A method and system for predicting failure of a power battery

Abstract

The application relates to the technical field of fault prediction, in particular to a power battery fault prediction method, which comprises the following steps: S1: collecting power battery operation data through a data acquisition module, pre-processing the collected data to obtain a feature sequence suitable for model input; S2: constructing a MoKAN fault prediction model, wherein the MoKAN fault prediction model outputs a predicted value of a battery operation parameter by using the input feature sequence; wherein the MoKAN fault prediction model comprises a mixed expert module, a KAN module and a prediction head submodule; S3: judging whether the power battery has a fault risk according to the error between the predicted value and a real-time measured value. The application provides a power battery fault prediction method and system with high precision, high adaptability and high efficiency.

Description

Technical Field

[0001] This application relates to the field of battery fault prediction technology, and in particular to a method and system for predicting power battery faults. Background Technology

[0002] The core of power battery fault prediction is to predict the future health status or fault occurrence time based on multi-parameter time-series data during battery operation, such as voltage, current, temperature, and internal resistance. This is a key technology to ensure the safe and reliable operation of systems such as new energy vehicles and energy storage power stations. In reality, power battery monitoring data exhibits typical heterogeneity and non-stationarity: in the short term, there are transient fluctuations such as peak charging current and sudden temperature rises; in the long term, there is a slow degradation trend such as capacity decay and increased internal resistance. Furthermore, the data patterns differ significantly under different operating conditions such as fast charging, low temperature, and high load. These characteristics make fault prediction a serious challenge.

[0003] Existing methods for predicting power battery failures have significant limitations. Traditional methods based on electrochemical models have clear physical meanings, but their accuracy is highly dependent on the accuracy of parameter identification, sensitive to individual battery differences and aging state changes, and computationally complex, making them difficult to apply in real-time to large-scale battery systems. Methods based on traditional machine learning, such as statistical regression and support vector machines, typically struggle to effectively describe the complex nonlinear aging mechanisms and long-term degradation trends of batteries, have limited predictive ability for sudden failures, and perform poorly in handling non-stationary sequences.

[0004] With the development of deep learning, related methods have gradually become mainstream. However, basic models such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs) have their own shortcomings in capturing the multi-scale time dependence and strong coupling relationships between parameters of battery data, making it difficult to efficiently model both short-term fluctuations and long-term degradation trends simultaneously. Transformer models have shown potential with their powerful global dependency modeling capabilities, but the secondary computational complexity of their self-attention mechanism is prominent, resulting in huge computational overhead when processing long-sequence battery data with high sampling rates, limiting their real-time application capabilities. In addition, existing deep learning methods generally face the problem of "insufficient specialized modeling": a single model needs to process multiple dynamic modes of the battery simultaneously, which can easily lead to the dominant mode masking subtle fault-related features, thus affecting the accuracy of fault prediction. Therefore, there is an urgent need for a fault prediction scheme that balances efficient computation, accurate capture of multi-mode features, and adaptation to the characteristics of battery data. Summary of the Invention

[0005] Therefore, it is necessary to provide a power battery fault prediction method and system with high accuracy, high adaptability, and high efficiency.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A method for predicting power battery faults includes the following steps:

[0008] S1: The power battery operation data is collected through the data acquisition module, and then the collected data is preprocessed to obtain the feature sequence for the adaptation model input; the power battery operation data is time series data, including the voltage, current, temperature and internal resistance of the power battery under different operating conditions;

[0009] S2: Construct a MoKAN fault prediction model. The MoKAN fault prediction model uses the input feature sequence to output the predicted values ​​of battery operating parameters. The MoKAN fault prediction model includes a hybrid expert module, a KAN module, and a prediction head sub-module.

[0010] S3: Based on the error between the predicted value and the real-time measurement value, determine whether there is a risk of failure in the power battery.

[0011] Furthermore, the preprocessing in step S1 includes channel normalization, Fourier adaptive patch partitioning, and patch and position embedding.

[0012] Furthermore, the method for preprocessing the data in step S1 is as follows:

[0013] S1.1: Normalize the collected battery data samples;

[0014] S1.2: Analyze the dominant period of the battery data through Fast Fourier Transform, determine the patch length based on the period, and divide the normalized long sequence into multiple semantically coherent short segments.

[0015] S1.3: Map each patch to a fixed-dimensional feature space using simple linear projection, and then add location information.

[0016] Furthermore, the specific steps for outputting the predicted values ​​of the battery operating parameters in step S2 are as follows:

[0017] S2.1: Feature extraction is performed using a hybrid expert module;

[0018] S2.2: The extracted features are refined using the KAN module to obtain refined features;

[0019] S2.3: The prediction head module outputs predicted values ​​of battery operating parameters based on refined features.

[0020] Furthermore, in step S2, the construction of the MoKAN fault prediction model specifically includes the following steps:

[0021] S2-1: Construct a hybrid expert module, which contains several expert models and a gating network. Each expert model adopts a dual-path structure that combines a KAN network with one-dimensional convolution. The gating network dynamically selects one or more expert models to process the input features.

[0022] S2-2: Construct a KAN module, which adopts a dual-path structure that combines KAN network and one-dimensional convolution, and refines the features output by the hybrid expert module.

[0023] S2-3: Construct a prediction head submodule, which uses a single-layer fully connected network to output predicted values ​​of battery operating parameters based on refined features.

[0024] Furthermore, in step S3, the method for determining whether the power battery has a fault risk based on the error between the predicted value and the real-time measured value specifically includes the following steps: calculating the error between the predicted value and the real-time measured value, setting a fault threshold, comparing the calculated error with the fault threshold, and determining that the power battery has a fault risk when the error exceeds the fault threshold.

[0025] Formula for calculating error:

[0026]

[0027] in, Indicates real-time measurement value, Indicates the predicted value. Indicates error;

[0028] Set a fault threshold, when the error When the fault threshold is exceeded, the power battery is deemed to be at risk of failure.

[0029] Furthermore, the data acquisition module is a sensor arranged at the end of the power battery pack system.

[0030] Furthermore, the method for collecting power battery operating data in step S1 is as follows: the voltage, current, temperature, and internal resistance data of the power battery under different operating conditions are obtained through the sensor.

[0031] This application also provides a power battery fault prediction system for executing the power battery fault prediction method described in any of the above embodiments, including:

[0032] The data acquisition module is used to collect operating data of the power battery;

[0033] The MoKAN fault prediction model includes a hybrid expert module, a KAN submodule, and a prediction head submodule; it is used to output predicted values ​​of battery performance parameters based on feature sequences generated from preprocessed battery operating data.

[0034] Furthermore, the preprocessing includes channel normalization, Fourier adaptive patch partitioning, and patch and position embedding operations; the hybrid expert module comprises several expert models and a gating network, each expert model employing a dual-path structure of parallel KAN network and one-dimensional convolution; the KAN module employs a dual-path structure of parallel KAN network and one-dimensional convolution; the prediction head submodule employs a single-layer fully connected network.

[0035] Understandably, the error between the predicted value and the real-time measurement value is used to determine whether there is a risk of failure in the power battery.

[0036] Among them, the KAN network is a lightweight and efficient modeling network that, based on mathematical theorems, can capture complex nonlinear relationships through simple function combinations. Compared with traditional networks, it has fewer parameters and faster computation, effectively improving efficiency. Combining the specialized division of labor approach of a mixture of experts (MoE), and further adapting the periodicity of battery data (such as charge-discharge cycles) through Fourier-guided adaptive patching, high-precision, high-efficiency, and highly adaptable power battery fault prediction is ultimately achieved.

[0037] Compared with existing technologies, the MoKAN fault prediction model in this invention's power battery fault prediction method is configured as a hybrid expert module, a KAN module, and a prediction head sub-module. Employing a lightweight and simplified KAN network structure, it eliminates the need for complex mathematical derivations, efficiently capturing nonlinear degradation relationships in batteries through simple function combinations. Its computation speed is over 30% faster than Transformer, fully meeting the needs of real-time applications such as vehicle-mounted and energy storage power stations. The multi-expert division of labor design in the hybrid expert module allows different experts to focus on specific patterns such as short-term fluctuations and long-term degradation, accurately capturing features from multiple models and avoiding the drawbacks of a single model overlooking certain aspects. It is particularly effective in accurately identifying subtle anomalies before faults, significantly improving the sensitivity of fault prediction and making the overall operation more efficient. The fault judgment logic is simple and intuitive, using a comparison of relative error and a threshold. The threshold can be flexibly calibrated according to different battery types such as ternary lithium and lithium iron phosphate, adapting to various power battery application scenarios with high adaptability, providing direct and reliable protection for battery safety. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of the present invention.

[0040] Figure 2 This is a flowchart illustrating the data preprocessing process in an embodiment of the present invention.

[0041] Figure 3 This is a flowchart illustrating the predicted values ​​of battery operating parameters in an embodiment of the present invention.

[0042] Figure 4 This is a framework diagram of the present invention.

[0043] Figure 5 This is a schematic diagram of time-series data segmentation in an embodiment of the present invention.

[0044] Figure 6 This is a schematic diagram of the structure of the MoKAN fault prediction model in an embodiment of the present invention. Detailed Implementation

[0045] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0046] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used in this application includes any and all combinations of one or more of the associated listed items.

[0047] This application provides a method for predicting power battery faults. See [link to relevant documentation]. Figures 1-6 This includes the following steps:

[0048] S1: Collect power battery operating data through the data acquisition module, preprocess the collected data to obtain the feature sequence for the adaptation model input; the power battery operating data is time series data, including the voltage, current, temperature and internal resistance of the power battery under different operating conditions;

[0049] S2: Construct the MoKAN fault prediction model. The MoKAN fault prediction model uses the input feature sequence to output the predicted values ​​of battery operating parameters. The MoKAN fault prediction model includes a hybrid expert module, a KAN module, and a prediction head sub-module.

[0050] S3: Based on the error between the predicted value and the real-time measurement value, determine whether there is a risk of failure in the power battery.

[0051] The MoKAN fault prediction model employs a lightweight and simplified KAN network structure, increasing computational speed by over 30%. Its multi-expert collaborative design with hybrid expert modules accurately captures features from multiple models and precisely identifies subtle anomalies preceding faults. By comparing relative error with a threshold, it adapts to various power battery application scenarios. This results in a fault prediction method that balances efficient computation, accurate capture of multi-mode features, and adaptation to battery data characteristics.

[0052] The data acquisition module consists of sensors deployed at the power battery pack system end. The sensor model is not limited, as long as it meets the acquisition requirements. The method for acquiring power battery operating data in step S1 is as follows: The sensors obtain data on the voltage, current, temperature, and internal resistance of the power battery under different operating conditions.

[0053] For details, see Figure 2 The preprocessing in step S1 includes channel normalization, Fourier adaptive patch partitioning, and patch and position embedding. Furthermore, the method for preprocessing the data is as follows:

[0054] S1.1: Normalize the collected battery data samples;

[0055] S1.2: Analyze the dominant period of the battery data through Fast Fourier Transform, determine the patch length based on the period, and divide the normalized long sequence into multiple semantically coherent short segments.

[0056] S1.3: Map each patch to a fixed-dimensional feature space through simple linear projection (patch embedding), and then add location information (location embedding).

[0057] Understandably, by using Fourier adaptive patching to precisely match the battery charge and discharge cycles, the problem of "segmented timing patterns" in traditional fixed patches is solved, ensuring that fault precursor features are not split, thus providing a solid data foundation for accurate prediction.

[0058] See Figure 3 Specifically, the steps for outputting the predicted values ​​of battery operating parameters in step S2 are as follows:

[0059] S2.1: Feature extraction is performed using a hybrid expert module;

[0060] S2.2: The extracted features are refined using the KAN module to obtain refined features;

[0061] S2.3: The prediction head module outputs predicted values ​​of battery operating parameters based on refined features.

[0062] Among them, battery operating parameters refer to the performance parameters of the battery during operation, specifically voltage, current, temperature, and internal resistance.

[0063] Furthermore, the specific steps for constructing the MoKAN fault prediction model in step S2 are as follows:

[0064] S2-1: Construct a hybrid expert module, which contains several expert models and a gating network. Each expert model adopts a dual-path structure that combines a KAN network with one-dimensional convolution. The gating network dynamically selects one or more expert models to process the input features. Here, "several" means more than two.

[0065] S2-2: Construct a KAN module, which adopts a dual-path structure that combines KAN network and one-dimensional convolution, and refines the features output by the hybrid expert module.

[0066] S2-3: Construct a prediction head submodule, which uses a single-layer fully connected network to output predicted values ​​of battery operating parameters based on refined features.

[0067] In step S3, the method for determining whether the power battery has a fault risk based on the error between the predicted value and the real-time measured value includes the following steps: calculating the error between the predicted value and the real-time measured value, setting a fault threshold, comparing the calculated error with the fault threshold, and determining that the power battery has a fault risk when the error exceeds the fault threshold.

[0068] Formula for calculating error:

[0069]

[0070] in, Indicates real-time measurement value, Indicates the predicted value. Indicates error;

[0071] Set a fault threshold, when the error When the fault threshold is exceeded, the power battery is deemed to be at risk of failure.

[0072] The following description uses a specific optional embodiment as an example.

[0073] S1.1: Sensors are deployed at the power battery pack system end to acquire historical time-series data such as voltage, current, temperature, and internal resistance of the power battery under different operating conditions. The collected battery data samples... After normalization, the expression is:

[0074]

[0075] in, This represents the normalized sample. and These represent the minimum and maximum values ​​in the feature dimension, respectively. Represents the original sample. and These represent the sample length and the number of features, respectively.

[0076] S1.2: Perform a Fast Fourier Transform (FFT) on the normalized time-series data to solve for the frequency distribution of the data. The FFT formula is as follows:

[0077]

[0078] in, For normalized time series data, The number of points in a single segment of time series data. For frequency indexing.

[0079] See Figure 5 By calculating the complex spectrum amplitude Exclude DC component ( After that, the frequency with the largest amplitude is selected as the dominant frequency. Convert the dominant frequency into a period ,Pick As patch length Set the sliding step size. The number of patches is calculated using the following formula:

[0080]

[0081] in, The formula represents the total number of data points in the historical time-series data of power batteries, and is used to divide the long sequence into... Each of the following short, semantically coherent segments (patches) has a dimension of [missing information]. Ensure that each patch contains a complete local pattern (such as a complete charge or discharge process).

[0082] S1.3: Through a trainable linear projection matrix ( For the embedding dimension), each patch ( For patch index, Mapped to In the feature space, we obtain the patch embedding. , dimension Introducing position embedding The temporal location features are encoded using trigonometric functions, as shown in the formula:

[0083]

[0084] in, For embedded dimension index ( The patch embedding is added to the position embedding to obtain the final feature sequence. , dimension This is used as input for subsequent models.

[0085] S2.1: Construct a hybrid expert module, which includes... Each expert model employs a dual-path structure combining a KAN network and one-dimensional convolutions.

[0086] The KAN network part adopts Polynomial basis functions of order 1, input dimension Output dimension The feature transformation formula is:

[0087]

[0088] in, (Normalize the input to [-1,1]). The learnable coefficient, For the first Chebyshev polynomials of order 3. The one-dimensional convolution part uses a kernel size of 3, a stride of 1, and padding of 1. Convolution, number of groups = The activation function is The output dimension is The final output of the expert model is the sum of the features from both paths:

[0089]

[0090] in, For the first The output feature vector of each patch after processing by the expert model For the KAN network to the first patch feature vectors The processing results ( (Dimensions).

[0091] The gating network calculates the affinity scores between the input features and each expert through linear projection, using the following formula:

[0092]

[0093] in, The linear projection weights of the gated network, For the first Each patch feature vector and Affinity score vector for each expert.

[0094] set up for The number of experts selected affects the score. conduct Select, keep previous The expert corresponding to the highest score The remaining scores are set as follows: The weight vector is obtained through Softmax normalization:

[0095]

[0096] in, The normalized expert weight vector, only the first... The weight corresponding to each selected expert is a non-zero value.

[0097] Set routing threshold Calculate the weight percentage for each expert. ( For expert indexing, If the proportion exceeds Then activate the expert. This is the expert activation flag (0 or 1). The output of the MoE module is a weighted aggregation of the activated experts:

[0098]

[0099] in, For the first The output feature vector of each patch after processing by the MoE module For the first The expert on the first The results of processing each patch.

[0100] To ensure balanced utilization of experts, an equilibrium loss is introduced. ,in To balance the loss weights, The coefficient of variation is 1. For the first The activation frequency percentage of each expert; and the introduction of routing loss. The auxiliary loss is , , The weighting coefficients are used to assist in the loss calculation.

[0101] S2.2: Construct the KAN module, which adopts the same dual-path structure as the expert model. The multinomial order and convolutional layer parameters of the KAN network are exactly the same as those of the expert model, and the output is fused through residual connections.

[0102]

[0103] in, For the first The feature vectors of each patch are refined by the KAN module. For the KAN network to output to the MoE module The processing results For one-dimensional convolutional layer pairs The processing results are all in dimensions. This module is used to refine complex non-linear dependency features across patches, further enhancing feature representation capabilities.

[0104] S2.3: Construct the prediction head submodule, which is a single-layer fully connected network with a weight matrix. Bias term Output predicted values ​​of battery operating parameters (such as remaining capacity and state of health, SOH):

[0105]

[0106] in, For the first Predicted battery operating parameters for each patch This is the refined feature vector output by the KAN module.

[0107] The entire MoKAN network model was trained using the Adam optimizer, with a learning rate set to [value missing]. The weight decays to The number of training rounds is The total loss function is ,in Mean square error, These are real-time measured values ​​of battery operating parameters. To determine the weight of the auxiliary loss in the total loss, a validation set is used for model performance evaluation during training, and an early stopping strategy is employed (validation set loss is continuous). If the number of wheels does not decrease, training should be stopped to ensure the model's generalization ability.

[0108] S3: Input the preprocessed feature sequence of the power battery to be detected into the trained MoKAN fault prediction model to obtain the predicted values ​​of the battery operating parameters. Calculate the relative error using the formula:

[0109]

[0110] in, These are the real-time measured values ​​of the operating parameters of the power battery to be tested. This represents the relative error.

[0111] Set a fault threshold, when the error When the fault threshold is exceeded, the power battery is deemed to have a fault risk. For fault prediction, data including normal operating conditions and some fault conditions are used as the training dataset. The MoKAN fault prediction model is trained using the error. The model hyperparameters are adjusted and the fault threshold is selected using the validation set. Finally, the test dataset is input into the trained model to predict faults. Based on the comparison between the error and the fault threshold, those exceeding the fault threshold are determined to have a fault risk.

[0112] This invention conducts comparative experiments on a publicly available dataset (NASA PCoE lithium-ion battery public dataset). The comparison method selects the mainstream deep learning models Transformer and LSTM, and the test results are as follows:

[0113] Table 1 Performance Test Table

[0114]

[0115] Among them, Precision (fault detection accuracy), Recall (recall rate), and F1-score (F1 score) are the core indicators for measuring model performance in power battery fault detection tasks: Precision reflects the accuracy of the model's detection results (i.e., the proportion of real fault samples among the samples predicted as faults), Recall reflects the model's ability to cover fault samples (i.e., the proportion of all real fault samples that are successfully detected), and F1-score is the harmonic mean of the two, which comprehensively reflects the overall detection performance of the model.

[0116] As can be seen from the comparison results in Table 1, the power battery fault prediction method disclosed in this embodiment is significantly better than the classic deep learning model (LSTM) and attention mechanism model (Transformer) in terms of the three core indicators of Precision, Recall and F1-score, which fully demonstrates that it has better detection accuracy and comprehensive performance in the power battery fault detection task.

[0117] This application also provides a power battery fault prediction system for executing the aforementioned power battery fault prediction method. The power battery fault prediction system includes a data acquisition module and a MoKAN fault prediction model. The data acquisition module is used to collect power battery operating data. Specifically, the data acquisition module is a sensor deployed at the power battery pack system end. The sensor model is not limited, as long as it meets the data acquisition requirements. The method for collecting power battery operating data involves obtaining voltage, current, temperature, and internal resistance data of the power battery under different operating conditions through the sensors. The MoKAN fault prediction model is used to output predicted values ​​of battery performance parameters based on a feature sequence generated from the preprocessed battery operating data. Finally, based on the error between the predicted value and the real-time measured value, it is determined whether the power battery has a fault risk. In this embodiment, a fault threshold can be set; when the error between the predicted value and the real-time measured value exceeds the fault threshold, it can be determined that the power battery has a fault risk.

[0118] Furthermore, the preprocessing includes channel normalization, Fourier adaptive patch partitioning, and patch and position embedding operations.

[0119] Specifically, the MoKAN fault prediction model comprises a hybrid expert module, a KAN submodule, and a prediction head submodule. The hybrid expert module contains several expert models and a gating network; each expert model employs a dual-path structure combining a KAN network and one-dimensional convolutions. The KAN module also uses a dual-path structure combining a KAN network and one-dimensional convolutions. The prediction head submodule uses a single-layer fully connected network.

[0120] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0121] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. For those skilled in the art, various modifications and improvements can be made without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the patent protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting power battery faults, characterized in that, Includes the following steps: S1: The power battery operation data is collected through the data acquisition module, and then the collected data is preprocessed to obtain the feature sequence for the adaptation model input; the power battery operation data is time series data, including the voltage, current, temperature and internal resistance of the power battery under different operating conditions; S2: Construct a MoKAN fault prediction model. The MoKAN fault prediction model uses the input feature sequence to output predicted values ​​of battery operating parameters. The MoKAN fault prediction model includes a hybrid expert module, a KAN module, and a prediction head sub-module. The construction of the MoKAN fault prediction model specifically includes the following steps: S2-1: Construct a hybrid expert module, which contains several expert models and a gating network. Each expert model adopts a dual-path structure that combines a KAN network and a one-dimensional convolution. The gating network dynamically selects one or more expert models to process the input features. S2-2: Construct a KAN module, which adopts a dual-path structure that combines KAN network and one-dimensional convolution, and refines the features output by the hybrid expert module. S2-3: Construct a prediction head submodule, which uses a single-layer fully connected network to output predicted values ​​of battery operating parameters based on refined features; S3: Based on the error between the predicted value and the real-time measurement value, determine whether there is a risk of failure in the power battery.

2. The power battery fault prediction method according to claim 1, characterized in that, The preprocessing in step S1 includes channel normalization, Fourier adaptive patch partitioning, and patch and position embedding.

3. The power battery fault prediction method according to claim 2, characterized in that, The method for preprocessing the data in step S1 is as follows: S1.1: Normalize the collected battery data samples; S1.2: Analyze the dominant period of the battery data through Fast Fourier Transform, determine the patch length based on the period, and divide the normalized long sequence into multiple semantically coherent short segments. S1.3: Map each patch to a fixed-dimensional feature space using simple linear projection, and then add location information.

4. The power battery fault prediction method according to claim 1, characterized in that, The specific steps for outputting the predicted values ​​of battery operating parameters in step S2 are as follows: S2.1: Feature extraction is performed using a hybrid expert module; S2.2: The extracted features are refined using the KAN module to obtain refined features; S2.3: The prediction head module outputs predicted values ​​of battery operating parameters based on refined features.

5. The power battery fault prediction method according to claim 1, characterized in that, In step S3, the method for determining whether the power battery has a fault risk based on the error between the predicted value and the real-time measured value specifically includes the following steps: calculating the error between the predicted value and the real-time measured value, setting a fault threshold, comparing the calculated error with the fault threshold, and determining that the power battery has a fault risk when the error exceeds the fault threshold.

6. The power battery fault prediction method according to claim 1, characterized in that, The data acquisition module is a sensor located at the end of the power battery pack system.

7. The power battery fault prediction method according to claim 6, characterized in that, The method for collecting power battery operating data in step S1 is as follows: the voltage, current, temperature, and internal resistance data of the power battery under different operating conditions are obtained through the sensor.

8. A power battery fault prediction system, characterized in that, A method for performing power battery fault prediction according to any one of claims 1-7 includes: The data acquisition module is used to collect operating data of the power battery; The MoKAN fault prediction model includes a hybrid expert module, a KAN submodule, and a prediction head submodule; it is used to output predicted values ​​of battery performance parameters based on feature sequences generated from preprocessed battery operating data.

9. A power battery fault prediction system according to claim 8, characterized in that, The preprocessing includes channel normalization, Fourier adaptive patch partitioning, and patch and position embedding operations; the hybrid expert module contains several expert models and a gating network, each expert model adopts a dual-path structure of KAN network and one-dimensional convolution in parallel; the KAN module adopts a dual-path structure of KAN network and one-dimensional convolution in parallel; the prediction head sub-module adopts a single-layer fully connected network.

Images