A Power Battery Fault Detection Method Based on Frequency Domain Strength-Weakness Contrast Learning Mamba

Abstract

A power battery fault detection method based on frequency domain contrastive learning (Mamba) includes the following steps: S1, collecting and preprocessing raw power battery data to generate time-series samples; S2, extracting frequency domain feature representations of the power battery using the constructed frequency domain contrastive learning Mamba model based on the time-series samples; S3, training the model by calculating the time-frequency reconstruction loss and contrastive learning loss of the frequency domain contrastive learning Mamba model; S4, using the trained frequency domain contrastive learning Mamba model to detect power battery faults. This invention integrates the contrastive learning mechanism with the Mamba model, constructing frequency domain strong and weak samples to guide the model in learning more discriminative features, thereby achieving efficient and accurate capture of dynamic features while maintaining low computational complexity. This enables efficient and accurate power battery fault detection under low computing power conditions.

Description

Technical Field

[0001] This invention relates to the field of power battery management technology, specifically to a power battery fault detection method based on frequency domain strength contrast learning Mamba. Background Technology

[0002] During the long-term operation of new energy vehicles, the performance of power batteries gradually degrades, often accompanied by increased internal resistance, prolonged charging time, and unstable discharge voltage. These degradation processes typically exhibit complex patterns. As the degradation deepens, the power battery pack may experience fault conditions during operation, leading to safety risks such as thermal runaway, insulation failure, and electrical short circuits. Therefore, timely and accurate identification of the operating mode characteristics of power batteries and fault detection are crucial for ensuring vehicle safety and extending battery life.

[0003] Currently, power battery fault detection methods are mainly divided into two categories: mechanistic modeling-based methods and data-driven methods. Mechanism-based modeling methods identify possible fault types by establishing mathematical models of the battery's electrochemical processes. However, due to the complexity of internal battery reactions and the variability of operating conditions, the model accuracy is easily affected by time and environmental changes, making it difficult to accurately describe nonlinear dynamic characteristics. Data-driven methods utilize a large amount of historical operating data for feature extraction and pattern recognition, exhibiting good generalization ability and fault detection performance. However, existing data-driven methods mostly rely on time-domain signals, resulting in limited accuracy in pattern-level fault identification and difficulty in comprehensively revealing the potential degradation patterns of the battery. Furthermore, the model structures are generally complex and computationally intensive, making it difficult to meet the real-time online monitoring needs of vehicle-mounted edge devices. Compared to time-domain analysis, frequency-domain features can more effectively reveal the potential periodicity and energy distribution information in the signal, helping to capture the differences between degradation modes.

[0004] Therefore, there is an urgent need for a power battery fault detection method that can fully exploit frequency domain features and has the advantages of low computing power and high precision, so as to ensure the safe and efficient operation of new energy vehicles. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and proposes a power battery fault detection method based on frequency domain strength-weak contrastive learning (Mamba). First, the time-series data of the power battery is converted to the frequency domain, leveraging its characteristics to enhance the ability to identify mode-level faults. Then, the contrastive learning mechanism is integrated with the Mamba model, constructing frequency domain strength-weak samples to guide the model in learning more discriminative features, thereby achieving efficient and accurate capture of dynamic features while maintaining low computational complexity. This invention can achieve efficient and accurate power battery fault detection under low computational power conditions.

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

[0007] A power battery fault detection method based on frequency domain strength-weak contrast learning Mamba includes the following steps:

[0008] S1. Collect and preprocess the raw data of the power battery to generate time series samples;

[0009] S2. Based on the time-series samples, extract the frequency domain feature representation of the power battery by constructing a frequency domain strength contrast learning Mamba model;

[0010] S3. Train the model by calculating the time-frequency reconstruction loss and the strength-weak contrast learning loss of the frequency domain strength-weak contrast learning Mamba model:

[0011] S4. Use the trained frequency domain strength contrast learning Mamba model to detect power battery faults.

[0012] Preferably, in S1, the preprocessing includes cleaning the original data of the power battery to remove noise points and invalid data; it also includes normalizing the cleaned valid data to make different feature values ​​fall within a uniform dimension range; and using a sliding window method to divide the normalized data into sequence segments of equal length according to time order to obtain the data sample.

[0013] Preferably, S2 includes:

[0014] S21. Construct a frequency domain strong-weak data augmentation layer, a frequency domain Mamba feature extraction network, and a time-frequency reconstruction network to build a frequency domain strong-weak contrastive learning Mamba model.

[0015] S22. Input the time series samples into the strong and weak data enhancement layer to generate strong sample pairs and weak sample pairs;

[0016] S23. The frequency domain Mamba feature extraction network performs Fourier transform operations on the time-series samples, the strong sample pairs, and the weak sample pairs respectively to obtain their respective frequency domain samples; after feature embedding and mapping each frequency domain sample to a higher dimension, frequency domain representation modeling is performed to generate frequency domain feature representation.

[0017] S24. Input the frequency domain feature representation into the time-frequency reconstruction network to perform frequency domain reconstruction and generate a time-frequency domain reconstruction output.

[0018] Preferably, the strong and weak data augmentation layer includes a parallel strong data augmentation module and a weak data augmentation module; in S22, the strong data augmentation module performs a noise injection operation on the time series samples to generate the strong sample pairs; the weak data augmentation module performs a motion smoothing operation on the time series samples to generate the weak sample pairs.

[0019] Preferably, the frequency domain Mamba feature extraction network includes a frequency domain transformation layer and a Mamba module; in S23, the frequency domain transformation layer performs Fourier transform operations on the time series samples, the strong sample pairs, and the weak sample pairs respectively to generate their respective corresponding frequency domain samples; the frequency domain samples are respectively input into the Mamba module for frequency domain representation modeling to obtain the frequency domain feature representations of the time series samples, the frequency domain feature representations of the strong sample pairs, and the frequency domain feature representations of the weak sample pairs.

[0020] Preferably, the Mamba module sequentially includes a linear layer, a convolutional layer, a SiLU activation layer, and a state-space model, used to extract the dependencies and dynamic relationships of frequency domain features; after feature embedding and mapping of each frequency domain sample to a higher dimension, the input samples of the Mamba module are generated; the input samples are sequentially input into two independent parallel linear layers for feature mapping, obtaining feature representation one and feature representation two respectively; the convolutional layer and the SiLU activation function perform nonlinear processing on feature representation one to obtain an intermediate feature representation; the intermediate feature representation is modeled by the state-space model to generate state-space features; feature representation two is activated by SiLU to obtain activation features; the state-space features and activation features are multiplied element-wise and then mapped by the linear layer to output the frequency domain feature representation.

[0021] Preferably, the time-frequency reconstruction network includes a fully connected layer and a time-domain restoration layer in sequence; in S24, the frequency domain feature representation is reconstructed in the frequency domain through the fully connected layer to obtain the frequency domain reconstruction output; the frequency domain reconstruction output is converted into the time domain reconstruction output by inverse Fourier transform through the time-domain restoration layer.

[0022] Preferably, S3 includes:

[0023] S31. Calculate the contrastive learning loss of the frequency domain contrastive learning Mamba model.

[0024] Strong-weak contrast learning is performed on each of the frequency domain feature representations. The strong-weak contrast learning loss is obtained by maximizing the similarity between the frequency domain feature representation of the time series sample and the frequency domain feature representation of the strong sample pair and the frequency domain feature representation of the weak sample pair.

[0025] S32. Calculate the time-frequency reconstruction loss of the frequency domain strong-weak contrast learning Mamba model; the time-frequency reconstruction loss includes time domain reconstruction loss and frequency domain reconstruction loss; use the sum of the time-frequency reconstruction loss and the strong-weak contrast learning loss as the total training loss to train the model.

[0026] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0027] 1. Unlike traditional data-driven methods that directly supervise learning on raw features, this invention employs a fusion of frequency domain strength-weak contrastive learning and the Mamba model. By constructing strong and weak samples to guide the model in learning discriminative frequency domain representations, and combining this with Mamba's advantages in efficient sequence modeling, efficient fault detection under low computational requirements is achieved. This approach improves the model's ability to capture dynamic features while maintaining real-time performance, exhibiting superior generalization performance and engineering deployability.

[0028] 2. Unlike traditional fault identification techniques that rely solely on time-domain features, this invention employs a combination of frequency-domain feature extraction and modeling. By converting the time-series signals of the power battery, such as voltage, current, and temperature, into the frequency domain, it fully exploits the energy distribution and periodicity characteristics of the signals, enabling the model to identify potential differences in degradation modes. This approach significantly improves the accuracy and robustness of fault detection, especially in capturing early degradation characteristics more accurately under complex operating conditions.

[0029] By employing the above technical solution, this invention utilizes a frequency-domain contrastive learning Mamba model to perform frequency-domain modeling and feature extraction on multi-source time-series data such as internal resistance, voltage, current, and temperature of the power battery. This effectively captures key spectral features during battery operation. By introducing a contrastive learning mechanism, the model's discriminative power and robustness are improved through feature comparison of different enhanced samples, ensuring stable detection performance even under noise interference or changing operating conditions. The lightweight Mamba architecture design significantly reduces the number of model parameters and computational complexity, enhancing real-time deployment capabilities at edge terminals. This enables efficient and accurate detection of mode-level faults in power batteries, ensuring system operational safety. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the present invention;

[0032] Figure 2 This is a schematic diagram of the frequency domain strength contrast learning Mamba model structure in Embodiment 1 of the present invention. Detailed Implementation

[0033] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.

[0034] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0035] Example 1:

[0036] like Figure 1 The method for detecting power battery faults based on frequency domain strength contrast learning Mamba, as shown, includes the following steps:

[0037] S1. Collect and preprocess the raw data of the power battery to generate time series samples;

[0038] High-precision sensors are used to monitor the operating status of the power battery system under different working conditions in real time, continuously collecting time-series data of key parameters such as battery internal resistance, voltage, current, and temperature. The collected data is preprocessed to remove invalid or noisy samples to ensure accuracy and completeness, resulting in cleaned raw data samples. ,in, and These represent the sample length and the number of features, respectively. Then, for the original data samples... Normalization is performed to bring all feature values ​​to a uniform scale, thereby improving the accuracy and stability of model training. The expression is as follows:

[0039]

[0040] in, This represents the normalized data sample. and Representing data samples respectively The minimum and maximum values ​​in.

[0041] To accommodate subsequent modeling and feature extraction, a sliding window operation is used to divide the normalized samples into segments of uniform length along the time dimension, resulting in time-series samples that can be used as model input. ,in and These represent the sliding window size and feature dimension of the time series samples, respectively.

[0042] S2, such as Figure 2 As shown, based on time-series samples, the frequency domain feature representation of the power battery is extracted by learning the Mamba model through frequency domain strength contrast.

[0043] S21. Construct a frequency domain strong-weak data augmentation layer, a frequency domain Mamba feature extraction network, and a time-frequency reconstruction network to build a frequency domain strong-weak contrastive learning Mamba model.

[0044] The strong and weak data augmentation layer consists of a strong data augmentation module and a weak data augmentation module in parallel. The strong data augmentation module enhances the robustness of features by adding perturbation noise to time-series samples. The weak data augmentation module generates contrast samples through smoothing or denoising operations, thus forming strong and weak sample pairs in the frequency domain.

[0045] The frequency domain Mamba feature extraction network consists of a frequency domain transformation layer and a Mamba module. The frequency domain transformation layer is used to convert time-series signals into frequency domain feature representations via Fast Fourier Transform. The Mamba module is composed of linear layers, convolutional layers, SiLU activation layers, and a state-space model, and is used to extract the dependencies and dynamic relationships of frequency domain features.

[0046] The time-frequency reconstruction network consists of a fully connected layer and a time-domain restoration layer, which are used to restore the frequency domain features to the time domain through inverse transformation, thereby realizing time-frequency mapping and feature reconstruction.

[0047] S22. Input the time series samples into the strong and weak data augmentation layer to generate strong sample pairs and weak sample pairs;

[0048] For the input time series samples Strong sample pairs are obtained after performing strong and weak data augmentation operations respectively. weak sample pairs :

[0049]

[0050]

[0051] in, This represents a noise injection operation. This represents the moving smoothing operation. Noise injection, by superimposing random perturbations onto the original signal, enhances the model's robustness to noise and perturbations, enabling it to maintain stable feature extraction capabilities even in the face of signal fluctuations or measurement errors. Moving smoothing, by applying a locally weighted average to the sample sequence over time, effectively weakens the impact of instantaneous fluctuations and high-frequency noise, thus preserving the sequence's main trend features and global dynamic information. The combination of strong and weak enhancements helps the model capture more robust and discriminative temporal feature representations during contrastive learning.

[0052] S23. The frequency domain Mamba feature extraction network performs Fourier transform operations on time-series samples, strong sample pairs, and weak sample pairs respectively to obtain their respective frequency domain samples; after feature embedding and mapping of each frequency domain sample to a higher dimension, frequency domain representation modeling is performed to generate frequency domain feature representation.

[0053] S231, Frequency Domain Conversion Layer

[0054] For time series samples Strong sample pairs weak sample pairs Perform Fourier transform operations on each sample to convert them to the frequency domain, and obtain their respective frequency domain samples. This enhances the pattern-level features in the sequence. Subsequently, the low-dimensional frequency domain samples are feature-embedded and mapped to a higher dimension to obtain the input samples for the Mamba module. :

[0055]

[0056]

[0057] in, This represents the Fourier transform operation, which converts the original time-domain signal into a frequency-domain representation that reflects the energy distribution and periodic characteristics. The representative feature embedding operation is used to map the input frequency domain samples to the feature space that the model can learn, thereby achieving an effective representation of frequency domain information.

[0058] S232, Mamba module

[0059] Input sample The samples are input into the Mamba module for frequency domain representation modeling, resulting in frequency domain feature representations of time series samples, strong sample pairs, and weak sample pairs. Among them, the input sample For example, the representation modeling process of Mamba modules is as follows:

[0060] The input samples are sequentially fed into two linear layers for feature mapping to obtain feature representations. and Subsequently, the convolutional layer and the SiLU activation function are applied to the features. Nonlinear processing is performed to obtain intermediate feature representations. Next, features Modeling is performed using a state-space model to generate features. .feature After SiLU activation, the features are obtained. Then features With features Features are obtained by element-wise multiplication. .final, After linear layer mapping, the output frequency domain feature representation modeling result is obtained. This provides high-quality feature representation for subsequent fault detection.

[0061] The specific implementation process is as follows:

[0062] First, input sample After two independent and parallel linear layer mappings, the features are obtained respectively. and characteristics :

[0063]

[0064] Subsequently, convolutional layers and the SiLU activation function are applied to the features. Processing is performed to obtain features. .feature After SiLU activation, the features are obtained. :

[0065]

[0066] feature Features are generated using a state-space model. and features Features are obtained by element-wise multiplication. :

[0067]

[0068]

[0069] final, After linear layer mapping, the representation modeling results of the Mamba module are output. :

[0070]

[0071] in, Represents a linear mapping function. Represents the convolution function. and Represents the activation function. This represents element-wise multiplication. The representative state-space model achieves accurate capture of mode-level features by modeling and dynamically updating the implicit state of the battery frequency domain signal, thereby improving the model's ability to perceive and model dynamic changes in battery features.

[0072] The specific implementation process of the state-space model is as follows:

[0073] Define the mapping process between input and output signals in a continuous state-space model:

[0074]

[0075]

[0076] in, Represents the hidden state function. Represents the input function. Represents the output function. This represents the learnable parameter matrix, used for mapping between the hidden state function, input function, and output function.

[0077] By discretizing the continuous state-space model using the zero-order hold method, the model can recursively learn features and dynamically model discrete battery data. This discretization process effectively preserves the frequency domain response characteristics of the power battery signal, making the model more stable and accurate in capturing the dynamic evolution of key features such as internal resistance, voltage, current, and temperature. The mapping relationship between the input and output signals of the discretized state-space model is as follows:

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] Discretized state-space models help improve training efficiency, reduce computational complexity, and adapt to various types of discrete data in practical applications. The step size represents the time interval used to sample continuous signals to generate discrete signals, and it determines the selection of sampling points. (Parameter) , , All parameters are learnable, obtained based on a 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.

[0084] S24. Input the frequency domain feature representation into the time-frequency reconstruction network to perform frequency domain reconstruction and generate the time-frequency domain reconstruction output.

[0085] Frequency domain feature representation After strong-weak contrast learning, frequency domain reconstruction is performed through a fully connected layer to obtain the frequency domain reconstructed output. Subsequently, the frequency domain reconstructed output is converted into the time domain reconstructed output via inverse Fourier transform. The expression is as follows:

[0086]

[0087]

[0088] in, Represents a fully connected layer. Represents the learnable weight matrix. This represents a learnable bias vector. This represents the inverse Fourier transform operation, used to restore the frequency domain reconstruction result to the corresponding time domain signal, realizing feature mapping and reconstruction from the spectrum space to the time space.

[0089] S3. Train the model by calculating the time-frequency reconstruction loss and the strength-weak contrast learning loss of the frequency domain strength-weak contrast learning Mamba model:

[0090] S31. Calculate the strong-weak contrast learning loss;

[0091] Frequency domain feature representation of the output of the frequency domain Mamba feature extraction network Weakness contrast learning is performed by maximizing the frequency domain feature representation. Frequency domain feature representation of strong and weak sample pairs The similarity between them enhances the model's ability to identify key spectral features, resulting in a strong-weak contrast learning loss. .

[0092] S32. Calculate the time-frequency reconstruction loss of the model and use it together with the strong-weak contrastive learning loss for model training:

[0093]

[0094]

[0095]

[0096] in, The symbols represent the time-domain reconstruction loss and the frequency-domain reconstruction loss, which are used to constrain the model to accurately recover the original signal features in both the time and frequency domains, thereby ensuring the consistency and integrity of the reconstruction results. Representing the total training loss, by jointly optimizing the contrastive learning loss and the time-frequency reconstruction loss, we can improve the model's ability to model the global and local structure of signals while maintaining feature discriminativeness.

[0097] Step 4: A power battery fault detection method based on frequency domain strength-weak contrast learning of the Mamba model.

[0098] The preprocessed power battery data samples were constructed into a standardized dataset and divided into training, validation, and test sets in a 7:2:1 ratio. The training set contains only data under normal operating conditions and is used to learn the operating mode of the power battery in a healthy state. The validation set is used for model parameter tuning and fault threshold determination. The test set is used to test the model's detection accuracy and generalization ability under different fault scenarios.

[0099] During fault detection, the internal resistance, voltage, current, and temperature data of the power battery are selected as model inputs. After reconstructing the input data, the model generates corresponding output reconstruction results. Subsequently, by comparing the input data with the output reconstruction results, the reconstruction error is calculated, thereby assessing the deviation between the sample and the normal mode and providing a basis for fault determination.

[0100]

[0101] in, Represents the reconstruction error, indicating the input data. With reconstructing data The differences between them are identified. During training, a training set containing only power battery data under normal operating conditions is used to optimize the frequency domain strength contrast learning Mamba model by minimizing the training loss. Subsequently, the sample reconstruction error distribution is calculated using the validation set to determine the fault detection threshold. Finally, in the testing phase, the reconstruction error of the sample is compared with the set threshold: when the reconstruction error exceeds the threshold, the corresponding operating condition of the sample is determined to be a fault state; otherwise, it is determined to be a normal state, thus achieving accurate identification of abnormal operating conditions of the power battery.

[0102] Example 2:

[0103] This invention conducted comparative experiments on multivariate time-series data fault detection on the publicly available MSL dataset, comparing it with the classic deep learning method Anomaly Transformer (Anomaly Trans). The test results are shown in Table 1.

[0104] Table 1:

[0105]

[0106] Among them, Precision, Recall, and F-score represent the precision, recall, and F1 score of fault detection, respectively, used to evaluate the detection performance of the model. Higher values ​​indicate better fault detection performance. Floating-point operations and the number of model parameters measure the computational complexity and storage scale of the model, reflecting its operating efficiency in resource-constrained environments. As shown in Table 1, the frequency domain strength contrast learning Mamba model proposed in this invention exhibits high detection accuracy and stability in power battery fault detection tasks. At the same time, compared with Anomaly Trans, it significantly reduces the floating-point operations and parameter scale, greatly improving the inference speed and energy efficiency of the model on edge devices, meeting the application requirements of real-time monitoring and low-power operation.

[0107] The embodiments of the present invention have been 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 can be made to these embodiments, including components, without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A power battery fault detection method based on frequency domain strong-weak contrast learning Mamba, characterized in that, The method comprises the following steps: S1, collecting and preprocessing the original data of the power battery to generate a time series sample; S2, extracting the frequency domain feature representation of the power battery according to the time series sample through the constructed frequency domain strong-weak contrast learning Mamba model; comprising: S21, constructing a frequency domain strong-weak contrast learning Mamba model comprising a strong-weak data enhancement layer, a frequency domain Mamba feature extraction network and a time-frequency reconstruction network; S22, inputting the time series sample into the strong-weak data enhancement layer to generate a strong sample pair and a weak sample pair; S23, the frequency domain Mamba feature extraction network respectively performs Fourier transform operation on the time series sample, the strong sample pair and the weak sample pair to obtain respective frequency domain samples; after the frequency domain samples are mapped to a higher dimension through feature embedding, frequency domain representation modeling is performed to generate a frequency domain feature representation; S24, inputting the frequency domain feature representation into the time-frequency reconstruction network for frequency domain reconstruction to generate a time-frequency domain reconstruction output; S3, training the model by calculating the time-frequency reconstruction loss and the strong-weak contrast learning loss of the frequency domain strong-weak contrast learning Mamba model; comprising: S31, calculating the strong-weak contrast learning loss of the frequency domain strong-weak contrast learning Mamba model; S32, calculating the time-frequency reconstruction loss of the frequency domain strong-weak contrast learning Mamba model; the time-frequency reconstruction loss comprises a time domain reconstruction loss and a frequency domain reconstruction loss; the sum of the time-frequency reconstruction loss and the strong-weak contrast learning loss is used as the total training loss to train the model; S4, using the trained frequency domain strong-weak contrast learning Mamba model to detect power battery faults. In S1, the preprocessing comprises cleaning the original data of the power battery to remove noise points and invalid data; further comprising normalizing the valid data generated after cleaning to make different characteristic values within a unified dimension range; adopting a sliding window method to divide the normalized data into equal-length sequence segments in time sequence to obtain the time series sample.

2. The method according to claim 1, wherein, The strong-weak data enhancement layer comprises parallel strong data enhancement modules and weak data enhancement modules; in S22, the strong data enhancement module performs noise injection operation on the time series sample to generate the strong sample pair; the weak data enhancement module performs moving smoothing operation on the time series sample to generate the weak sample pair.

3. The method according to claim 1, characterized in that, The frequency domain Mamba feature extraction network comprises a frequency domain conversion layer and a Mamba module; in S23, the frequency domain conversion layer performs Fourier transform operation on the time series sample, the strong sample pair and the weak sample pair to generate respective frequency domain samples; the frequency domain samples are input into the Mamba module for frequency domain representation modeling to obtain the frequency domain feature representation of the time series sample, the frequency domain feature representation of the strong sample pair and the frequency domain feature representation of the weak sample pair, respectively.

4. The method according to claim 1, wherein, ​ 5. The method according to claim 4, characterized in that, The Mamba module sequentially comprises a linear layer, a convolution layer, a SiLU activation layer and a state space model, is used for extracting the dependent and dynamic relationship of the frequency domain features; the frequency domain samples are mapped to higher dimensions after feature embedding to generate each input sample of the Mamba module; the input sample is sequentially input to two independent parallel linear layers for feature mapping to obtain feature representation one and feature representation two respectively; The convolution layer and the SiLU activation function perform nonlinear processing on the feature representation one to obtain an intermediate feature representation; The intermediate feature representation is modeled by the state space model to generate a state space feature; the feature representation two is activated by the SiLU activation to obtain an activated feature; and the state space feature and the activated feature are multiplied element by element and then mapped by a linear layer to output the frequency domain feature representation. The time-frequency reconstruction network sequentially comprises a fully connected layer and a time domain restoration layer; in S24, the frequency domain feature representation is reconstructed by the fully connected layer to obtain a frequency domain reconstruction output; and the frequency domain reconstruction output is converted into a time domain reconstruction output by inverse Fourier transform through the time domain restoration layer.

6. The method according to claim 1, wherein, ​

Images