A Neural Network-Based Battery Impedance Spectrum Denoising Method

By employing a neural network-based battery impedance spectroscopy noise reduction method and dynamically matching the noise reduction output level, the problem of insufficient signal fidelity of electrochemical impedance spectroscopy under dynamic operating conditions is solved, achieving accurate noise reduction of impedance spectra and reliable battery state assessment.

CN123066567APending Publication Date: 2026-11-10ZHEJIANG TIANNENG NEW ENERGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611000950.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-07
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing technologies, electrochemical impedance spectroscopy has insufficient signal fidelity under dynamic operating conditions. Traditional filters and algorithms cannot adapt to complex and ever-changing noise environments, resulting in the loss of key features or evaluation bias.

Method used

A neural network-based battery impedance spectrum denoising method is adopted. By using a noise intensity evaluation gating network and a student autoencoder model, the denoising exit level is dynamically matched. Combined with the teacher autoencoder model for knowledge distillation training, adaptive denoising is achieved.

Benefits of technology

It achieves accurate noise reduction of impedance spectrum in noisy environments, ensuring that the output results are valid both mathematically and physically, and improving the reliability of battery state assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123066567A_ABST
    Figure CN123066567A_ABST
Patent Text Reader

Abstract

This invention relates to the field of neural network technology, specifically to a method for denoising battery impedance spectra based on neural networks. The method includes the following steps: acquiring impedance spectrum data of the battery to be processed; inputting the impedance spectrum data into a noise intensity evaluation gating network to obtain an impedance spectrum signal-to-noise ratio (SNR) evaluation result; determining the target denoising exit level in a student autoencoder model based on the SNR evaluation result; and inputting the impedance spectrum data into the student autoencoder model. This invention introduces electrochemical expertise to impose strong physical constraints, ensuring that the phase angle of the output impedance conforms to the electrochemical reaction mechanism, while simultaneously guaranteeing that the real and imaginary parts of the impedance satisfy the physical causal relationship defined by the Cramer-Kronen relation. In this way, the resulting denoising result is not only mathematically clean but also physically valid, improving the reliability of subsequent battery state assessments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to a method for reducing noise in battery impedance spectrum based on neural networks. Background Technology

[0002] While electrochemical impedance spectroscopy (EIS) is a powerful tool for analyzing the internal state of batteries, its practical deployment, especially in dynamic operating environments such as electric vehicles and energy storage power stations, faces challenges in signal fidelity. During battery system operation, high-frequency switching of the inverter, drastic load fluctuations, and coupling from the ambient electromagnetic field inject a large amount of non-stationary, wideband, and strong noise into the weak impedance measurement circuit. Traditional signal processing methods typically employ filters with fixed parameters or algorithms based on ideal models for noise reduction. The inherent limitation of these methods lies in their rigid design, which cannot adapt to complex and variable real-world noise environments. For example, a hardware filter designed to filter out interference at a specific frequency is ineffective against sudden noise in other frequency bands; simultaneously, if the filter design is too conservative, it will filter out some high-frequency impedance information characterizing early battery aging along with the noise, resulting in the permanent loss of crucial features. Similarly, averaging or smoothing algorithms often lead to blurring or shifting of key feature points in the impedance spectrum, such as the apex of a semicircle or the inflection point of a diffusion tail, when dealing with nonlinear noise. This directly results in the failure of subsequent equivalent circuit model fitting and deviations in health status assessment. Therefore, improvements are needed. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a battery impedance spectrum noise reduction method based on neural networks.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a battery impedance spectrum noise reduction method based on neural networks, comprising the following steps:

[0005] Obtain the impedance spectrum data of the battery to be processed;

[0006] The impedance spectrum data of the battery to be processed is input into a noise intensity evaluation gating network to obtain the impedance spectrum signal-to-noise ratio evaluation result.

[0007] The target noise reduction exit level is determined in the student autoencoder model based on the impedance spectrum signal-to-noise ratio evaluation results.

[0008] The battery impedance spectrum data to be processed is input into the student autoencoder model, and the impedance spectrum noise reduction result is output using the target noise reduction output level.

[0009] Obtain a historical impedance spectrum training set with clean labeled data, and use the historical impedance spectrum training set to train the initial teacher network model to obtain the teacher autoencoder model.

[0010] The student network model is trained by knowledge distillation using the historical impedance spectrum training set and the teacher autoencoder model to obtain the student autoencoder model.

[0011] Preferably, the steps for obtaining the impedance spectrum data of the battery to be processed are as follows:

[0012] Trigger the impedance acquisition command configured in the battery management system, and control the excitation generation circuit to inject a wideband voltage excitation signal into the target battery in the running state according to the impedance acquisition command;

[0013] The broadband current response signal generated by the target battery in response to the broadband voltage excitation signal is acquired synchronously;

[0014] The wideband voltage excitation signal and the wideband current response signal are transformed and mapped using the Fast Fourier Transform algorithm to obtain a frequency domain voltage complex sequence and a frequency domain current complex sequence.

[0015] Complex division is performed on each frequency element in the frequency domain voltage complex sequence and the corresponding frequency element in the frequency domain current complex sequence to calculate the discrete impedance complex sequence.

[0016] The mean value of the local region of the discrete impedance complex sequence is calculated using a sliding window of a preset width. The original discrete values ​​in the discrete impedance complex sequence are replaced with the mean value of the local region to obtain a smoothed impedance complex sequence.

[0017] The smoothed impedance complex sequence is separated into independent impedance real part characteristic sequences and impedance imaginary part characteristic sequences;

[0018] The impedance real part feature sequence and the impedance imaginary part feature sequence are concatenated along the first feature dimension to generate a multidimensional impedance feature vector.

[0019] The values ​​of each item in the multidimensional impedance feature vector are mapped to a preset value range using a normalization function, and the multidimensional impedance feature vector after mapping the value range is used as the impedance spectrum data of the battery to be processed.

[0020] Preferably, the step of inputting the impedance spectrum data of the battery to be processed into a noise intensity evaluation gating network to obtain the impedance spectrum signal-to-noise ratio evaluation result is as follows:

[0021] A pre-constructed lightweight convolutional neural network is extracted as the noise intensity evaluation gating network;

[0022] The impedance spectrum data of the battery to be processed is input into the first-dimensional convolutional feature layer of the noise intensity evaluation gating network to perform shallow spatial feature extraction, and the gating feature mapping matrix is ​​obtained.

[0023] The gated feature mapping matrix is ​​processed using a nonlinear activation processing function to obtain the activated gated feature mapping matrix;

[0024] The activated gating feature mapping matrix is ​​input into the global average pooling layer in the noise intensity evaluation gating network for feature compression to obtain the global environment feature vector.

[0025] The information entropy discreteness feature scalar is calculated based on the global environment feature vector and used as the impedance spectrum signal-to-noise ratio continuity index. The formula is as follows: ,in This indicates the continuity index of the impedance spectrum signal-to-noise ratio. This represents the total number of dimensions of the global environment feature vector. This indicates that the first element in the global environment feature vector is... The probability distribution values ​​after normalizing the feature values ​​at each position. This indicates the prevention of extremely small positive numbers whose logarithm is zero. Represents the logarithmic function with base ten;

[0026] After expanding the dimension of the impedance spectrum signal-to-noise ratio continuous index as an environmental complexity feature condition, it is concatenated and fused with the global environmental feature vector to obtain the fused environmental feature vector.

[0027] The fused environment feature vector is input into the fully connected classification output layer in the noise intensity evaluation gating network for high-dimensional nonlinear mapping, and the output is a discrete classification probability distribution feature set containing a preset noise level identifier;

[0028] The maximum classification probability value is extracted from the discrete classification probability distribution feature set;

[0029] Extract the specific noise level identifier that uniquely corresponds to the maximum classification probability value, and use the specific noise level identifier as the impedance spectrum signal-to-noise ratio evaluation result.

[0030] Preferably, the step of determining the target noise reduction exit level in the student autoencoder model based on the impedance spectrum signal-to-noise ratio evaluation results is as follows:

[0031] The multi-level network decoding structure contained in the pre-constructed student autoencoder model is analyzed to determine the first-level dynamic noise reduction output, the second-level dynamic noise reduction output, and the third-level dynamic noise reduction output deployed and configured inside the student autoencoder model.

[0032] Extract the noise intensity level corresponding to the impedance spectrum signal-to-noise ratio evaluation result;

[0033] The noise intensity level is compared with a first preset characteristic noise threshold and a second preset characteristic noise threshold pre-stored in a local database.

[0034] When the noise intensity level is less than or equal to the first preset characteristic noise threshold, the first level of dynamic noise reduction exit is determined as the target noise reduction exit level;

[0035] When the noise intensity level is greater than the first preset characteristic noise threshold and less than or equal to the second preset characteristic noise threshold, the second-level dynamic noise reduction output is determined as the target noise reduction output level.

[0036] If the noise intensity level is greater than the second preset characteristic noise threshold, the third-level dynamic noise reduction outlet is determined as the target noise reduction outlet level.

[0037] Preferably, the specific steps for inputting the battery impedance spectrum data to be processed into the student autoencoder model and using the target noise reduction output level to output the impedance spectrum noise reduction result are as follows:

[0038] The one-dimensional depthwise separable convolutional extraction encoder deployed at the front end of the student autoencoder model is invoked to perform feature dimensionality reduction on the battery impedance spectrum data to be processed.

[0039] The feature dimension reduction operation is used to compress and transform the feature dimension of the battery impedance spectrum data to be processed, so as to obtain a low-dimensional compressed impedance hidden feature representation.

[0040] The compressed impedance hidden feature representation is input into the back end of the student autoencoder model, which contains a multi-level hierarchical decoder network with a progressive structure, for upsampling feature reconstruction.

[0041] When the target noise reduction exit level is determined to be the first level dynamic noise reduction exit, only the first layer shallow scale decoder in the multi-level layer decoder network is activated to perform a single feature reconstruction calculation on the compressed impedance hidden feature representation. The feature value output by the first layer shallow scale decoder is used as the impedance spectrum noise reduction result and output to the outside through the first level dynamic noise reduction exit.

[0042] When the target noise reduction output level is determined to be the second-level dynamic noise reduction output, the first-level shallow-scale decoder and the second-level medium-scale decoder in the multi-level decoder network are cascaded and activated to perform progressive feature reconstruction calculation on the compressed impedance hidden feature representation in sequence. The feature value output by the second-level medium-scale decoder is used as the impedance spectrum noise reduction result and output to the outside through the second-level dynamic noise reduction output.

[0043] When the target noise reduction exit level is determined to be the third-level dynamic noise reduction exit, all available depth decoders in the multi-level layer decoder network are fully activated to perform complete layer-by-layer feature reconstruction on the compressed impedance hidden feature representation. The feature value output by the final layer depth decoder at the final depth is used as the impedance spectrum noise reduction result and output through the third-level dynamic noise reduction exit.

[0044] Preferably, the steps for obtaining a historical impedance spectrum training set with clean labeled data, and then using the historical impedance spectrum training set to train the initial teacher network model to obtain the teacher autoencoder model are as follows:

[0045] Impedance signal measurements were performed on multiple test electric vehicles using an impedance analysis instrument in a controlled, shielded laboratory environment to obtain a set of characteristic labels for the baseline pure target impedance spectrum under low ambient background noise.

[0046] The actual background operating noise spectrum of multiple test electric vehicles of the same model under driving conditions was automatically collected by a cloud data collection server.

[0047] The operating noise features are extracted using the actual background noise spectrum, and the operating noise features are randomly injected into the benchmark pure target impedance spectrum feature label set according to the preset signal-to-noise ratio gradient, thereby simulating and generating a set of noisy test historical impedance spectrum samples of a preset number of samples.

[0048] Each independent noisy sample in the noisy test history impedance spectrum sample set is paired and bound with the corresponding real reference pure impedance spectrum label data in the benchmark pure target impedance spectrum feature label set to construct the history impedance spectrum training set with pure label data.

[0049] Construct the initial teacher network model with multiple self-attention feature extraction operation units;

[0050] Noisy sample input data for the current training batch is extracted from the historical impedance spectrum training set containing clean labeled data;

[0051] The noisy sample input data of the current training batch is input into the initial teacher network model to execute the network forward propagation logic reasoning process, and the teacher forward denoising prediction and reconstruction feature sequence corresponding to the batch is output.

[0052] Obtain the real reference pure impedance spectrum label data corresponding to the noisy sample input data of the current training batch in the historical impedance spectrum training set with pure label data;

[0053] Calculate the mean square error feature loss value between the teacher forward noise reduction prediction reconstruction feature sequence and the real reference pure impedance spectrum label data;

[0054] The optimizer is used to update the internal node weight parameters of the initial teacher network model using gradient backpropagation based on the mean squared error feature loss value.

[0055] The complete optimization process of repeatedly iterating and extracting noisy sample input data of the current training batch and updating the internal node weight parameters continues until the mean squared error feature loss value reaches the convergence boundary state. The final network model with all feature parameters solidified is then extracted as the teacher autoencoder model.

[0056] Preferably, the specific steps for training the initial student network model using the historical impedance spectrum training set and the teacher autoencoder model to obtain the student autoencoder model are as follows:

[0057] Deploy and construct the initial student network model, which consists of a combination of deeply decoupled separable convolutional feature layers and multi-level dynamic noise reduction outputs;

[0058] The basic training sample stream data in the historical impedance spectrum training set with pure labeled data are input in parallel to the teacher autoencoder model in the weight parameter frozen state and the initial student network model in the weight parameter freely variable state.

[0059] The teacher hidden layer mapping feature tensor is extracted after performing feature extraction and inference on the network architecture of the teacher autoencoder model.

[0060] Simultaneously extract the set of multiple student stage hidden layer mapping feature tensors that overflow from the multi-level dynamic noise reduction output of the initial student network model during the staged feature decoding process;

[0061] Calculate the network-based self-reconstruction error loss index between the final network denoised generation sequence output by the initial student network model and the real reference pure impedance spectrum label data in the historical impedance spectrum training set with pure label data.

[0062] Extract the real and imaginary parts of the denoised impedance from the final network denoised generation sequence, and extract the real and imaginary parts of the tag impedance from the real reference pure impedance spectrum tag data;

[0063] The generated impedance phase angle is calculated based on the real part and the imaginary part of the noise-reduced impedance. The tag impedance phase angle is calculated based on the real part and the imaginary part of the tag impedance. The difference between the generated impedance phase angle and the tag impedance phase angle is calculated as an index of electrochemical phase consistency loss.

[0064] The real part of the noise-reduced impedance is calculated based on the preset Cramer-Kroni relation to obtain the corresponding theoretical causal impedance imaginary part sequence. The deviation between the theoretical causal impedance imaginary part sequence and the noise-reduced impedance imaginary part is calculated as an electrochemical-physical causal constraint loss index.

[0065] Calculate the knowledge distillation structural feature-level matching difference loss index between the teacher hidden layer mapping feature tensor and the set of multiple student stage hidden layer mapping feature tensors;

[0066] Assign a first control weight coefficient to the network foundation self-reconstruction error loss index, a second control weight coefficient to the knowledge distillation structure feature-level matching difference loss index, a third control weight coefficient to the electrochemical phase consistency loss index, and a fourth control weight coefficient to the electrochemical physical causal constraint loss index.

[0067] The weighted loss indexes of the network foundation self-reconstruction error, knowledge distillation structure feature-level matching difference, electrochemical phase consistency, and electrochemical physical causal constraint are added together to obtain the joint training feedback loss function.

[0068] The student autoencoder model is obtained by iteratively updating and optimizing all deep decoupled separable convolution weights in the initial student network model using the backpropagation algorithm based on the joint training feedback loss function, and the student autoencoder model is obtained after detecting that the value of the joint training feedback loss function is minimized.

[0069] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0070] This invention intelligently senses the noise intensity carried by the impedance spectrum data of the battery to be processed in advance, providing accurate prior information for subsequent noise reduction processing, and then dynamically matching the optimal depth of computing resources. For signals with light noise pollution, a lightweight, shallow feature reconstruction path is adopted to quickly output results and effectively save computing power; while for signals overwhelmed by strong noise, a deeper and more complex network structure is invoked for refined reconstruction to ensure thoroughness and accuracy of noise reduction. This adaptive resource scheduling mechanism achieves a balance between noise reduction effect and computational efficiency. First, a complex and powerful network learns deeply from massive amounts of data and masters the mapping rules from noisy data to clean data, then the core knowledge and judgment capabilities extracted are transferred to a more streamlined and efficient execution network. During training, not only is the final output noise reduction result required to be highly consistent with the real clean data numerically, but also strong physical constraints are introduced through electrochemical expertise to ensure that the phase angle of the output impedance conforms to the electrochemical reaction mechanism, while ensuring that the real and imaginary parts of the impedance satisfy the physical causal relationship defined by the Cramer-Kronenich relation. In this way, the noise reduction result is not only mathematically clean, but also physically effective, which improves the reliability of subsequent battery state assessment. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0073] Please see Figure 1 This invention provides a technical solution, a battery impedance spectrum noise reduction method based on neural networks, comprising the following steps:

[0074] Obtain the impedance spectrum data of the battery to be processed;

[0075] The impedance spectrum data of the battery to be processed is input into the noise intensity evaluation gating network to obtain the impedance spectrum signal-to-noise ratio evaluation result.

[0076] The target noise reduction exit level is determined in the student autoencoder model based on the impedance spectrum signal-to-noise ratio evaluation results.

[0077] The impedance spectrum data of the battery to be processed is input into the student autoencoder model, and the impedance spectrum denoising result is output using the target denoising output level.

[0078] Obtain a training set of historical impedance spectra with clean labels, and use the training set of historical impedance spectra to train the initial teacher network model to obtain the teacher autoencoder model.

[0079] The initial student network model was trained by knowledge distillation using the historical impedance spectrum training set and the teacher autoencoder model to obtain the student autoencoder model.

[0080] In this embodiment, the steps for acquiring the impedance spectrum data of the battery to be processed are as follows: triggering the impedance acquisition command configured in the battery management system, and controlling the excitation generation circuit to inject a wideband voltage excitation signal into the target battery in the running state according to the impedance acquisition command; synchronously acquiring the wideband current response signal generated by the target battery in response to the wideband voltage excitation signal; using the fast Fourier transform algorithm to transform and map the wideband voltage excitation signal and the wideband current response signal respectively to obtain the frequency domain voltage complex sequence and the frequency domain current complex sequence; performing complex division operation on each frequency element in the frequency domain voltage complex sequence and the corresponding frequency element in the frequency domain current complex sequence. The discrete impedance complex sequence is calculated; the local mean of the discrete impedance complex sequence is calculated using a sliding window of preset width, and the original discrete values ​​in the discrete impedance complex sequence are replaced with the local mean to obtain a smoothed impedance complex sequence; the smoothed impedance complex sequence is separated into independent impedance real part feature sequences and impedance imaginary part feature sequences; the impedance real part feature sequences and impedance imaginary part feature sequences are concatenated along the first feature dimension to generate a multidimensional impedance feature vector; the values ​​of each term in the multidimensional impedance feature vector are mapped to a preset value interval using a normalization function, and the multidimensional impedance feature vector after mapping the value interval is used as the battery impedance spectrum data to be processed.

[0081] Specifically, triggering the impedance acquisition command configured in the battery management system involves sending a start signal to the control port of the excitation generation circuit. This command includes the parameter configuration of the excitation signal, such as selecting a swept-frequency sine wave with a frequency scanning range of 0.01Hz to 10000Hz, or a wideband signal generated based on a pseudo-random binary sequence, and setting its injection amplitude to 5mV. The excitation generation circuit then injects the configured wideband voltage excitation signal into the positive and negative terminals of the target battery in operation according to the command. Simultaneously, a high-precision shunt or Hall effect sensor located in the current loop... The sensor synchronously acquires the broadband current response signal generated by the target battery in response to the excitation signal at a sampling frequency of 50kHz. The acquired continuous-time voltage and current signals are converted into discrete time-series data via an analog-to-digital converter. A Fast Fourier Transform algorithm with a Hanning window function is applied to the acquired voltage and current time series to reduce spectral leakage, mapping the time-domain signals to the frequency domain, resulting in a frequency-domain complex voltage sequence and a frequency-domain complex current sequence containing amplitude and phase frequency information. Each frequency point in the frequency-domain complex voltage sequence is traversed, and the complex value at that frequency point is read. Read the complex values ​​of the corresponding frequency points in the frequency domain current complex number sequence, perform complex division (dividing the complex voltage by the complex current) to calculate the discrete impedance complex value at that frequency point. After calculating all frequency points, aggregate them into a discrete impedance complex number sequence. Set a sliding window with a width of 5 frequency points, and slide it across the discrete impedance complex number sequence with a step size of 1 frequency point. Calculate the average of the real part and the average of the imaginary part of the 5 impedance complex values ​​within each window. Replace the original impedance complex value at the center frequency point of the window with the calculated local average value to complete the smoothing process of the entire sequence, obtaining a smoothed impedance complex number sequence. Each complex value in the smoothed impedance complex sequence is separated into its corresponding real and imaginary parts, forming independent impedance real part feature sequences and impedance imaginary part feature sequences. For example, if the impedance spectrum contains 128 frequency points, then the dimensions of the two sequences are both 1x128. The two sequences are spliced ​​along the new feature dimension to generate a multidimensional impedance feature vector with a dimension of 2x128. Finally, the minimum-maximum normalization function is applied to linearly map each value in the multidimensional impedance feature vector to the closed interval [0,1]. The multidimensional impedance feature vector after mapping the numerical interval is used as the battery impedance spectrum data to be processed.

[0082] In this embodiment, the steps for inputting the battery impedance spectrum data to be processed into a noise intensity assessment gating network to obtain the impedance spectrum signal-to-noise ratio assessment result are as follows: A pre-constructed lightweight convolutional neural network is extracted as the noise intensity assessment gating network; the battery impedance spectrum data to be processed is input into the first-dimensional convolutional feature layer of the noise intensity assessment gating network for shallow spatial feature extraction to obtain a gating feature mapping matrix; the gating feature mapping matrix is ​​processed using a nonlinear activation function to obtain an activated gating feature mapping matrix; the activated gating feature mapping matrix is ​​input into the global average pooling layer of the noise intensity assessment gating network for feature compression to obtain a global environment feature vector; the information entropy discreteness feature scalar is calculated based on the global environment feature vector as the continuous index of the impedance spectrum signal-to-noise ratio, as shown in the formula: ,in This indicates the continuity of the impedance spectrum signal-to-noise ratio. This represents the total number of dimensions of the global environment feature vector. This represents the first element in the global environment feature vector. The probability distribution values ​​after normalizing the feature values ​​at each position. This indicates the prevention of extremely small positive numbers whose logarithm is zero. The function represents a logarithmic function with base 10. The impedance spectrum signal-to-noise ratio (SNR) continuous index is used as a feature condition for environmental complexity. After dimensional expansion, it is concatenated and fused with the global environmental feature vector to obtain a fused environmental feature vector. This fused environmental feature vector is input into the fully connected classification output layer of the noise intensity assessment gating network for high-dimensional nonlinear mapping, outputting a discrete classification probability distribution feature set containing a preset noise level identifier. The maximum classification probability value in the discrete classification probability distribution feature set is extracted. The specific noise level identifier uniquely corresponding to the maximum classification probability value is extracted and used as the impedance spectrum SNR assessment result.

[0083] Specifically, a pre-constructed lightweight convolutional neural network is extracted as the noise intensity evaluation gating network. This network consists of an input layer, two one-dimensional convolutional layers, a global average pooling layer, and a fully connected output layer. The input layer receives 2x128 dimension battery impedance spectrum data to be processed. The first one-dimensional convolutional layer is configured with 16 3x1 kernels with a stride of 1, and the second one-dimensional convolutional layer is configured with 32 3x1 kernels with a stride of 1. Both convolutional layers use the ReLU linear rectified function as the activation function. The battery impedance spectrum data to be processed is input into the first-dimensional convolutional feature layer of this network. Through parallel convolution operations of 16 kernels, the local spatial characteristics of the impedance spectrum in the frequency domain are extracted. Inter-correlation features are used to generate a 16x126-dimensional gated feature map matrix. Then, the ReLU activation function is applied to non-linearly process all elements in this matrix, setting all values ​​less than 0 to 0. This processed, activated gated feature map matrix is ​​then input into a second one-dimensional convolutional layer to further extract deeper features. The resulting feature map is then input into a global average pooling layer in the network. This layer calculates the average value of all values ​​in each feature map (channel), compressing the 32 feature maps into a one-dimensional vector containing 32 elements. This is the global environment feature vector. Based on this global environment feature vector, the Softmax function is first applied to convert it into a probability distribution, and then the information entropy formula is used... Calculate the information entropy discrete feature scalar, where This represents a scalar value that serves as a continuous index of the impedance spectrum signal-to-noise ratio. This represents the total dimension of the global environment feature vector, which is 32. This indicates the global environment feature vector after processing by the Softmax function. The probability distribution values ​​for each position. To prevent the logarithmic operation parameter from being zero, a very small positive number is set, specifically 1e-8, which will be used to calculate the scalar index. Dimension expansion is performed by copying the value 32 times to form a 32-dimensional vector. This vector is then concatenated with the original 32-dimensional global environment feature vector along the feature channel dimension to form a 64-dimensional fused environment feature vector. This fused environment feature vector is input into the last fully connected classification output layer of the network. This layer has 3 neurons, corresponding to three preset noise levels: low, medium, and high. The Softmax activation function is applied to output a discrete classification probability distribution feature set containing 3 probability values. The maximum probability value in this set is selected, and the noise level label corresponding to the maximum probability value is extracted. For example, if the output is [0.85, 0.10, 0.05], the maximum probability is 0.85, and its corresponding level is "low noise". This specific noise level label is used as the impedance spectrum signal-to-noise ratio evaluation result.

[0084] In this embodiment, the specific steps for determining the target noise reduction exit level in the student autoencoder model based on the impedance spectrum signal-to-noise ratio evaluation results are as follows: Analyzing the multi-level network decoding structure contained within the pre-built student autoencoder model to determine the first-level dynamic noise reduction exit, the second-level dynamic noise reduction exit, and the third-level dynamic noise reduction exit deployed within the student autoencoder model; extracting the noise intensity level corresponding to the impedance spectrum signal-to-noise ratio evaluation results; comparing the noise intensity level with the first preset characteristic noise threshold and the second preset characteristic noise threshold pre-stored in the local database; if the noise intensity level is less than or equal to the first preset characteristic noise threshold, determining the first-level dynamic noise reduction exit as the target noise reduction exit level; if the noise intensity level is greater than the first preset characteristic noise threshold and less than or equal to the second preset characteristic noise threshold, determining the second-level dynamic noise reduction exit as the target noise reduction exit level; if the noise intensity level is greater than the second preset characteristic noise threshold, determining the third-level dynamic noise reduction exit as the target noise reduction exit level.

[0085] Specifically, the multi-level network decoding structure within the pre-constructed student autoencoder model is analyzed. This structure consists of a series of upsampling layers and convolutional layers stacked together, with three exit points at different depths of the decoder, labeled as the first-level dynamic noise reduction exit, the second-level dynamic noise reduction exit, and the third-level dynamic noise reduction exit. The first-level exit is located in the shallowest layer, the second-level in the middle layer, and the third-level in the deepest layer. The noise intensity level corresponding to the impedance spectrum signal-to-noise ratio (SNR) evaluation result obtained in the previous step is extracted. This level is one of three discrete classification labels: "low," "medium," and "high." The noise intensity level is compared with the first and second preset feature noise thresholds stored in the local database. These two thresholds are empirical boundary values ​​based on the analysis of a large amount of historical data. The setting process is as follows: tens of thousands of impedance spectrum data points with SNR labels are collected. Samples with an SNR higher than 25 dB are classified as low noise samples, and samples with an SNR between 10 dB and 25 dB are classified as high noise samples. Samples with noise levels between dB are classified as medium noise samples, and those below 10 dB are classified as high noise samples. These samples are input into a noise intensity assessment gating network to obtain the continuous impedance spectrum signal-to-noise ratio index E for each sample. The mean of index E in the boundary region between all low-noise and medium-noise samples is calculated as the first preset characteristic noise threshold, and the mean of index E in the boundary region between all medium-noise and high-noise samples is calculated as the second preset characteristic noise threshold. For example, the first preset characteristic noise threshold is calculated to be 0.9, and the second preset characteristic noise threshold is calculated to be 1.8. The specific comparison and judgment logic is as follows: if the currently acquired noise intensity level is determined to be "low", the system automatically determines the first-level dynamic noise reduction output as the target noise reduction output level; if the currently acquired noise intensity level is determined to be "medium", the second-level dynamic noise reduction output is determined as the target noise reduction output level; and if the currently acquired noise intensity level is determined to be "high", the third-level dynamic noise reduction output is determined as the target noise reduction output level.

[0086] In this embodiment, the steps of inputting the battery impedance spectrum data to be processed into the student autoencoder model and outputting the impedance spectrum denoising result using the target denoising exit level are as follows: The one-dimensional depthwise separable convolutional extractor deployed at the front end of the student autoencoder model is invoked to perform feature dimensionality reduction on the battery impedance spectrum data to be processed; the feature dimensions of the battery impedance spectrum data to be processed are compressed and transformed through feature dimensionality reduction to obtain a low-dimensional compressed impedance hidden feature representation; the compressed impedance hidden feature representation is input into a multi-level hierarchical decoder network containing a progressively layered structure at the back end of the student autoencoder model for upsampling feature reconstruction; when the target denoising exit level is determined to be the first-level dynamic denoising exit level, only the first-level shallow-scale decoder in the multi-level hierarchical decoder network is activated to perform a single feature reconstruction calculation on the compressed impedance hidden feature representation, thus converting the first-level shallow-scale decoder into a low-dimensional compressed impedance hidden feature representation. The feature values ​​output by the layer-scale decoder are used as impedance spectrum denoising results and output outward through the first-level dynamic denoising exit. When the target denoising exit level is determined to be the second-level dynamic denoising exit, the first-layer shallow-scale decoder and the second-layer mid-scale decoder in the cascaded multi-level layer-scale decoder network are activated to perform progressive feature reconstruction calculations on the compressed impedance hidden feature representation in sequence. The feature values ​​output by the second-layer mid-scale decoder are used as impedance spectrum denoising results and output outward through the second-level dynamic denoising exit. When the target denoising exit level is determined to be the third-level dynamic denoising exit, all available depth decoders in the multi-level layer-scale decoder network are fully activated to perform complete layer-by-layer feature reconstruction on the compressed impedance hidden feature representation. The feature values ​​output by the final-layer depth decoder at the final depth are used as impedance spectrum denoising results and output outward through the third-level dynamic denoising exit.

[0087] Specifically, the one-dimensional depthwise separable convolutional encoder deployed at the front end of the student autoencoder model performs feature dimensionality reduction on the battery impedance spectrum data to be processed. This encoder consists of multiple stacked depthwise separable convolutional blocks, each containing a depthwise convolutional layer and a pointwise convolutional layer. Through a series of convolution and downsampling operations, the input 2x128-dimensional battery impedance spectrum data is progressively compressed, extracting its core features in the high-dimensional space to obtain a low-dimensional compressed impedance hidden feature representation, for example, with a dimension of 8x16. This compressed impedance hidden feature representation is then input to the back end of the student autoencoder model, which contains a multi-level hierarchical decoder network with a progressively layered structure, for upsampling feature reconstruction. This decoder is symmetrical to the encoder structure and consists of multiple upsampling layers and depthwise separable convolutional blocks. Based on the target noise reduction exit level determined in the previous step, a selective decoding path is executed. If the target noise reduction exit level is determined to be the first-level dynamic noise reduction exit level, only the first-level shallow-scale decoder in the multi-level hierarchical decoder network is activated. This decoder contains one upsampling operation and one convolution operation to perform single-level decoding on the compressed impedance hidden feature representation. The feature reconstruction calculation restores the dimension of the feature values ​​output by the decoder to 2x128 and outputs them directly through the first-level dynamic noise reduction output as the impedance spectrum noise reduction result. When the target noise reduction output level is determined to be the second-level dynamic noise reduction output, the first-level shallow-scale decoder and the second-level mid-scale decoder in the multi-level layer decoder network are cascaded and activated. The compressed impedance hidden feature representation is first input into the first-level decoder, and its output is not used as the final output, but continues to be used as the input of the second-level decoder for a second progressive feature reconstruction calculation. The feature values ​​output by the second-level mid-scale decoder are used as the impedance spectrum noise reduction result and output through the second-level dynamic noise reduction output. When the target noise reduction output level is determined to be the third-level dynamic noise reduction output, all available depth decoders in the multi-level layer decoder network are fully activated, allowing the compressed impedance hidden feature representation to pass completely and layer by layer through all upsampling and convolutional layers in the decoder network for the most complete layer-by-layer feature reconstruction. The feature values ​​output by the final depth decoder at the final depth are used as the impedance spectrum noise reduction result and output through the third-level dynamic noise reduction output.

[0088] In this embodiment, the steps for obtaining a historical impedance spectrum training set with clean label data and training the initial teacher network model using the historical impedance spectrum training set to obtain the teacher autoencoder model are as follows: Impedance signal measurements are performed on multiple test electric vehicles using an impedance analysis instrument in a controlled, shielded laboratory environment to obtain a benchmark clean target impedance spectrum feature label set under low environmental background noise; Actual background operating noise maps automatically recorded by multiple test electric vehicles of the same model under driving conditions are collected through a cloud data collection server; Operating condition noise features are extracted using the actual background operating noise maps, and these features are randomly injected into the benchmark clean target impedance spectrum feature label set according to a preset signal-to-noise ratio gradient, simulating the generation of a preset number of noisy test historical impedance spectrum sample sets; Each independent noisy sample in the noisy test historical impedance spectrum sample set is paired and bound with the corresponding real reference clean impedance spectrum label data in the benchmark clean target impedance spectrum feature label set to construct a historical impedance spectrum training set with clean label data; A multi-... The initial teacher network model consists of a self-attention feature extraction unit; noisy sample input data for the current training batch is extracted from the historical impedance spectrum training set with clean labeled data; the noisy sample input data of the current training batch is input into the initial teacher network model to execute the network forward propagation logic inference process, and outputs the teacher forward denoising prediction reconstruction feature sequence corresponding to the batch; the real reference clean impedance spectrum label data corresponding to the noisy sample input data of the current training batch is obtained from the historical impedance spectrum training set with clean labeled data; the mean square error feature loss value between the teacher forward denoising prediction reconstruction feature sequence and the real reference clean impedance spectrum label data is calculated; the optimizer is used to update the internal node weight parameters of the initial teacher network model with gradient backpropagation based on the mean square error feature loss value; the complete optimization process of extracting noisy sample input data of the current training batch and updating internal node weight parameters is repeated iteratively until the mean square error feature loss value reaches the convergence boundary state, and the final network model with all feature parameters solidified is extracted as the teacher autoencoder model.

[0089] Specifically, in an electromagnetically shielded controlled laboratory, an electrochemical workstation was used to measure the impedance signals of the battery packs of multiple test electric vehicles in a healthy state. The frequency sweep range covered 0.01Hz to 10000Hz, obtaining a series of reference pure target impedance spectrum feature labels under extremely low ambient background noise. Simultaneously, through a data collection server deployed in the cloud, background voltage and current noise spectra automatically recorded by the battery management systems of multiple test electric vehicles of the same model operating on actual roads were collected. Pure operating condition noise was extracted from these actual background operating noise spectra. The acoustic feature sequence is generated, and these operating condition noise features are randomly injected into the previously acquired benchmark pure target impedance spectrum feature label set according to a preset signal-to-noise ratio gradient, for example from 0dB to 30dB, at 5dB intervals. Simulation generates a noisy test historical impedance spectrum sample set containing 100,000 samples. Each independent noisy sample in this set is paired and bound with its corresponding original pure label data before generation, constructing a historical impedance spectrum training set with pure label data. An initial teacher network model is then constructed, which uses a Transformer architecture containing 8 encoder layers and 8 decoder layers. The autoencoder architecture consists of an encoder layer containing a multi-head self-attention mechanism unit and a feedforward neural network, and a decoder layer with an additional cross-attention mechanism unit inserted between them. Noisy sample input data for the current training batch is extracted from the constructed historical impedance spectrum training set in batches of 256 samples each. This batch of data is input into the initial teacher network model, and the complete forward propagation logic inference process is executed. The output is the teacher forward denoising prediction and reconstruction feature sequence corresponding to that batch. Simultaneously, the real reference clean impedance spectrum label data corresponding one-to-one with the noisy samples in the training set is obtained. The mean squared error feature loss value between the predicted reconstruction feature sequence and the real label data is calculated. Using the Adam optimizer with a learning rate of 0.0001, the gradient backpropagation values ​​of the weight parameters of all nodes within the initial teacher network model are updated based on the calculated mean squared error feature loss value. This complete optimization process of extracting batch data, forward propagation, calculating loss, and backpropagation is repeated until the mean squared error feature loss value on the validation set no longer decreases for 10 consecutive cycles, reaching the convergence boundary state. At this point, the finalized network model with all feature parameters solidified is extracted as the teacher autoencoder model.

[0090] In this embodiment, the steps for training the initial student network model using knowledge distillation with the historical impedance spectrum training set and the teacher autoencoder model to obtain the student autoencoder model are as follows: Deploying and constructing an initial student network model composed of a deep decoupled separable convolutional feature layer and a multi-level dynamic noise reduction output; inputting the basic training sample stream data from the historical impedance spectrum training set with clean labeled data into the teacher autoencoder model (in a frozen weight parameter state) and the initial student network model (in a freely variable weight parameter state); extracting the teacher hidden layer mapping feature tensor output after feature extraction and inference of the network architecture of the teacher autoencoder model; and simultaneously extracting the initial student network model... The multi-stage dynamic denoising output of this type overflows the set of multiple student-stage hidden layer mapping feature tensors during the staged feature decoding process; the network-based self-reconstruction error loss index is calculated between the final network denoised generation sequence output by the initial student network model and the real reference pure impedance spectrum label data in the historical impedance spectrum training set with pure label data; the real and imaginary parts of the denoised impedance are extracted from the final network denoised generation sequence, and the real and imaginary parts of the label impedance are extracted from the real reference pure impedance spectrum label data; the generated impedance phase angle is calculated based on the real and imaginary parts of the denoised impedance, and the label impedance is calculated based on the real and imaginary parts of the label impedance. The phase angle is calculated, and the difference between the generated impedance phase angle and the tag impedance phase angle is used as an electrochemical phase consistency loss index. Based on the preset Cramer-Kroni relation, the real part of the denoised impedance is calculated to obtain the corresponding theoretical causal impedance imaginary part sequence. The deviation between the theoretical causal impedance imaginary part sequence and the denoised impedance imaginary part is calculated as an electrochemical physical causal constraint loss index. The knowledge distillation structure feature-level matching difference loss index is calculated between the teacher's hidden layer mapping feature tensor and the set of multiple student stage hidden layer mapping feature tensors. A first control weight coefficient is assigned to the network foundation self-reconstruction error loss index, and a second control weight coefficient is assigned to the knowledge distillation structure feature-level matching difference loss index. The electrochemical phase consistency loss index is assigned a third control weight coefficient, and the electrochemical physical causal constraint loss index is assigned a fourth control weight coefficient. The weighted network self-reconstruction error loss index, the weighted knowledge distillation structure feature-level matching difference loss index, the weighted electrochemical phase consistency loss index, and the weighted electrochemical physical causal constraint loss index are added together to obtain the joint training feedback loss function. The backpropagation algorithm is used to iteratively update and optimize all deep decoupled separable convolution weights in the initial student network model based on the joint training feedback loss function. The student autoencoder model is obtained after the value of the joint training feedback loss function is minimized.

[0091] Specifically, an initial student network model is constructed, consisting of a deeply decoupled separable convolutional feature layer and a multi-level dynamic noise reduction output. This model structure is lighter than the teacher network. Both the encoder and decoder use deeply separable convolutional blocks instead of standard convolutions. Basic training sample stream data from the constructed historical impedance spectrum training set with clean labels is batch-wise input into both the teacher autoencoder model (with frozen weights) and the initial student network model (with freely variable weights). The teacher hidden layer mapping feature tensor, output from the last layer of the encoder in the teacher autoencoder model after feature extraction inference, is extracted. Simultaneously, the initial student network model's phased feature extraction is also extracted. During the decoding process, the three student-stage hidden layer mapping feature tensors output by the decoder layer corresponding to the three dynamic denoising outputs form a set. The mean square error between the final network denoised generation sequence output by the initial student network model and the corresponding real reference pure impedance spectrum label data in the training set is calculated as the network's own reconstruction error loss index. The real and imaginary parts of the denoised impedance in the denoised generation sequence, as well as the real and imaginary parts of the label impedance in the real label data, are extracted respectively. Based on these two sets of real and imaginary parts, the impedance phase angle at each frequency point is calculated. Then, the mean absolute error between the two sets of phase angle sequences is calculated as the electrochemical phase consistency loss index. Subsequently, based on the Kramers-Kronig relation, an integral transformation is performed on the real part sequence of the denoised impedance to calculate the theoretically corresponding imaginary part sequence of causal impedance. Then, the mean square error between this theoretically causal imaginary part sequence and the actually generated imaginary part sequence of the denoised impedance is calculated as an electrochemical physics causal constraint loss index. Simultaneously, the mean square error between the teacher's hidden layer mapping feature tensor and the student's stage-specific hidden layer mapping feature tensor set, and the tensor that best matches the teacher's dimension, is calculated as a knowledge distillation structure feature-level matching difference loss index. Control weight coefficients are assigned to these four loss indices, their settings based on the importance of each loss. For example, the network foundation itself reconstruction error loss index is assigned a weight... The first control weight coefficient is assigned to 1.0, the second control weight coefficient to the knowledge distillation structure feature-level matching difference loss index is assigned to 0.6, the third control weight coefficient to the electrochemical phase consistency loss index is assigned to 0.2, and the fourth control weight coefficient to the electrochemical physical causal constraint loss index is assigned to 0.2. The four weighted loss indices are added together to obtain the joint training feedback loss function. The Adam backpropagation algorithm is used to iteratively update and optimize all deep decoupled separable convolution weights in the initial student network model based on this joint training feedback loss function. After detecting that the value of the joint training feedback loss function on the validation set no longer decreases for 20 consecutive cycles and reaches its minimum, the student autoencoder model is obtained.

[0092] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A battery impedance spectrum noise reduction method based on neural networks, characterized in that, Includes the following steps: Obtain the impedance spectrum data of the battery to be processed; The impedance spectrum data of the battery to be processed is input into a noise intensity evaluation gating network to obtain the impedance spectrum signal-to-noise ratio evaluation result. The target noise reduction exit level is determined in the student autoencoder model based on the impedance spectrum signal-to-noise ratio evaluation results. The battery impedance spectrum data to be processed is input into the student autoencoder model, and the impedance spectrum noise reduction result is output using the target noise reduction output level. Obtain a historical impedance spectrum training set with clean labeled data, and use the historical impedance spectrum training set to train the initial teacher network model to obtain the teacher autoencoder model. The student network model is trained by knowledge distillation using the historical impedance spectrum training set and the teacher autoencoder model to obtain the student autoencoder model.

2. The battery impedance spectrum noise reduction method based on neural networks according to claim 1, characterized in that, The specific steps for obtaining the impedance spectrum data of the battery to be processed are as follows: Trigger the impedance acquisition command configured in the battery management system, and control the excitation generation circuit to inject a wideband voltage excitation signal into the target battery in the running state according to the impedance acquisition command; The broadband current response signal generated by the target battery in response to the broadband voltage excitation signal is acquired synchronously; The wideband voltage excitation signal and the wideband current response signal are transformed and mapped using the Fast Fourier Transform algorithm to obtain a frequency domain voltage complex sequence and a frequency domain current complex sequence. Complex division is performed on each frequency element in the frequency domain voltage complex sequence and the corresponding frequency element in the frequency domain current complex sequence to calculate the discrete impedance complex sequence. The mean value of the local region of the discrete impedance complex sequence is calculated using a sliding window of a preset width. The original discrete values ​​in the discrete impedance complex sequence are replaced with the mean value of the local region to obtain a smoothed impedance complex sequence. The smoothed impedance complex sequence is separated into independent impedance real part characteristic sequences and impedance imaginary part characteristic sequences; The impedance real part feature sequence and the impedance imaginary part feature sequence are concatenated along the first feature dimension to generate a multidimensional impedance feature vector. The values ​​of each item in the multidimensional impedance feature vector are mapped to a preset value range using a normalization function, and the multidimensional impedance feature vector after mapping the value range is used as the impedance spectrum data of the battery to be processed.

3. The battery impedance spectrum noise reduction method based on neural networks according to claim 1, characterized in that, The specific steps for inputting the impedance spectrum data of the battery to be processed into a noise intensity evaluation gating network to obtain the impedance spectrum signal-to-noise ratio evaluation result are as follows: A pre-constructed lightweight convolutional neural network is extracted as the noise intensity evaluation gating network; The impedance spectrum data of the battery to be processed is input into the first-dimensional convolutional feature layer of the noise intensity evaluation gating network to perform shallow spatial feature extraction, and the gating feature mapping matrix is ​​obtained. The gated feature mapping matrix is ​​processed using a nonlinear activation processing function to obtain the activated gated feature mapping matrix; The activated gating feature mapping matrix is ​​input into the global average pooling layer in the noise intensity evaluation gating network for feature compression to obtain the global environment feature vector. The information entropy discreteness feature scalar is calculated based on the global environment feature vector and used as the impedance spectrum signal-to-noise ratio continuity index. The formula is as follows: ,in This indicates the continuity index of the impedance spectrum signal-to-noise ratio. This represents the total number of dimensions of the global environment feature vector. This indicates that the first element in the global environment feature vector is... The probability distribution values ​​after normalizing the feature values ​​at each position. This indicates the prevention of extremely small positive numbers whose logarithm is zero. Represents the logarithmic function with base ten; After expanding the dimension of the impedance spectrum signal-to-noise ratio continuous index as an environmental complexity feature condition, it is concatenated and fused with the global environmental feature vector to obtain the fused environmental feature vector. The fused environment feature vector is input into the fully connected classification output layer in the noise intensity evaluation gating network for high-dimensional nonlinear mapping, and the output is a discrete classification probability distribution feature set containing a preset noise level identifier; The maximum classification probability value is extracted from the discrete classification probability distribution feature set; Extract the specific noise level identifier that uniquely corresponds to the maximum classification probability value, and use the specific noise level identifier as the impedance spectrum signal-to-noise ratio evaluation result.

4. The battery impedance spectrum noise reduction method based on neural networks according to claim 1, characterized in that, The specific steps for determining the target noise reduction exit level in the student autoencoder model based on the impedance spectrum signal-to-noise ratio evaluation results are as follows: The multi-level network decoding structure contained in the pre-constructed student autoencoder model is analyzed to determine the first-level dynamic noise reduction output, the second-level dynamic noise reduction output, and the third-level dynamic noise reduction output deployed and configured inside the student autoencoder model. Extract the noise intensity level corresponding to the impedance spectrum signal-to-noise ratio evaluation result; The noise intensity level is compared with a first preset characteristic noise threshold and a second preset characteristic noise threshold pre-stored in a local database. When the noise intensity level is less than or equal to the first preset characteristic noise threshold, the first level of dynamic noise reduction output is determined as the target noise reduction output level; When the noise intensity level is greater than the first preset characteristic noise threshold and less than or equal to the second preset characteristic noise threshold, the second-level dynamic noise reduction output is determined as the target noise reduction output level. If the noise intensity level is greater than the second preset characteristic noise threshold, the third-level dynamic noise reduction outlet is determined as the target noise reduction outlet level.

5. The battery impedance spectrum noise reduction method based on neural networks according to claim 4, characterized in that, The specific steps for inputting the battery impedance spectrum data to be processed into the student autoencoder model and using the target noise reduction output level to output the impedance spectrum noise reduction result are as follows: The one-dimensional depthwise separable convolutional extraction encoder deployed at the front end of the student autoencoder model is invoked to perform feature dimensionality reduction operations on the battery impedance spectrum data to be processed. The feature dimension reduction operation is used to compress and transform the feature dimension of the battery impedance spectrum data to be processed, so as to obtain a low-dimensional compressed impedance hidden feature representation. The compressed impedance hidden feature representation is input into the back end of the student autoencoder model, which contains a multi-level hierarchical decoder network with a progressive structure, for upsampling feature reconstruction. When the target noise reduction exit level is determined to be the first level dynamic noise reduction exit, only the first layer shallow scale decoder in the multi-level layer decoder network is activated to perform a single feature reconstruction calculation on the compressed impedance hidden feature representation. The feature value output by the first layer shallow scale decoder is used as the impedance spectrum noise reduction result and output to the outside through the first level dynamic noise reduction exit. When the target noise reduction output level is determined to be the second-level dynamic noise reduction output, the first-level shallow-scale decoder and the second-level medium-scale decoder in the multi-level decoder network are cascaded and activated to perform progressive feature reconstruction calculation on the compressed impedance hidden feature representation in sequence. The feature value output by the second-level medium-scale decoder is used as the impedance spectrum noise reduction result and output to the outside through the second-level dynamic noise reduction output. When the target noise reduction exit level is determined to be the third-level dynamic noise reduction exit, all available depth decoders in the multi-level layer decoder network are fully activated to perform complete layer-by-layer feature reconstruction on the compressed impedance hidden feature representation. The feature value output by the final layer depth decoder at the final depth is used as the impedance spectrum noise reduction result and output through the third-level dynamic noise reduction exit.

6. The battery impedance spectrum noise reduction method based on neural networks according to claim 1, characterized in that, The specific steps for obtaining a historical impedance spectrum training set with clean labeled data, and then using this historical impedance spectrum training set to train the initial teacher network model to obtain the teacher autoencoder model are as follows: Impedance signal measurements were performed on multiple test electric vehicles using an impedance analysis instrument in a controlled, shielded laboratory environment to obtain a set of characteristic labels for the baseline pure target impedance spectrum under low ambient background noise. The actual background operating noise spectrum of multiple test electric vehicles of the same model under driving conditions was automatically collected by a cloud data collection server. The operating noise features are extracted using the actual background noise spectrum, and the operating noise features are randomly injected into the benchmark pure target impedance spectrum feature label set according to the preset signal-to-noise ratio gradient, thereby simulating and generating a set of noisy test historical impedance spectrum samples of a preset number of samples. Each independent noisy sample in the noisy test history impedance spectrum sample set is paired and bound with the corresponding real reference pure impedance spectrum label data in the benchmark pure target impedance spectrum feature label set to construct the history impedance spectrum training set with pure label data. Construct the initial teacher network model with multiple self-attention feature extraction operation units; Noisy sample input data for the current training batch is extracted from the historical impedance spectrum training set containing clean-label data; The noisy sample input data of the current training batch is input into the initial teacher network model to execute the network forward propagation logic reasoning process, and the teacher forward denoising prediction and reconstruction feature sequence corresponding to the batch is output. Obtain the real reference pure impedance spectrum label data corresponding to the noisy sample input data of the current training batch in the historical impedance spectrum training set with pure label data; Calculate the mean square error feature loss value between the teacher forward noise reduction prediction reconstruction feature sequence and the real reference pure impedance spectrum label data; The optimizer is used to update the internal node weight parameters of the initial teacher network model using gradient backpropagation based on the mean squared error feature loss value. The complete optimization process of repeatedly iterating and extracting noisy sample input data of the current training batch and updating the internal node weight parameters continues until the mean squared error feature loss value reaches the convergence boundary state. The final network model with all feature parameters solidified is then extracted as the teacher autoencoder model.

7. The battery impedance spectrum noise reduction method based on neural networks according to claim 1, characterized in that, The specific steps for training the initial student network model using the historical impedance spectrum training set and the teacher autoencoder model to obtain the student autoencoder model are as follows: Deploy and construct the initial student network model, which consists of a combination of deeply decoupled separable convolutional feature layers and multi-level dynamic noise reduction outputs; The basic training sample stream data in the historical impedance spectrum training set with pure labeled data are input in parallel to the teacher autoencoder model in the weight parameter frozen state and the initial student network model in the weight parameter freely variable state. The teacher hidden layer mapping feature tensor is extracted after performing feature extraction and inference on the network architecture of the teacher autoencoder model. Simultaneously extract the set of multiple student stage hidden layer mapping feature tensors that overflow from the multi-level dynamic noise reduction output of the initial student network model during the staged feature decoding process; Calculate the network-based self-reconstruction error loss index between the final network denoised generation sequence output by the initial student network model and the real reference pure impedance spectrum label data in the historical impedance spectrum training set with pure label data. Extract the real and imaginary parts of the denoised impedance from the final network denoised generation sequence, and extract the real and imaginary parts of the tag impedance from the real reference pure impedance spectrum tag data; The generated impedance phase angle is calculated based on the real part and the imaginary part of the noise-reduced impedance. The tag impedance phase angle is calculated based on the real part and the imaginary part of the tag impedance. The difference between the generated impedance phase angle and the tag impedance phase angle is calculated as an index of electrochemical phase consistency loss. The real part of the noise-reduced impedance is calculated based on the preset Cramer-Kroni relation to obtain the corresponding theoretical causal impedance imaginary part sequence. The deviation between the theoretical causal impedance imaginary part sequence and the noise-reduced impedance imaginary part is calculated as an electrochemical-physical causal constraint loss index. Calculate the knowledge distillation structural feature-level matching difference loss index between the teacher hidden layer mapping feature tensor and the set of multiple student stage hidden layer mapping feature tensors; Assign a first control weight coefficient to the network foundation self-reconstruction error loss index, a second control weight coefficient to the knowledge distillation structure feature-level matching difference loss index, a third control weight coefficient to the electrochemical phase consistency loss index, and a fourth control weight coefficient to the electrochemical physical causal constraint loss index. The weighted loss indexes of the network foundation self-reconstruction error, knowledge distillation structure feature-level matching difference, electrochemical phase consistency, and electrochemical physical causal constraint are added together to obtain the joint training feedback loss function. The student autoencoder model is obtained by iteratively updating and optimizing all deep decoupled separable convolution weights in the initial student network model using the backpropagation algorithm based on the joint training feedback loss function, and the student autoencoder model is obtained after detecting that the value of the joint training feedback loss function is minimized.