Power battery fault detection method based on bidirectional Mama architecture

Abstract

The invention discloses a power battery fault detection method based on a bidirectional Mama architecture, and the method comprises the steps: firstly collecting the voltage, current, temperature and internal resistance of a power battery as original data, and carrying out the preprocessing of the original data, and generating a data set; secondly, the bidirectional Mama feature extraction network and the weighted reconstruction network are connected in series to construct a bidirectional Mama network model, bidirectional representation modeling of the time dimension and the feature dimension is carried out on the power battery data in the data set, and reconstruction output is carried out through the weighted reconstruction network. And finally, taking the time sequence data of the power battery under the normal working condition as the input of the bidirectional Mamba network for training. And according to a reconstruction error between the output data and the input data of the weighted reconstruction network, a fault threshold is set, and fault detection is realized. The key information in the time sequence data of the power battery is accurately extracted, potential faults in the power battery are efficiently and comprehensively detected, and the safety of the power battery is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power battery management, and in particular to a power battery fault detection method based on a bidirectional Mamba architecture. Background Art

[0002] Power battery fault detection can be divided into two categories: model-based methods and data-driven methods. Model-based methods use the physical properties of the battery, such as voltage, current, temperature, and internal resistance, to construct mathematical models to deduce the normal operating state of the battery, thereby detecting whether a fault occurs. However, due to the complex working environment of the battery, it is difficult for the model to cover multiple types of operating conditions, so it is difficult to build a detection model with strong versatility and high accuracy. Data-driven methods use the historical data of power batteries to construct a fault feature representation model. Among them, deep learning methods are widely used due to their strong generalization ability and feature extraction performance. However, deep learning methods usually require a lot of computing resources for training and reasoning, which limits their widespread deployment on edge devices with limited computing resources. In addition, the physical relationship inside the power battery is complex, and there are complex physical associations between features, but traditional data-driven methods usually only model the time dimension of historical data, ignoring the correlation between feature dimensions, which limits the fault detection accuracy of the model. Summary of the invention

[0003] The present invention aims to overcome the shortcomings of the prior art and proposes a power battery fault detection method based on a bidirectional Mamba architecture. Through the selective scanning mechanism of the Mamba architecture, the key features in the power battery data are focused on, and redundant scanning and computational burdens are reduced, thereby improving fault detection efficiency and real-time response capabilities. The bidirectional Mamba architecture performs bidirectional modeling on the power battery time series data in the time dimension and feature dimension, effectively capturing complex time patterns and feature associations between time series, and achieving efficient and accurate power battery fault detection.

[0004] In order to solve the above technical problems, the technical solution of the present invention is:

[0005] A power battery fault detection method based on a bidirectional Mamba architecture comprises the following steps:

[0006] S1. Collect the raw data of the power battery such as voltage, current, temperature and internal resistance, and generate a data set for training and testing the fault detection model after preprocessing.

[0007] S2, construct a bidirectional Mamba network model, which is composed of a bidirectional Mamba feature extraction network and a weighted reconstruction network connected in series. The bidirectional Mamba feature extraction network is composed of time dimension and feature dimension Mamba modules in parallel order, and the information of input data in time and feature dimensions is modeled respectively. Each Mamba module includes a linear layer, a convolutional layer, a SiLU activation layer and a state space model (SSM) connected in sequence. The weighted reconstruction network is composed of a weighted fusion layer, a fully connected layer and a dimensionality reduction and dimensionality transformation layer in serial order.

[0008] S3. Through the bidirectional Mamba feature extraction network, the power battery data in the data set is modeled in both the time dimension and the feature dimension to generate features with rich semantic information. Then, the extracted power battery features are reconstructed and output through the weighted reconstruction network.

[0009] S4. Use the time series data of the voltage, current, temperature and internal resistance of the power battery under normal working conditions as the input of the bidirectional Mamba network for model training. Set the fault threshold according to the reconstruction error between the output and input data of the weighted reconstruction network to achieve fault detection.

[0010] Preferably, in step S1, the method for collecting the original data of the power battery is: using high-precision sensors to monitor the operating status of the power battery system under various working conditions in real time, and continuously collecting time series data of key parameters such as voltage, current, temperature and internal resistance through a sensor network.

[0011] Preferably, the preprocessing method of the power battery raw data collected in step S1 is: cleaning the collected power battery data samples, removing invalid data, and normalizing them so that each feature value is within the same scale range to improve the accuracy and reliability of model training. Then, using a sliding window operation, the normalized data samples are divided into segments of the same length according to the time dimension to obtain Where W and C represent the sliding window size and feature dimension of the data sample respectively.

[0012] Preferably, in step S1, the method for preparing the power battery data set is: organizing the preprocessed data into a data set, and dividing it into a training set, a validation set, and a test set in proportion. The training set only contains the power battery data under normal working conditions, which is used for the model to learn the power battery mode under normal working conditions; the validation set is used to adjust the model parameters and set the fault threshold; and the test set is used to evaluate the fault detection effect of the model.

[0013] Preferably, the specific process of step S3 is:

[0014] S3-1. After performing a time series data Patch segmentation operation on the input time series sample x in the time dimension, we can obtain Where N and P represent the number and length of patches respectively.

[0015] S3-2. For sample X p Perform Patch feature embedding operation to map low-dimensional feature embedding to higher dimensions and obtain the input sample of the bidirectional Mamba feature extraction network

[0016] z p =Embedding(X p )=TokenEmbedding(X p )+PositionalEmbedding(X p )

[0017] Among them, Embedding(·) represents the feature embedding operation, TokenEmbedding(·) represents the word embedding operation, which converts discrete time series data into continuous vector representation, and PositionalEmbedding(·) represents the position embedding operation, which adds position information to each time step in the sequence data.

[0018] S3-3. Input samples through bidirectional Mamba feature extraction network Feature extraction is performed on the time dimension and feature dimension respectively.

[0019] Preferably, the step S3-3 includes the following sub-steps:

[0020] S3-3-1. For the time dimension, input samples Input into the time dimension Mamba module to model the time dimension representation and obtain the time dimension representation For the feature dimension, the input sample Perform dimension transformation to obtain Then input it into the feature dimension Mamba module to model the feature dimension representation and obtain the feature dimension representation Among them, the input sample z in the time dimension p For example, the representation modeling process of the Mamba module is:

[0021] First, input sample z p After two linear layer mappings, feature z1 and feature z2 are obtained respectively. Subsequently, the convolution layer and SiLU activation function process feature z1 to obtain feature z1′. After feature z2 is activated by SiLU, feature z′2 is obtained. Feature z1′ generates feature Then multiply it by element by feature z′2 to get feature z t . Finally, z t After linear layer mapping, the representation modeling result x of the time dimension Mamba module is obtained t ′. The implementation process of the state space model is as follows:

[0022] First, define the mapping process of the continuous state space model for input signals and output signals:

[0023] h′(k)=Ah(k)+Bx(k)

[0024] y(k)=Ch(k)+Dx(k)

[0025] Among them, h(k) represents the hidden state function, x(k) represents the input function, y(k) represents the output function, A, B, C and D represent learnable parameter matrices used for mapping between the hidden state function, input function and output function.

[0026] Secondly, the continuous state space model is discretized by using the zero-order hold method, so that it can use discrete time series data for feature learning and modeling in a recursive form during training. The specific implementation process of the mapping process between the input signal and the output signal of the discretized state space model is as follows:

[0027]

[0028] Where Δ represents the step size, which is the time interval used to sample a continuous signal to generate a discrete signal, and determines the choice of sampling points. Δ are all learnable parameters, which are obtained based on the linear mapping of the input and can be dynamically adjusted according to the contextual semantics of the input time series data, thereby realizing the selective scanning mechanism of the state space model. In order to improve efficiency, the residual connection mapping matrix used for input to output is is 0.

[0029] S3-3-2. Output of the Mamba module in the time dimension Residual connections are made to enhance the gradient fluidity of the model, and layer normalization operations are applied to improve the stability and convergence speed of the output to obtain the time representation At the same time, the feature dimension Mamba module output After dimensional transformation, residual connection and layer normalization are performed to obtain feature representation The expression is as follows:

[0030] x t =LayerNorm[Add(z p ,x′ t )]

[0031] x f =LayerNorm[Add(z p ,Transpose(x′ f ,0,1))]

[0032] Among them, LayerNorm(·) represents the layer normalization operation, Add(·) represents the residual connection operation, and Transpose(x′ f ,0,1) represents the dimension conversion operation, which converts the feature dimension into x′ f The first dimension is swapped with the second dimension.

[0033] Preferably, the step S4 includes the following sub-steps:

[0034] S4-1. Time representation of the output of the bidirectional Mamba feature extraction network x t and feature representation x f Use learnable weights to fuse and get a bidirectional representation The expression is as follows:

[0035] x dual =W t x t +W f x f

[0036] in, Represents the learnable weight value.

[0037] S4-2, after weighted fusion, the bidirectional representation x is transformed through the fully connected layer dual Convert to Patch block timing samples Then, the final weighted reconstructed data is obtained through dimensionality reduction and dimensionality transformation operations. The expression is as follows:

[0038] x dp =MLP(x dual )=Wx dual +b

[0039]

[0040] Among them, MLP(·) represents a fully connected layer, W represents a learnable weight matrix, and b represents a learnable bias vector. Flatten(·,1) represents a dimensionality reduction operation, merging the second and third dimensions. Transpose(·,0,1) represents a dimensionality conversion operation, exchanging the first and second dimensions.

[0041] S4-3. Calculate the reconstruction error based on the input data and output reconstruction:

[0042]

[0043] in, represents the reconstruction error, Represents input and output data;

[0044] The fault threshold is determined by calculating the reconstruction error of the validation set data; by comparing the reconstruction error with the fault threshold, data samples that exceed the threshold will be judged as faults.

[0045] Preferably, the bidirectional Mamba network model constructed in step S2 has its hyperparameters determined by a cross-validation method, specifically including sliding window size, patch size, embedding dimension size and batch size.

[0046] The present invention has the following characteristics and beneficial effects:

[0047] Adopting the above technical solution, the present invention adopts a bidirectional Mamba feature extraction network to perform bidirectional feature representation and modeling on the time dimension and feature dimension of the time series data such as voltage, current, temperature and internal resistance of the power battery. At the same time, combined with the weighted reconstruction network, accurate reconstruction of the data is achieved. Through the efficient Mamba architecture, the deployment problem of deep learning methods on the edge platform can be effectively solved, and its selective mechanism can be used to accurately extract key information from the power battery time series data. A more comprehensive power battery feature representation is constructed through bidirectional modeling of time and features, which helps to efficiently and comprehensively detect potential faults in the power battery and improve the safety of the power battery. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0049] Figure 1 A power battery fault detection method based on a bidirectional Mamba architecture according to an embodiment of the present invention;

[0050] Figure 2 A schematic diagram of a bidirectional Mamba network structure in an embodiment of the present invention;

[0051] Figure 3 It is a schematic diagram of the Mamba architecture structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention provides a power battery fault detection method based on a bidirectional Mamba architecture, such as Figure 1 As shown, the following steps are included:

[0053] Step (1): Power battery raw data collection and preprocessing.

[0054] High-precision sensors are used to monitor the operating status of the power battery system in real time under various working conditions. Through the sensor network, the time series data of key parameters such as battery voltage, current, temperature and internal resistance are continuously collected. The collected power battery data samples are preprocessed to filter out invalid and noisy data to ensure the accuracy and integrity of the data and obtain the cleaned power battery raw data samples. Among them, L and C represent the length and number of features of the sample respectively. Then, Normalization is performed to make each eigenvalue within the same scale range, thus improving the accuracy and reliability of model training. The expression is as follows:

[0055]

[0056] in, represents the data sample after normalization, min(·) and max(·) represent the data sample X battery The minimum and maximum values ​​in .

[0057] Using the sliding window operation, the normalized data samples are divided into segments of the same length according to the time dimension to obtain

[0058] Step (2) extracts power battery features based on the bidirectional Mamba feature extraction network, such as Figure 2 As shown in Figure 1, the bidirectional Mamba feature extraction network consists of a time dimension Mamba module and a feature dimension Mamba module in parallel order, which respectively model the information of the input data after patch segmentation in the time and feature dimensions. Figure 3 As shown in the figure, the Mamba module contains linear layers, convolutional layers, SiLU activation layers, and state space models (SSM). Specifically:

[0059] First, the input time series sample X is segmented into time series data patches in the time dimension to obtain Where N and P represent the number and length of patches respectively. Dividing time series data into patches can effectively capture local features and improve the computational efficiency and robustness of the model.

[0060] Secondly, for the sample X pPerform Patch feature embedding operation to map low-dimensional feature embedding to higher dimensions, thereby capturing more nonlinear relationships and complex patterns, and obtaining the input samples of the bidirectional Mamba feature extraction network.

[0061] Then, the input sample is extracted through the bidirectional Mamba feature extraction network. Feature extraction is performed on the time dimension and feature dimension respectively.

[0062] For the time dimension, the input sample Input into the time dimension Mamba module to model the representation in the time dimension and obtain the time dimension representation For the feature dimension, the input sample Perform dimension transformation to obtain Then input it into the feature dimension Mamba module to model the representation on the feature dimension and obtain the feature dimension representation Bidirectional modeling of the time dimension and feature dimension of power battery time series data can fully capture the dynamic relationship and global time dependency between various features and improve the accuracy of fault detection. p For example, the representation modeling process of the Mamba module is:

[0063] First, input sample z p After two linear layer mappings, feature z1 and feature z2 are obtained respectively:

[0064] z1=Linear(z p ),z2=Linear(z p )

[0065] Subsequently, the convolutional layer and SiLU activation function process feature z1 to obtain feature z1′. After feature z2 is activated by SiLU, feature z′2 is obtained:

[0066]

[0067] Feature z1′ generates features through the state space model And multiply it by feature z′2 element by element to get feature z t :

[0068]

[0069] Finally, t After linear layer mapping, the output time dimension Mamba module representation modeling result x t ′:

[0070] x t ′=Linear(zt )

[0071] Among them, Linear(·) represents the linear mapping function, ConV(·) represents the convolution function, SiLU(·) and Sigmoid(·) represent the activation function, Represents the element-by-element multiplication operation. SSM(·) stands for the state space model, which is a framework for optimizing time series data modeling by dynamically selecting relevant states, improving the ability to capture dynamic changes in complex systems. The state space model implementation process is as follows:

[0072] First, define the mapping process of the continuous state space model for input signals and output signals:

[0073] h′(k)=Ah(k)+Bx(k)

[0074] y(k)=Ch(k)+Dx(k)

[0075] Among them, h(k) represents the hidden state function, x(k) represents the input function, y(k) represents the output function, and A, B, C, and D represent learnable parameter matrices used for mapping between the hidden state function, the input function, and the output function.

[0076] Secondly, the continuous state space model is discretized by using the zero-order hold method, so that it can use discrete time series data for feature learning and modeling in a recursive form during training. The specific implementation process of the mapping process between the input signal and the output signal of the discretized state space model is as follows:

[0077]

[0078]

[0079] Discretized state space models help improve the training efficiency of the model, reduce computational complexity, and adapt to multiple types of discrete data in practical applications. Where Δ represents the step size, that is, the time interval used to sample continuous signals to generate discrete signals, which determines the selection of sampling points. Parameters Δ are all learnable parameters, which are obtained based on the linear mapping of the input and can be dynamically adjusted according to the contextual semantics of the input time series data, thereby realizing the selective scanning mechanism of the state space model. In order to improve efficiency, the residual connection mapping matrix used for input to output is is 0.

[0080] Mamba module output for the time dimension Perform residual connections and layer normalization to obtain temporal representation At the same time, the feature dimension Mamba module output After dimensional transformation, residual connection and layer normalization are performed to obtain feature representation The expression is as follows:

[0081] x t =LayerNorm[Add(z p ,x′ t )]

[0082] x f =LayerNorm[Add(z p ,Transpose(x′ f ,0,1))]

[0083] Among them, LayerNorm(·) represents the layer normalization operation, Add(·) represents the residual connection operation, and Transpose(·,0,1) represents the dimension conversion operation, which represents the feature dimension x′ f The first dimension is swapped with the second dimension.

[0084] Step (3): Reconstruct the extracted bidirectional feature representation based on the weighted reconstruction network. The weighted reconstruction network is composed of a weighted fusion layer, a fully connected layer, and a dimensionality reduction and dimensionality transformation layer in a serial order. Specifically:

[0085] First, the temporal representation x of the output of the bidirectional Mamba feature extraction network is t and feature representation x f Use learnable weights to fuse and get a bidirectional representation The expression is as follows:

[0086] x dual =W t x t +W f x f

[0087] in, Represents a learnable weight value. By weighted fusion of the bidirectional representation of time and feature dimensions, the accuracy of the model can be improved, the feature expression ability can be optimized, and noise interference can be reduced.

[0088] After weighted fusion, the bidirectional representation x is transformed through the fully connected layer. dual Convert to Patch Block Sample Then, the final weighted reconstructed data is obtained through dimensionality reduction and dimensionality transformation operations. The expression is as follows:

[0089] x dp =MLP(x dual )=Wx dual +b

[0090]

[0091] Among them, MLP(·) represents a fully connected layer, W represents a learnable weight matrix, and b represents a learnable bias vector. Flatten(·,1) represents a dimensionality reduction operation, merging the second and third dimensions. Transpose(·,0,1) represents a dimensionality conversion operation, exchanging the first and second dimensions.

[0092] Step (4) A power battery fault detection method based on a bidirectional Mamba architecture.

[0093] The preprocessed data is organized into a data set and divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training set only contains the power battery data under normal working conditions, which is used for the model to learn the power battery mode under normal working conditions; the validation set is used to adjust the model parameters and set the fault threshold; the test set is used to evaluate the effect of fault detection.

[0094] During the fault detection process, the voltage, current, temperature and internal resistance data of the power battery are selected as the input data of the model, and the output is reconstructed through the bidirectional Mamba network. The reconstruction error is calculated based on the input data and output reconstruction:

[0095]

[0096] in, represents the reconstruction error, Represents input and output data. During local training, a training set containing only power battery data under normal conditions is used to train the bidirectional Mamba network through reconstruction error. Subsequently, the validation set is used to determine the threshold for fault detection. After training, the model and the selected fault threshold are uploaded to the cloud server, and fault detection is performed on data that may contain anomalies. By comparing the reconstruction error with the fault threshold, data that exceeds the threshold will be judged as a fault.

[0097] Step (5) The present invention conducts a comparative experiment on multivariate time series data fault detection on the public PSM (Pooled Server Metrics) dataset and compares the classic deep learning method Transformer. The test results are shown in Table 1:

[0098] Table 1

[0099] Detection Model Precision(%) Recall(%) F-score(%) Floating point operations (M) Model parameters (K) Transformer 69.46 81.27 74.90 173.99 2403.12 The present invention 98.52 95.74 97.11 2.13 97.85

[0100] Among them, Precision, Recall and F-score represent the accuracy, recall rate and F1 score of fault detection respectively, reflecting the performance of the model in fault detection. Floating-point operations and model parameters measure the computational complexity and parameter scale of the model, reflecting its computational efficiency and storage requirements. As can be seen from Table 1, the bidirectional Mamba architecture proposed in the present invention shows high accuracy and excellent performance in the task of power battery fault detection. At the same time, compared with Transformer, it greatly reduces the number of floating-point operations and model parameters, improves the inference speed of cloud detection, reduces resource consumption, and is more suitable for scenarios with limited computing resources.

[0101] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A power battery fault detection method based on a bidirectional Mamba architecture, characterized in that: The steps include: S1. Collect the voltage, current, temperature and internal resistance of the power battery as raw data, and perform preprocessing to generate a data set; S2, construct a bidirectional Mamba network model, which consists of a bidirectional Mamba feature extraction network and a weighted reconstruction network connected in series; S3, through the bidirectional Mamba feature extraction network, the power battery data in the data set is modeled in the bidirectional representation of the time dimension and the feature dimension, and then reconstructed and outputted through the weighted reconstruction network; S4. Use the time series data of the voltage, current, temperature and internal resistance of the power battery under normal working conditions as the input of the bidirectional Mamba network for training; set the fault threshold according to the reconstruction error between the output and input data of the weighted reconstruction network to realize fault detection.

2. The power battery fault detection method based on the bidirectional Mamba architecture according to claim 1 is characterized in that: The preprocessing in step S1 is specifically as follows: cleaning the collected raw data of the power battery, removing invalid data, and normalizing it; then, using a sliding window operation, dividing the normalized data samples into segments of the same length according to the time dimension to obtain Where W and C represent the sliding window size and feature dimension of the data sample respectively; The specific process of generating the data set is: organizing the preprocessed data into a data set, and dividing it into a training set, a validation set and a test set in proportion; wherein the training set only contains the power battery data under normal working conditions, and is used to learn the power battery mode under normal working conditions; the validation set is used to adjust the model parameters and set the fault threshold; and the test set is used to evaluate the fault detection effect.

3. The power battery fault detection method based on the bidirectional Mamba architecture according to claim 2 is characterized in that: The bidirectional Mamba feature extraction network is composed of time dimension and feature dimension Mamba modules in parallel order, and models the information of input data in time and feature dimensions respectively; each Mamba module includes a linear layer, a convolutional layer, a SiLU activation layer and a state space model SSM connected in sequence; the weighted reconstruction network is composed of a weighted fusion layer, a fully connected layer and a dimensionality reduction and dimensionality transformation layer in serial order.

4. The power battery fault detection method based on the bidirectional Mamba architecture according to claim 3 is characterized in that: The specific implementation process of step S3 is as follows: S3-1. After performing a time series data Patch segmentation operation on the input time series sample x in the time dimension, we can obtain Among them, N and P represent the number and length of patches respectively; S3-2. For sample X p Perform Patch feature embedding operation to obtain the input sample of the bidirectional Mamba feature extraction network z p =Embedding(X p )=TokenEmbedding(X p )+PositionalEmbedding(X p ) Among them, Embedding(·) represents the feature embedding operation, TokenEmbedding(·) represents the word embedding operation, which converts discrete time series data into continuous vector representation, and PositionalEmbedding(·) represents the position embedding operation, which adds position information to each time step in the sequence data. S3-3. Input samples through bidirectional Mamba feature extraction network Feature extraction is performed on the time dimension and feature dimension respectively.

5. The power battery fault detection method based on the bidirectional Mamba architecture according to claim 4 is characterized in that: The step S3-3 includes the following sub-steps: S3-3-1. For the time dimension, input samples Input into the time dimension Mamba module to model the time dimension representation and obtain the time dimension representation For the feature dimension, the input sample Perform dimension transformation to obtain Then input it into the feature dimension Mamba module to model the feature dimension representation and obtain the feature dimension representation Input sample z in the time dimension p , the representation modeling process of the Mamba module is: First, input sample z p After two linear layer mappings, feature z1 and feature z2 are obtained respectively; then, the convolution layer and SiLU activation function process feature z1 to obtain feature z1′; feature z2 is activated by SiLU to obtain feature z′2; feature z1′ generates feature Then multiply it by element by feature z′2 to get feature z t ; Finally, z t After linear layer mapping, the representation modeling result x of the time dimension Mamba module is obtained t ′; S3-3-2. Output of the Mamba module in the time dimension Perform residual connection and apply layer normalization operation to obtain time representation At the same time, the feature dimension Mamba module output After dimensional transformation, residual connection and layer normalization are performed to obtain feature representation The expression is as follows: x t =LayerNorm[Add(z p ,x′ t )] x f =LayerNorm[Add(z p ,Transpose(x′ f ,0,1))] Among them, LayerNorm(·) represents the layer normalization operation, Add(·) represents the residual connection operation, and Transpose(x′ f ,0,1) represents a dimension conversion operation, exchanging the first dimension and the second dimension of the feature dimension representation x′f.

6. The power battery fault detection method based on the bidirectional Mamba architecture according to claim 5 is characterized in that: The state space model implementation process is as follows: First, define the mapping process of the continuous state space model for input signals and output signals: h′(k)=Ah(k)+Bx(k) y(k)=Ch(k)+Dx(k) Among them, h(k) represents the hidden state function, x(k) represents the input function, y(k) represents the output function, and A, B, C and D represent the learnable parameter matrices; Secondly, the continuous state space model is discretized by using the zero-order hold method. The specific implementation process of the mapping process between the input signal and the output signal of the discretized state space model is as follows: Among them, Δ represents the step size, that is, the time interval used to sample the continuous signal to generate a discrete signal, which determines the selection of the sampling point; the parameter and Δ are both learnable parameters, obtained based on the linear mapping of the input, and dynamically adjusted according to the contextual semantics of the input time series data; in order to improve efficiency, the residual connection mapping matrix used for input to output is 0.

7. The power battery fault detection method based on the bidirectional Mamba architecture according to claim 6 is characterized in that: The specific implementation process of step S4 is as follows: S4-1. Time representation of the output of the bidirectional Mamba feature extraction network x t and feature representation x f Use learnable weights to fuse and get a bidirectional representation The expression is as follows: x dual =W t x t +W f x f in, Represents the learnable weight value; S4-2, after weighted fusion, the bidirectional representation x is transformed through the fully connected layer dual Convert to Patch block timing samples Then, the final weighted reconstructed data is obtained through dimensionality reduction and dimensionality transformation operations. The expression is as follows: x dp =MLP(x dual )=Wx dual +b Among them, MLP(·) represents the fully connected layer, W represents the learnable weight matrix, b represents the learnable bias vector, and Flatten(·,1) represents the dimensionality reduction operation, merging the second and third dimensions; S4-3. Calculate the reconstruction error based on the input data and output reconstruction: in, represents the reconstruction error, Represents input and output data; The fault threshold is determined by calculating the reconstruction error of the validation set data; by comparing the reconstruction error with the fault threshold, data samples that exceed the threshold will be judged as faults.

Images