fault prediction based on the fault-related features of the charge domain pseudo-period to obtain the fault prediction result corresponding to the charge domain pseudo-period. That is, in this embodiment, the fault prediction model performs prediction cycle by cycle, obtaining the fault prediction result of the battery corresponding to each charge domain pseudo-period. The predicted charge domain pseudo-period can be at least one charge domain pseudo-period after the current charge domain pseudo-period, or it can be the current charge domain pseudo-period.

[0129] The fault-related features corresponding to the charge domain pseudo-period include temporal fault-related features and spatial fault-related features. The temporal fault-related features include intra-period local features and cross-period global features. The intra-period local features are extracted from the current segment of the charge domain pseudo-period, while the cross-period global features are extracted from the current segment of the charge domain pseudo-period and reference segments from other charge domain periods. The spatial fault-related features are obtained by fusing the spatial fault-related features corresponding to each moment (corresponding to each charge quantity in the period) within the charge domain pseudo-period. Specific implementation details are provided above and will not be elaborated upon here.

[0130] In some embodiments, the fault prediction model can also perform fault prediction based on data from multiple charge domain pseudo-periods in the current charge domain sequence, or it can perform fault prediction using data from the current charge domain sequence and the new charge domain sequence when a new charge domain sequence is input subsequently.

[0131] In some embodiments, the steps of feature extraction from the charge domain sequence and fault prediction using battery fault-related features are performed by a fault prediction model. The fault prediction model needs to be trained before inference can be performed using it.

[0132] In some embodiments, please refer to Figure 4 , Figure 4 This is a flowchart illustrating an embodiment of the fault prediction model training method provided in this application. This embodiment includes: S41: Obtain the sample charge domain sequence of the sample battery.

[0133] The method for obtaining the sample charge domain sequence of the sample battery can be found in the section on obtaining the charge domain sequence above.

[0134] S42: The fault prediction model is used to process the sample charge domain sequence to obtain the sample fault prediction result.

[0135] In some embodiments, the fault prediction model sequentially outputs the sample fault prediction results corresponding to each sample period segment in the sample charge domain sequence. Each sample period segment corresponds to a sample charge domain pseudo-period in the sample charge domain sequence. The sample fault prediction results include the sample fault warning probability of the sample battery.

[0136] The method for determining each periodic segment of the sample charge domain sequence can be found in the description of the relevant schemes for determining each periodic segment of the charge domain sequence above; the method for sequentially outputting the sample fault prediction results corresponding to each periodic segment in the sample charge domain sequence using the fault prediction model can be found in the description of sequentially outputting the fault prediction results corresponding to each periodic segment above.

[0137] In other embodiments, the fault prediction model may also perform fault prediction based on the entire sample charge domain sequence or multiple sample periodic segments in the sample charge domain sequence.

[0138] S43: Determine the prediction loss of the fault prediction model based on the difference between the sample fault prediction results and the fault labeling results of the sample charge domain sequence.

[0139] S44: Determine the actual electrochemical degradation state of the battery based on the sample fault prediction results, and determine the physical penalty loss of the battery electrochemical degradation mechanism based on the difference between the actual electrochemical degradation state and the theoretical electrochemical degradation state of the battery.

[0140] In one embodiment, the true electrochemical degradation state can be obtained based on the rate of change of the current sample fault warning probability relative to the historical sample fault warning probability; wherein, the current sample fault warning probability is the sample fault warning probability output for the current sample period segment, and the historical sample fault warning probability is the sample fault warning probability output for the previous sample period segment.

[0141] In one embodiment, the theoretical electrochemical degradation state of the battery is represented by the Arrhenius reaction rate, and the corresponding physical penalty loss of the battery electrochemical degradation mechanism is expressed as follows:

[0142] in, Indicates physical penalty loss. This indicates the probability of a fault warning for the current sample. This represents the rate of change of the current sample's fault warning probability relative to the historical sample's fault warning probability (the rate of change of battery health over time), used to characterize the actual electrochemical degradation state. The Arrhenius reaction rate is used to characterize the theoretical electrochemical degradation state of a battery (characterizing the theoretical aging rate), and A represents the prefactor. R represents the activation energy of the reaction (an empirical value), R represents the universal gas constant, and T represents the temperature (the temperature at which the data is collected from a single cell or battery).

[0143] It is understandable that when the battery temperature rises, it will accelerate battery degradation, and the corresponding model prediction of the current fault warning probability should increase. Therefore, if the current fault warning probability decreases, it does not conform to the electrochemical mechanism.

[0144] S45: Combine the predicted loss and the physical penalty loss to obtain the total loss.

[0145] In this embodiment, the sum of the prediction loss and the physical weighting result is taken as the total loss. The physical weighting result is the product of the physical penalty loss and the physical penalty weight.

[0146] For example, the total loss is represented as follows:

[0147] in, Indicates the total loss. Indicates the predicted loss. Indicates physical penalty loss. This represents the physical penalty weight.

[0148] In one implementation scenario, This is an empirical value determined based on experience. In another implementation scenario, It is a dynamic value whose magnitude is related to the environment in which the battery is located; among them, during the training of the fault prediction model, the physical penalty weight can be automatically increased in extreme environments such as extreme cold and extreme heat to constrain the prediction boundary of the model and ensure the physical rationality of the fault prediction results under extreme environmental conditions.

[0149] S46: Adjust the parameters of the fault prediction model based on the total loss.

[0150] For example, the Adam optimizer can be used to minimize the total loss in order to adjust the network parameters of the fault prediction model.

[0151] It should be noted that this embodiment takes into account that traditional deep learning models rely too much on the statistical distribution of historical data. When faced with extreme conditions (such as extreme cold, extreme heat, and high-rate charging and discharging) that are less common in the training set (sample charge domain sequence), their output results often deviate from the basic electrochemical mechanism.

[0152] To address this issue, this embodiment embeds battery electrochemical degradation mechanisms, such as the Arrhenius equation, as physical regularization penalty terms into the loss function corresponding to the total loss of the fault prediction model. This forcibly constrains the parameter update space of the fault prediction model, ensuring that the fault prediction results strictly follow the basic laws of thermodynamics and electrochemistry even under extreme long-tail conditions. This effectively avoids erroneous predictions from purely data-driven models that violate physical common sense, significantly improving the robustness and early warning confidence of the fault prediction model.

[0153] Figure 5 This is a schematic diagram of a framework of an embodiment of the fault prediction device provided in this application. In this embodiment, the fault prediction device 50 includes a memory 51 and a processor 52 coupled to each other.

[0154] The memory 51 stores program instructions, and the processor 52 executes the program instructions stored in the memory 51 to implement the steps of any of the above-described method implementations. In a specific implementation scenario, the battery failure prediction device 50 may include, but is not limited to, a microcomputer or a server. In addition, the battery failure prediction device 50 may also include mobile devices such as laptops and tablets, which are not limited here.

[0155] Specifically, processor 52 controls itself and memory 51 to implement the steps of any of the above embodiments. Processor 52 may also be referred to as a CPU (Central Processing Unit). Processor 52 may be an integrated circuit chip with signal processing capabilities. Processor 52 may also be a general-purpose processor, 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. A general-purpose processor may be a microprocessor or any conventional processor. Furthermore, processor 52 may be implemented using integrated circuit chips.

[0156] Please see Figure 6 , Figure 6 This is a schematic diagram of the framework of the computer-readable storage medium provided in this application. The computer-readable storage medium 60 of this application embodiment stores program instructions 61, which, when executed, implement the methods provided in any embodiment or any non-conflicting combination of the above-described methods. The program instructions 61 can form a program file and be stored in the computer-readable storage medium 60 in the form of a software product, so that a computer device (which may be a personal computer, server, or network device, etc.) can execute all or part of the steps of the methods of various embodiments of this application. The aforementioned computer-readable storage medium 60 includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0157] The above scheme, which maps time-series data to the charge domain, can unify the fragmented battery usage process with varying durations using charge as the measurement benchmark. This eliminates data scale differences caused by different battery usage conditions, such as fast / slow charging and differences in initial SOC. It enables horizontal comparison of battery data under different usage conditions, making it easier to extract effective fault features and improve the accuracy of battery fault prediction.

[0158] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0159] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0160] 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, depending on actual needs.

[0161] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0162] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A battery fault prediction method, characterized in that, include: Acquire time-series data about multiple battery parameters, wherein the time-series data is collected during battery use and includes the first parameter values ​​of the multiple battery parameters at different collection times; The time-series data is mapped to the charge domain to obtain a charge domain sequence, wherein the charge domain sequence includes second parameter values ​​corresponding to different charge amounts for the plurality of battery parameters; Feature extraction is performed on the charge domain sequence to obtain the fault-related features of the battery; Fault prediction is performed using the fault-related features of the battery to obtain the fault prediction result of the battery.

2. The method according to claim 1, characterized in that, The plurality of battery parameters include current; the step of mapping the time-series data into the charge domain to obtain a charge domain sequence includes: The current value at each acquisition time is obtained from the time series data; The charge of the battery at each sampling time is obtained by integrating the current values ​​at each sampling time. Using the charge amount at each acquisition time and the first parameter value of the battery parameters at each acquisition time, multiple sets of parameters corresponding to different charge amounts are generated, wherein the parameter sets include the second parameter value of each battery parameter mapped to the charge amount.

3. The method according to claim 2, characterized in that, The process of generating multiple sets of parameters corresponding to different charge amounts using the charge amount at each acquisition time and the first parameter value of the battery parameters at each acquisition time includes: Multiple mapping groups are formed by using the charge amount at each acquisition time and the first parameter value of the battery parameters at each acquisition time. Each mapping group includes the charge amount at the same acquisition time and the first parameter value of each battery parameter. By utilizing the mapping relationship between the charge quantity represented by the multiple mapping sets and each of the battery parameters, multiple parameter sets corresponding to different charge quantities are generated.

4. The method according to claim 1, characterized in that, The charge domain sequence includes multiple sets of parameters corresponding to each charge node during battery use. Each charge node takes values ​​at preset charge step intervals. The parameter set includes the second parameter value of each battery parameter corresponding to the charge node.

5. The method according to claim 1, characterized in that, Before performing feature extraction on the charge domain sequence to obtain the fault-related features of the battery, the method further includes: Using the charge domain sequence, at least one electrochemical feature sequence of the battery in the charge domain is generated, and each electrochemical feature sequence includes a second parameter value corresponding to a different charge amount; The at least one electrochemical feature sequence is added to the charge domain sequence.

6. The method according to claim 1, characterized in that, The plurality of battery parameters include voltage; generating at least one electrochemical characteristic sequence of the battery in the charge domain using the charge domain sequence includes: By using the voltage values ​​corresponding to different charges in the charge domain sequence, the incremental capacity sequence and differential voltage sequence in the charge domain are obtained; The incremental capacity sequence and the differential voltage sequence are used as the electrochemical characteristic sequence.

7. The method according to claim 1, characterized in that, The step of extracting features from the charge domain sequence to obtain fault-related features of the battery includes: Identify the charge domain pseudo-period of the charge domain sequence, and perform a first feature extraction on the first charge domain sequence based on the charge domain pseudo-period to obtain time-series fault-related features. The time-series fault-related features include at least one of intra-period local features and cross-period global features. The first charge domain sequence includes second parameter values ​​corresponding to different charge amounts for first battery parameters. The first battery parameter is at least one of the battery parameters included in the charge domain sequence; and / or, Based on the physical coupling relationship of each cell in the battery, a second feature is extracted from the second charge domain sequence to obtain spatial fault-related features. The second charge domain sequence includes second parameter values ​​corresponding to different charge amounts of the second battery parameters, and the second battery parameters are at least one of the battery parameters included in the charge domain sequence.

8. The method according to claim 7, characterized in that, The identification of the charge domain pseudo-period of the charge domain sequence includes: Perform a Fast Fourier Transform on the charge domain sequence to obtain the charge domain spectrum; The pseudo-period of the charge domain is obtained by converting a preset number of frequency points with the largest amplitude in the charge domain spectrum.

9. The method according to claim 7, characterized in that, The timing fault-related features include local features within a cycle and global features across cycles; The first feature extraction based on the pseudo-period of the charge domain sequence to obtain time-series fault-related features includes: According to the pseudo-period of the charge domain, the first charge domain sequence is divided into several periodic segments to obtain a segment sequence; The first convolutional submodule of the temporal feature extraction module extracts local features within a period from the segment sequence, and the second convolutional submodule of the temporal feature extraction module extracts global features across periods from the segment sequence, wherein the convolutional kernel of the first convolutional submodule is smaller than the convolutional kernel of the second convolutional submodule; By fusing the local features within the cycle and the global features across the cycle, the timing fault-related features are obtained.

10. The method according to claim 9, characterized in that, The first convolutional submodule of the temporal feature extraction module extracts local features within a period from the segment sequence, and the second convolutional submodule of the temporal feature extraction module extracts global features across periods from the segment sequence, including: Each of the aforementioned periodic segments is taken as the current segment; The first convolutional submodule is used to extract features from the current segment to obtain local features within the period of the current segment. The second convolutional submodule is used to extract features based on the current segment and a reference segment to obtain global features across the period of the current segment. The reference segment is at least one period segment other than the current segment. The fusion of the intra-period local features and the cross-period global features to obtain the time-series fault-related features includes: By fusing the intra-period local features and cross-period global features of the current segment, the timing fault-related features corresponding to the current segment are obtained.

11. The method according to claim 9, characterized in that, The segment sequence is a two-dimensional tensor, which is obtained by folding the several periodic segments.

12. The method according to claim 7, characterized in that, The local features within the period are the fluctuation features within the period, and the global features across the period are the global degradation features across the period.

13. The method according to claim 7, characterized in that, The first battery parameters include voltage, temperature, incremental capacity, and differential voltage.

14. The method according to claim 7, characterized in that, Based on the physical coupling relationship of each cell within the battery, a second feature extraction is performed on the second charge domain sequence to obtain spatial fault-related features, including: The data corresponding to the same moment in the second charge domain sequence are taken as a data frame, and the second battery parameter values ​​of each cell in the data frame are obtained, which are respectively used as the initial frame features of the cell. Each of the aforementioned battery cells is taken as a target battery cell. Based on the initial frame features of the target battery cell and the initial frame features of the neighboring battery cells of the target battery cell, the influence weight of the neighboring nodes on the target battery cell is determined. Then, using the initial frame features of each neighboring battery cell of the target battery cell and the corresponding influence weight, the updated frame features of the target battery cell are obtained. By fusing the update frame features of each of the battery cells, the spatial fault-related features of the battery at the same time are obtained.

15. The method according to claim 14, characterized in that, The step of obtaining the second battery parameter values ​​of each cell in the data frame, and using them as the initial frame features of the cell, includes: Based on the physical coupling relationship of each cell in the battery, a physical topology graph of the battery is constructed, wherein each node in the physical topology graph represents each cell in the battery, and the edges between the nodes represent the physical adjacency relationship between the cells corresponding to the nodes. The second battery parameter values ​​of each cell in the data frame are used as the initial features of the corresponding node; The process involves taking each of the aforementioned battery cells as a target battery cell, determining the influence weight of the adjacent nodes on the target battery cell based on the initial frame features of the target battery cell and the initial frame features of its neighboring battery cells, and obtaining the updated frame features of the target battery cell using the initial frame features and corresponding influence weights of each neighboring battery cell. This includes: The initial frame features of each node in the physical topology graph are processed by a graph attention network to obtain the influence weights between adjacent nodes in the physical topology graph, wherein the adjacent nodes are nodes connected by the edges. For each node, the initial frame features of each of the node's neighboring nodes are weighted and aggregated using the influence weights corresponding to each of the node's neighboring nodes to obtain the updated frame features of the node.

16. The method according to claim 14, characterized in that, The second battery parameters include voltage, temperature, internal resistance, and state of charge.

17. The method according to claim 1, characterized in that, The steps of feature extraction from the charge domain sequence and fault prediction using the fault-related features of the battery are performed by the fault prediction model. The fault prediction model sequentially performs the following processing on each charge domain pseudo-period of the charge domain sequence: extracting the fault-related features corresponding to the charge domain pseudo-period, and performing fault prediction based on the fault-related features of the charge domain pseudo-period to obtain the fault prediction result corresponding to the charge domain pseudo-period. The fault-related features corresponding to the charge domain pseudo-period include the temporal fault-related features and the spatial fault-related features. The temporal fault-related features include intra-period local features and cross-period global features. The intra-period local features are extracted from the current segment of the charge domain pseudo-period. The cross-period global features are extracted from the current segment of the charge domain pseudo-period and reference segments of other charge domain periods. The spatial fault-related features are obtained by fusing the spatial fault-related features corresponding to each moment in the charge domain pseudo-period.

18. The method according to claim 1, characterized in that, The fault-related features include temporal fault-related features and spatial fault-related features; The step of using the fault-related characteristics of the battery to perform fault prediction and obtain the fault prediction result of the battery includes: By fusing the temporal fault-related features and spatial fault-related features, a fused feature is obtained; Based on the fusion features, the fault prediction result is obtained.

19. The method according to claim 1, characterized in that, The steps of feature extraction from the charge domain sequence and fault prediction using the fault-related features of the battery are performed by the fault prediction model. The method further includes: Obtain the sample charge domain sequence of the sample battery; The fault prediction model is used to process the sample charge domain sequence to obtain the sample fault prediction result; Based on the difference between the sample fault prediction results and the fault labeling results of the sample charge domain sequence, the prediction loss of the fault prediction model is determined. Based on the sample fault prediction results, the actual electrochemical degradation state of the battery is determined, and based on the difference between the actual electrochemical degradation state and the theoretical electrochemical degradation state of the battery, the physical penalty loss of the battery electrochemical degradation mechanism is determined. The total loss is obtained by combining the predicted loss and the physical penalty loss. Based on the total loss, the parameters of the fault prediction model are adjusted.

20. The method according to claim 19, characterized in that, The fault prediction model sequentially outputs the sample fault prediction results corresponding to each sample period segment in the sample charge domain sequence. Each sample period segment corresponds to a sample charge domain pseudo-period in the sample charge domain sequence. The sample fault prediction results include the sample fault warning probability of the sample battery. Determining the true electrochemical degradation state of the battery based on the sample fault prediction results includes: The true electrochemical degradation state is obtained based on the rate of change of the current sample fault warning probability relative to the historical sample fault warning probability, wherein the current sample fault warning probability is the sample fault warning probability output for the current sample period segment, and the historical sample fault warning probability is the sample fault warning probability output for the previous sample period segment.

21. The method according to claim 19, characterized in that, The theoretical electrochemical degradation state of the battery is the Arrhenius reaction rate.

22. The method according to claim 19, characterized in that, The total loss, obtained by combining the prediction loss and the physical penalty loss, includes: The sum of the predicted loss and the physical weighted result is taken as the total loss. The physical weighted result is the product of the physical penalty loss and the physical penalty weight. The magnitude of the physical penalty weight is related to the environment in which the battery is located.

23. The method according to claim 1, characterized in that, The fault prediction results include the fault type and fault warning probability of the battery; After performing fault prediction using the fault-related features of the battery to obtain the fault prediction result of the battery, the method further includes: In response to the fault warning probability exceeding a first threshold, a potential anomaly warning is triggered; In response to the fault warning probability exceeding a second threshold, a high-risk fault warning is triggered, and fault prompt information is generated, wherein the second threshold is greater than the first threshold, and the fault prompt information includes at least one of fault location, fault type, and fault cause analysis.

24. The method according to claim 1, characterized in that, The battery is a vehicle battery, and the time-series data is extracted from vehicle operation data.

25. A fault prediction device, characterized in that, Including interconnected memory and processor, The memory stores program instructions; The processor is configured to execute program instructions stored in the memory to implement the method according to any one of claims 1-24.

26. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that can be executed by a processor to implement the method of any one of claims 1-24.

Images