Battery early fault identification method and system based on acoustic fingerprints

Abstract

The invention belongs to the technical field of battery fault recognition, and provides a battery early fault recognition method and system based on acoustic fingerprints, and the method comprises the steps: collecting an output signal of a piezoelectric sensor on the surface of a battery at a high-frequency sampling rate, removing background noise in a mode of combining a time domain spectral subtraction method with a statistical threshold, and dynamically updating noise spectrum estimation; performing parallel calculation on short-time Fourier transform, continuous wavelet transform and Mel-frequency cepstral coefficients to form a feature tensor; learning a low-dimensional fingerprint vector through a sparse auto-encoder, wherein a loss function comprises a reconstruction error term and a sparse constraint term; and a mahalanobis distance or a single-class support vector machine is used to measure a deviation degree between the fingerprint vector and the health fingerprint database, a fault confidence coefficient is output, and when the confidence coefficient exceeds a preset threshold value, early fault early warning is triggered. According to the method, characterization processing can be carried out on the acoustic data of the battery within milliseconds, the fault confidence coefficient is given, and high-sensitivity detection, easy migration, early warning and real-time performance are achieved.

Description

Technical Field

[0001] This invention belongs to the field of battery fault identification technology, specifically relating to a method and system for early battery fault identification based on acoustic fingerprinting, which uses acoustic fingerprinting (AF) to identify early faults in secondary batteries such as lithium-ion batteries. Background Technology

[0002] Batteries are prone to early defects during service, such as electrode material expansion, current collector cracking, SEI film damage, and electrolyte decomposition. These defects often do not affect external voltage or current in their early stages, but they can generate weak mechanical vibrations or acoustic emission signals. Traditional electrochemical methods (such as dQ / dV and EIS) struggle to capture such transient information within milliseconds; while thermal imaging and other methods are limited by resolution and response time, making timely warnings difficult. Acoustic detection, due to its extreme sensitivity to mechanical stress and gas microbubbles, is considered an important supplement to battery safety monitoring. However, acoustic signals contain significant environmental noise and are highly correlated with battery type and assembly stress; without robust fingerprint algorithms, accuracy is difficult to guarantee. Therefore, it is necessary to propose a systematic acoustic fingerprint modeling and matching identification method to achieve accurate identification in the early stages of fault development. Summary of the Invention

[0003] The purpose of this invention is to overcome the existing defects and provide a method and system for early battery fault identification based on acoustic fingerprinting.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0005] The first objective of this invention is to provide a method for early battery fault identification based on acoustic fingerprinting, comprising:

[0006] (1) Acoustic signal acquisition: The output signal of the piezoelectric sensor on the battery surface is acquired at a high frequency sampling rate;

[0007] (2) Adaptive noise suppression: Background noise is removed by combining time-domain spectral subtraction with statistical threshold. Residual noise is suppressed by adjusting the over-subtraction coefficient and statistical threshold, and the noise spectrum estimate is dynamically updated to adapt to environmental changes.

[0008] (3) Construction of multi-resolution feature tensors: Parallel calculation of short-time Fourier transform, continuous wavelet transform and Mel cepstral coefficients to form multi-resolution time-frequency features, which are then spliced ​​together into feature tensors after being aligned in time.

[0009] (4) Fingerprint encoding: Low-dimensional fingerprint vectors are learned through sparse autoencoders. The loss function includes reconstruction error terms and sparse constraint terms. The target sparsity is set to a specific range to balance information preservation and model simplification.

[0010] (5) Anomaly calculation: The deviation of the fingerprint vector from the healthy fingerprint database is measured using Mahalanobis distance or single-class support vector machine, and the fault confidence is output. When the confidence exceeds the preset threshold, an early fault warning is triggered.

[0011] Furthermore, in step (2), let the original acoustic signal be s(n), and the background noise be estimated as... Spectral subtraction yields

[0012]

[0013] Where γ is the over-decrease coefficient, the adaptive threshold Suppress residual noise, where L is the frame length.

[0014] Furthermore, the background noise estimation steps include:

[0015] Silence segment detection: The stable segment is determined by the short-time energy E(k) and spectral entropy H(k). If E(k) ≤ θ E And H(k)≥θ H If so, it is considered a silent frame;

[0016] Noise spectrum initialization: Calculate the average spectrum during the first 100ms silent period as the initial noise estimate;

[0017] Dynamic update: The noise spectrum is estimated frame by frame using the minimum statistical method, so that the initial noise estimate is corrected as the environment changes.

[0018] Residual noise suppression: Noise estimation is subtracted from the observed spectrum, and a threshold is set to suppress residual noise.

[0019] Furthermore, the construction of the multi-resolution feature tensor includes:

[0020] STFT spectrum:

[0021]

[0022] in, Let w(·) represent the time-domain signal after spectral subtraction and noise reduction; w(·) represent the windowing function; L represent the length of each frame; R represent the shift between adjacent frames; and k represent the frequency index, k = 0, 1, ..., K. STFT -1; m represents the time frame index; X k,m This represents the STFT spectral coefficients at the m-th frame and the k-th frequency point.

[0023] CWT coefficient:

[0024]

[0025] in, ψ(t) represents the continuous-time signal after noise reduction; a represents the mother wavelet function; b represents the scale parameter, which controls the frequency resolution; and W(a,b) represents the translation parameter, which controls the time positioning.

[0026] MFCC vector:

[0027]

[0028] Where P(f) represents the power spectral density; H m (f) represents the frequency response of the m-th filter in the Mel filter bank; M represents the number of Mel filter banks; N represents the normalization factor; q represents the index of the cepstral coefficients; Q represents the number of MFCC dimensions used; C q This represents the q-th MFCC coefficient;

[0029] Align the STFT spectrum, CWT coefficients, and MFCC vector in time and concatenate them into a tensor.

[0030]

[0031] Where T represents the multiresolution feature tensor; K STFT K represents the STFT dimension; CWT Q represents the CWT feature dimension; Q represents the MFCC feature dimension; M represents the total number of frames.

[0032] Furthermore, the sparse autoencoder in step (4) is represented as:

[0033]

[0034] Where N is the number of samples; This represents the i-th input feature vector; is the sample reconstructed by the autoencoder; W,b is the neural network weight matrix and bias; is the weight coefficient of the sparse constraint; ρ represents the Kullback–Leibler divergence, used to constrain the average activation probability; ρ is the target sparsity, ρ = 0.05; This represents the average activation value of the j-th hidden unit;

[0035] Encoded output:

[0036] z i =f(x) i )=σ(Wx i +b)

[0037] Where σ(·) represents the Sigmoid activation function, z i This represents a sparse fingerprint.

[0038] Furthermore, in step (5), the Mahalanobis distance is expressed as:

[0039]

[0040] in, Represents the fingerprint vector to be tested; Represents the mean vector of the health fingerprint database; D represents the covariance matrix of the health fingerprint database; M (z) represents the degree of deviation of the fingerprint vector z from the health distribution; The threshold value is represented by a chi-square distribution with m degrees of freedom and a confidence level of 1-α.

[0041] A single-class support vector machine is represented as:

[0042]

[0043] Where w is the normal vector of the classification hyperplane; ξ i ρ is a slack variable that allows some outliers; ρ is the offset that determines the position of the hyperplane; v∈(0,1] is a parameter used to control the proportion of outliers; φ(·) is the kernel function mapping (RBF kernel is often chosen); N is the number of training samples.

[0044] Another object of the present invention is to provide a battery early fault identification system based on acoustic fingerprinting, comprising:

[0045] The acoustic signal acquisition module is used to acquire the output signal of the piezoelectric sensor on the battery surface at a high-frequency sampling rate.

[0046] The adaptive noise suppression module is used to remove background noise by combining time-domain spectral subtraction with statistical thresholding. It suppresses residual noise by adjusting the over-subtraction coefficient and statistical threshold, and dynamically updates the noise spectrum estimate to adapt to environmental changes.

[0047] The multi-resolution feature tensor construction module is used to perform parallel computation of short-time Fourier transform, continuous wavelet transform and Mel-frequency cepstral coefficients to form multi-resolution time-frequency features. After being aligned in time, these features are concatenated into a feature tensor.

[0048] The fingerprint encoding module is used to learn low-dimensional fingerprint vectors through a sparse autoencoder. The loss function includes a reconstruction error term and a sparse constraint term. The target sparsity is set to a specific range to balance information preservation and model simplification.

[0049] The anomaly calculation module is used to measure the deviation of fingerprint vectors from the healthy fingerprint database using Mahalanobis distance or single-class support vector machine, and outputs the fault confidence. When the confidence exceeds the preset threshold, an early fault warning is triggered.

[0050] Another object of the present invention is to provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the battery early fault identification method based on acoustic fingerprinting provided by the first object of the present invention.

[0051] Another object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method for early battery fault identification based on acoustic fingerprinting provided by the first object of the present invention.

[0052] Another object of the present invention is to provide a server including at least one processor and a memory communicatively connected to the processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the processor to cause the at least one processor to perform the battery early fault identification method based on acoustic fingerprinting provided in the first object of the present invention.

[0053] In combination with the above technical solutions, the beneficial effects of the present invention compared with the prior art are as follows:

[0054] This invention establishes a five-stage algorithm flow of "acquisition—noise reduction—feature encoding—fingerprint matching—anomaly measurement," which performs feature processing on battery acoustic data within milliseconds and provides fault confidence levels, enabling:

[0055] High-sensitivity detection: non-invasive, capable of identifying weak acoustic emission events with signal amplitudes only 1–2 dB of ambient noise;

[0056] Robustness and portability: Fingerprint codes can be migrated across models, reducing recalibration costs and facilitating migration;

[0057] Early warning: Potential defects can be detected when capacity decay is less than 2% and internal resistance increase is less than 5%, with a false alarm rate of less than 1.2%;

[0058] Real-time performance: Single frame processing time ≤ 5ms, adaptable to the 200Hz sampling requirements of vehicle BMS.

[0059] Applicability: This invention is applicable to online safety monitoring of power battery packs, energy storage power stations and consumer electronics batteries. It can be used as a BMS software upgrade module and can also be used in warehouse screening and extreme fast charging scenarios. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1This is a flowchart of the battery early fault identification method based on acoustic fingerprinting provided in the embodiments of the present invention;

[0062] Figure 2 This is a fault warning timing comparison diagram provided in the embodiments of the present invention;

[0063] Figure 3 This is a schematic diagram of the structure of the battery early fault identification system based on acoustic fingerprinting provided in an embodiment of the present invention. Detailed Implementation

[0064] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0065] Example 1:

[0066] like Figure 1 The image shows an embodiment of the battery early fault identification method and system based on acoustic fingerprinting provided by the present invention, which specifically includes the following steps:

[0067] S1: Acoustic signal acquisition: Acquire the output signal of the piezoelectric sensor on the battery surface at a high-frequency sampling rate;

[0068] S2: Adaptive noise suppression: Background noise is removed by combining time-domain spectral subtraction with statistical thresholding. Residual noise is suppressed by adjusting the over-subtraction coefficient and statistical threshold, and the noise spectrum estimate is dynamically updated to adapt to environmental changes.

[0069] S3: Construction of multi-resolution feature tensors: Parallel computation of short-time Fourier transform, continuous wavelet transform and Mel-frequency cepstral coefficients to form multi-resolution time-frequency features, which are then time-aligned and spliced ​​into feature tensors;

[0070] S4: Fingerprint encoding: Low-dimensional fingerprint vectors are learned through a sparse autoencoder. The loss function includes a reconstruction error term and a sparse constraint term. The target sparsity is set to a specific range to balance information preservation and model simplification.

[0071] S5: Anomaly Calculation: Use Mahalanobis distance or single-class support vector machine to measure the deviation of fingerprint vectors from the healthy fingerprint database, output the fault confidence, and trigger an early fault warning when the confidence exceeds a preset threshold.

[0072] Specifically, this invention proposes an early battery fault identification method based on acoustic fingerprinting, constructing a complete five-stage algorithm flow of "acquisition—noise reduction—feature encoding—fingerprint matching—anomaly measurement": high-frequency acoustic signal acquisition to ensure the capture of weak acoustic events; adaptive noise suppression (spectral subtraction + statistical threshold + dynamic update) to effectively remove time-varying background noise; multi-resolution feature tensor construction (parallel fusion of STFT, CWT, and MFCC) to retain multi-scale time-frequency features; sparse autoencoder fingerprint encoding to obtain a low-dimensional, transferable acoustic fingerprint vector; and anomaly calculation (Mahathano distance / single-class SVM) to output fault confidence and trigger an early warning.

[0073] The invention will be further illustrated below with specific examples.

[0074] 1. Acoustic signal acquisition: Acquire the output of the piezoelectric sensor on the battery surface at a sampling rate of over 200kHz;

[0075] Piezoelectric sensor selection and performance requirements:

[0076] Key performance indicators of piezoelectric sensors

[0077]

[0078] Signal conditioning chain: piezoelectric element → charge amplifier (10M, 20dB) → 5th order Butterworth bandpass (20k–800kHz) → 16-bit ADC (≥2MS / s) → DMA cache → MCU / DSP. Even with the sampling frequency set to 200kHz, the front-end bandwidth still retains a ×4 oversampling margin, which can reduce aliasing and support harmonic analysis.

[0079] 2. Adaptive noise suppression: Background noise is removed using time-domain Spectral Subtraction and statistical threshold η;

[0080] Let the original acoustic signal be s(n), and the background noise estimate be... Spectral subtraction yields

[0081]

[0082] Where γ is the over-decay coefficient. Adaptive threshold. Suppress residual noise (L is the frame length).

[0083] The over-reduction coefficient γ is determined by fixed calibration, adaptive adjustment, or gradual adjustment based on the residual noise variance.

[0084] In spectral subtraction, s(n) and These represent the original acoustic signal and the estimated background noise, respectively. They do not appear directly in the final frequency domain formula because after Fourier transform, they are converted into |S(ω)| and |S(ω)|, respectively.

[0085] The core principle of spectral subtraction is to subtract the noise-estimated spectrum from the amplitude spectrum |S(ω)| of the noisy signal. Furthermore, by combining the over-reduction coefficient γ with the adaptive threshold η for adjustment, the intrinsic acoustic characteristics of the battery are highlighted.

[0086] Therefore, s(n) provides complete information about the observed signal and is the input source for spectral subtraction; Providing a reference for environmental noise serves as the basis for subtraction; the Fourier transform results of both directly participate in spectral subtraction, determining the separation effect between effective acoustic information and noise.

[0087] (a) Fixed calibration (offline calibration)

[0088] In the laboratory anechoic chamber, using different signal-to-noise ratios (SNR)... in Sound data were collected for three types of power batteries: 10Ah, 50Ah, and 100Ah, ∈{-10,-5,0,5,10}dB. PESQ2 and SSNR (Seg.SNR) were used as objective indicators. The results show that when γ∈[1.6,2.2], PESQ can be improved by ≥0.4 points and SSNR can be improved by ≥7dB. Therefore, under static or semi-static conditions, γ can be fixed at γ=2 as the empirical optimal value.

[0089] (b) Adaptive adjustment based on posterior SNR

[0090] If the battery compartment noise exhibits significant time-varying characteristics, it can be introduced...

[0091] γ t (ω)=γ min +(γ max -γ min )exp(-ξ t (ω)),

[0092] in

[0093] The prior SNR is estimated using the Ephraim–Malah minimum mean square error estimate.

[0094] γ min =1.0,γ max =3.0 is the upper and lower limits of experience.

[0095] When ξ t When (ω)→0 (extremely low SNR), γ t (ω)→γ max To prevent residual noise; when ξ t When (ω)→∞ (high SNR), γ t (ω)→γ minThis avoids excessive attenuation of semantic information. In field tests, this strategy can improve SSNR by an additional 1.2–1.8 dB compared to a fixed γ.

[0096] (c) Gradual adjustment based on residual noise variance

[0097] Alternatively, it can be based on the residual noise power spectral density. Perform gradual decay:

[0098]

[0099] If the residual noise is still high after noise reduction If γ increases gradually, then γ decreases slowly.

[0100] Recommended implementation:

[0101] Static benchmarking – offline calibration γ=2;

[0102] Field Deployment – ​​Enable Adaptive Gamma t (ω), take γ min =1.0,γ max =3.0, and update ξ every 10 frames. t (ω).

[0103] Background noise estimation (MS + silent segment sampling)

[0104] Silent period detection: Within each 1s window, the electrochemical stable region is determined by the short-time energy E(k) and spectral entropy H(k); if E(k) ≤ θ E And H(k)≥θ H If it is, then it is considered a silent frame.

[0105] Noise spectrum initialization: Calculate the average spectrum N0(f) for the first 100ms silent segment.

[0106] Dynamic update: Minimum statistical method is used frame by frame.

[0107] N t+1 (f)=αN t (f)+(1-α)min(S t (f),γN t (f)), α=0.8, γ=1.5,

[0108] Where S t (f) represents the current power spectrum.

[0109] Spectral subtraction + statistical threshold η

[0110]

[0111] Dynamic threshold

[0112]

[0113] in This represents the standard deviation of the current noise spectrum. The more stable the background noise, the higher the η. t The lower the value, the better; when sudden structural noise occurs, η t Improving can prevent over-suppression.

[0114] 3. Multi-resolution feature extraction: Parallel computation of short-time Fourier transform (STFT), continuous wavelet transform (CWT) and Mel-frequency cepstral coefficients (MFCC) to form the original feature tensor;

[0115] 3.1 Construction of Multi-resolution Feature Tensors

[0116] (1) STFT (Short Time Fourier Transform)

[0117]

[0118] The time-domain signal after spectral subtraction and noise reduction;

[0119] w(·): Windowing function (such as Hamming window or Hanning window);

[0120] L: Length of each frame (number of sampling points);

[0121] R: Frame shift (i.e., the displacement between adjacent frames, usually R) <L);

[0122] k: Frequency index, k = 0, 1, ..., K STFT -1;

[0123] m: Time frame index;

[0124] X k，m : STFT spectral coefficients at the m-th frame and the k-th frequency point; K STFT STFT feature dimension (number of bandwidths).

[0125] (2) CWT (Continuous Wavelet Transform)

[0126]

[0127] The noise-reduced continuous-time signal (corresponding to the discretization) ); ψ(t): Mother wavelet function (such as Morlet or Morse wavelet);

[0128] a: Scale parameter, which controls frequency resolution;

[0129] b: Translation parameter, controls time positioning;

[0130] W(a, b): Wavelet coefficients at scale a and time b;

[0131] K CWT CWT feature dimension (number of scales).

[0132] (3) MFCC (Melch Cepstral Coefficients)

[0133]

[0134] P(f): Power spectral density;

[0135] H m (f): Frequency response of the m-th filter in the Mel filter bank; M: Number of Mel filter banks;

[0136] N: Normalization factor (usually M to ensure a suitable numerical scale for the coefficients); q: Index of the cepstral coefficients;

[0137] Q: Number of MFCC dimensions used;

[0138] C q : The qth MFCC coefficient.

[0139] (4) Feature tensor splicing

[0140]

[0141] T: Multiresolution feature tensor;

[0142] K STFT STFT dimension;

[0143] K CWT CWT dimensions;

[0144] Q: MFCC dimensions;

[0145] M: Total number of frames (time dimension length).

[0146] 3.2 Complementarity of the three types of features

[0147] STFT / CWT / MFCC Complementary Relationship

[0148]

[0149] 3.3 Parallel Computing and Synchronization Mechanisms:

[0150] Unified frame-level timestamps: The denoised signal is divided into L=512 frames and R=64 shift steps.

[0151] Shared cache queue: CPU writes to lock-free FIFO; GPU / DSP executes kernel_STFT→A in parallel. STFTkernel_CWT→C CWT kernel_MFCC→C MFCC .

[0152] Barrier synchronization: Assembled by frame number n

[0153]

[0154] Tensor stacking: spliced ​​along the time dimension And it is fed into the sparse autoencoder SAE.encode(X).

[0155] GPU parallelism can control the latency of single-frame feature calculation to <1ms, ensuring the real-time requirement of <5ms for the entire chain of single frames.

[0156] 4. Fingerprint Encoding (AF Encoding): Learning the mapping through a Sparse Autoencoder (SAE) Obtain the low-dimensional fingerprint vector z;

[0157] Sparse autoencoder minimization

[0158]

[0159] KL divergence sparsity constraint is applied, with a target sparsity ρ = 0.05, where:

[0160] N: Number of samples;

[0161] The i-th input feature vector;

[0162] Samples reconstructed using an autoencoder;

[0163] W, b: Neural network weight matrix and bias;

[0164] λ: Weighting coefficient of sparse constraints;

[0165] Kullback–Leibler divergence is used to constrain the average activation probability;

[0166] ρ = 0.05: Target sparsity;

[0167] The average activation value of the j-th hidden unit.

[0168] Encoded output:

[0169] z i =f(x) i )=σ(Wx i +b)

[0170] Where σ(·) represents the Sigmoid activation function, z i This represents a sparse fingerprint.

[0171] Determining the target sparsity ρ:

[0172] SAE loss

[0173]

[0174] Where KL is the Kullback–Leibler divergence.

[0175] Empirical prior: ρ∈[0.01, 0.1] can improve "sparse-selectivity" while maintaining information content; industrial anomaly detection often takes ρ=0.05 as a compromise.

[0176] Five-fold cross-validation: Reconstruction error, fingerprint clustering separability (Silhouette score), and anomaly detection AUC were used as joint metrics, with ρ∈{0.02,0.05,0.08,0.1}. The results are shown in Fig.4-3: The F1 score was highest (0.92) and training was stable when ρ=0.05, therefore ρ=0.05 was adopted.

[0177] Input size of a sparse autoencoder (SAE):

[0178] Tensors obtained in the multi-resolution feature extraction stage

[0179]

[0180] in:

[0181] T—Number of frames (e.g., L) seg =1s, L=512 points, R=64 points T = 301);

[0182] F s —STFT bandwidth (128);

[0183] F c —CWT scale number (64);

[0184] D-MFCC dimension (13).

[0185] The previous section described the construction of the feature tensor T (composed of STFT / CWT / MFCC). This section further quantifies the tensor's dimension, explaining how the input size of the SAE is determined by the feature extraction stage.

[0186] Flatten X frame by frame and concatenate them to obtain

[0187]

[0188] Output a definite value for the fingerprint dimension m.

[0189] The SAE structure is

[0190]

[0191] Candidate interval: take m∈{32, 64, 128, 256};

[0192] Evaluation metric: Reconstruction error Comparison with anomaly detection AUC;

[0193] Cross-validation: In the 5-fold Leave-One-Battery-Out experiment, when m=128, the reconstruction error drops to the saturation value and the AUC improves by <1% (see Fig. 4-2), so m=128 is finally selected.

[0194] The SAE loss function contains L1 sparse terms.

[0195]

[0196] This section employs L1 regularized sparsity constraints, which, while not entirely identical in form to the KL divergence sparsity constraints used previously, share the same fundamental goal: to obtain a low-dimensional fingerprint representation through reconstruction error plus sparsity constraints. The KL sparsity constraints used earlier theoretically better align with the classic definition of a sparse autoencoder; this section uses L1 sparsity terms, representing an implementation-level alternative. Functionally, both are equivalent, serving to "suppress the simultaneous activation of too many neurons and ensure fingerprint sparsity."

[0197] 5. Anomaly Calculation: Use Mahalanobis distance or one-class SVM to measure the deviation of z from the healthy fingerprint database and output the early failure confidence ρ.

[0198] 5.1 Mahalanobis distance

[0199]

[0200] Methods for calculating the mean μ and covariance matrix Σ:

[0201] Health fingerprint database Rapid anomaly removal (ρ<τ) low After that, calculate the parameters using the following formula:

[0202]

[0203] in:

[0204] Fingerprint vector to be tested;

[0205] The mean vector of the health fingerprint database;

[0206] The covariance matrix of the health fingerprint database;

[0207] D M (z): The degree of deviation of the fingerprint vector z from the health distribution;

[0208] The threshold follows a chi-square distribution with m degrees of freedom and a confidence level of 1-α.

[0209] Online update: If the new z satisfies ρ<τ low Then it is weighted by index.

[0210] μ t+1 =(1-β)μ t +βz, β = 10 -3

[0211] This recursion avoids the overhead of repeated inverse operations on large matrices while also taking into account long-term mean drift.

[0212] 5.2 Single-class SVM

[0213]

[0214] in:

[0215] w: The normal vector of the classification hyperplane;

[0216] ξ i Slack variables allow for some outliers;

[0217] ρ: Offset, determines the position of the hyperplane;

[0218] v∈(0,1]: Parameter used to control the proportion of outliers;

[0219] φ(·): Kernel function mapping (RBF kernel is often chosen);

[0220] N: Number of training samples.

[0221] This section presents the standard optimization form, which is the definition for the training phase of a single-class SVM. It describes the mathematical definition and training constraints of OCSVM.

[0222] Choice of kernel function φ(·) for a single-class SVM:

[0223] Candidate kernel functions: linear, polynomial, RBF (Gaussian), SIGMOID. The Leave-One-Battery-Out validation was used to evaluate AUC and false alarm rate (FPR). 95 ):

[0224]

[0225] Conclusion: RBF nuclear energy captures non-linear boundaries in high-dimensional sparse spaces, significantly reducing false alarms. Therefore,

[0226]

[0227] Healthy fingerprint database Construction:

[0228] Sample source: 20 power batteries with a rated capacity of 50–100 Ah, collected during the first 100 cycles of charge and discharge at 25 ± 2°C and 0.5C; 2-hour acoustic signals were randomly sampled from each battery, and after noise reduction and feature-fingerprint extraction, ~10 4 z items were obtained.

[0229] Initial library size:

[0230] Online update: If the new vector z satisfies ρ < τ low (τ low taking the 90th percentile), and the OCV and temperature are within the normal range during the same period, then add and update the mean μ and covariance ∑ with exponentially decaying weights (or incrementally train One-Class SVM).

[0231] Mahalanobis distance and One-Class SVM:

[0232] Mahalanobis distance (default)

[0233]

[0234] Setting of the threshold τ: Assume that the healthy fingerprints follow a distribution, and when m = 128, τ ≈ 167. If ρ MD > τ, trigger an "early fault warning".

[0235] One-Class SVM (optional)

[0236] ρ SVM = 1 - σ(f OCSVM (z)),

[0237] This is the formula for calculating the fault confidence level during the inference stage, that is, normalizing the discriminant function f OCSVM (z) of OCSVM through the Sigmoid function to obtain the confidence level in the range of [0, 1]. It is used to illustrate how to obtain an interpretable confidence level output from the trained model.

[0238] The kernel function used is RBF, with hyperparameters γ = 1 / m and v = 0.05. The threshold τ is determined by the 1-v quantiles on the training set.

[0239] Output early fault confidence:

[0240] Final confidence level

[0241]

[0242] If ρ>τ, then output an Alarm and record the fingerprint vector and timestamp for subsequent root cause analysis.

[0243] 5.3 Online update mechanism for the health fingerprint database:

[0244] Initial library construction:

[0245] Acoustic data was continuously collected from only healthy battery cells (capacity 50–100Ah) during the first 100 cycles at 25±2℃ and 0.5C charge / discharge; 2 hours of acoustic data were collected from each cell, resulting in a total of N0≈2.1×10⁻⁶. 5 For each fingerprint vector, calculate μ0 and ∑0 as basic parameters.

[0246] Incremental update:

[0247] Calculate ρ = MD(z) for the fingerprint z of the new frame:

[0248] ρ<τ low =0.9τ The concept of "health" is considered, and a weight β = 10 is used. -3 Recursion (μ,∑);

[0249] 0.9τ≤ρ<τ=5.4 Enter the observation pool; do not put it into storage yet.

[0250] ρ≥τ (or SVM confidence level <0.09) Alarm.

[0251] Wide-area operating condition coverage:

[0252] Random selection every month For batteries in service, 10-minute data points are collected for each of the following operating conditions: {-10, 25, 45}℃ and {0.1C, 1C, 3C}, to supplement the database, making ‖∑‖ F Drift <5% within six months.

[0253] Alarm threshold rules: In addition to the statistical threshold τ, an empirical lower limit is also given in the engineering implementation: "ρ≥0.9τ(≈0.9) triggers the warning", to ensure that the low signal-to-noise ratio scene can still respond quickly.

[0254] The following explanation will be further illustrated with specific experiments.

[0255] Early failure examples:

[0256] Battery cell model: NCM811 / 50Ah (soft pack), sampling rate 250kHz;

[0257] Fault induction: Applying 5C constant current fast charging (I=250A, V) at 80% SoC cut =4.20V), inducing lithium dendrite growth.

[0258] Acoustic sampling rate: f s =250kHz.

[0259] Result: D M The confidence level increased from 1.8 to 7.2 (threshold 5.4), with a single-class SVM confidence level of 0.91, providing an early warning 18 minutes earlier than traditional electrochemical indicators.

[0260] like Figure 2 As shown, to compare the alarm triggering time differences between the AF and EIS methods, the AF method (solid orange line) triggers the alarm (marked by the vertical line) at 12.4 minutes. Before triggering, it is in a "normal" state and remains in the "alarm" state after triggering. The EIS method (dashed orange line) triggers the alarm after 30.5 minutes. Before triggering, it is in a "normal" state and enters the "alarm" state after triggering.

[0261] Conclusion: AF can issue an early warning approximately 18 minutes earlier than EIS. This indicates that AF has higher sensitivity and can provide a more sufficient time window for system fault handling and safety protection.

[0262] Example 2:

[0263] like Figure 3 As shown, another objective of the present invention is to provide a battery early fault identification system based on acoustic fingerprinting, comprising:

[0264] The acoustic signal acquisition module is used to acquire the output signal of the piezoelectric sensor on the battery surface at a high-frequency sampling rate.

[0265] The adaptive noise suppression module is used to remove background noise by combining time-domain spectral subtraction with statistical thresholding. It suppresses residual noise by adjusting the over-subtraction coefficient and statistical threshold, and dynamically updates the noise spectrum estimate to adapt to environmental changes.

[0266] The multi-resolution feature tensor construction module is used to perform parallel computation of short-time Fourier transform, continuous wavelet transform and Mel-frequency cepstral coefficients to form multi-resolution time-frequency features. After being aligned in time, these features are concatenated into a feature tensor.

[0267] The fingerprint encoding module is used to learn low-dimensional fingerprint vectors through a sparse autoencoder. The loss function includes a reconstruction error term and a sparse constraint term. The target sparsity is set to a specific range to balance information preservation and model simplification.

[0268] The anomaly calculation module is used to measure the deviation of fingerprint vectors from the healthy fingerprint database using Mahalanobis distance or single-class support vector machine, and outputs the fault confidence. When the confidence exceeds the preset threshold, an early fault warning is triggered.

[0269] Example 3: Another object of the present invention is to provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the battery early fault identification method based on acoustic fingerprint provided by the first object of the present invention.

[0270] Example 4: Another object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the battery early fault identification method based on acoustic fingerprint provided by the first object of the present invention.

[0271] Example 5: Another object of the present invention is to provide a server, including at least one processor and a memory communicatively connected to the processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the processor to cause the at least one processor to perform the battery early fault identification method based on acoustic fingerprint provided in the first object of the present invention.

[0272] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in the present invention, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0273] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0274] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early battery failure identification based on acoustic fingerprint, characterized in that, The method comprises: (1) acoustic signal acquisition: collecting the output signal of the piezoelectric sensor on the surface of the battery at a high sampling rate; (2) adaptive noise suppression: removing background noise by combining time-domain spectral subtraction and statistical thresholding, suppressing residual noise by adjusting the subtraction coefficient and statistical threshold, and dynamically updating the noise spectrum estimate to adapt to environmental changes; (3) multi-resolution feature tensor construction: parallel computing short-time Fourier transform, continuous wavelet transform and mel-frequency cepstral coefficient, forming multi-resolution time-frequency features, aligning them by time, and concatenating them into a feature tensor; (4) fingerprint encoding: learning a low-dimensional fingerprint vector through a sparse autoencoder, the loss function including a reconstruction error term and a sparse constraint term, and the target sparsity being set to a specific range to balance information retention and model simplification; (5) anomaly degree calculation: using Mahalanobis distance or one-class support vector machine to measure the deviation of the fingerprint vector from the healthy fingerprint library, and outputting the fault confidence, which triggers early fault warning when the confidence exceeds a preset threshold. 2.The battery early failure identification method based on acoustic fingerprint according to claim 1, characterized in that, In step (2), let s(n) be the original acoustic signal and the background noise estimate be The spectral subtraction method gives where γ is an over-reduction factor, and the adaptive threshold suppressing the residual noise, L is the frame length. 3.The battery early failure identification method based on acoustic fingerprint according to claim 2, characterized in that, The background noise estimation step comprises: Silent section detection: Determine stable section by short-time energy E(k) and spectral entropy H(k), if E(k)≤θ E and H(k)≥θ H , then consider as silent frame; Noise spectrum initialization: average spectrum in the first 100 ms of silence is taken as the initial noise estimate; Dynamic update: recursively update the noise spectrum estimate frame by frame using the minimum statistics method, so that the initial noise estimate is corrected as the environment changes. Residual noise suppression: subtract the noise estimate from the observed spectrum and set a threshold to suppress residual noise. 4.The battery early failure identification method based on acoustic fingerprint according to claim 1, wherein, The multi-resolution feature tensor construction comprises: STFT spectrum: wherein, denotes the spectrum subtracted and denoised time-domain signal; w(·) denotes a windowing function; L denotes the length of each frame; R denotes the displacement of adjacent frames; k denotes the frequency index, k = 0, 1, …, K STFT - 1; m denotes the time frame index; X k,m denotes the STFT spectrum coefficient at the m-th frame, k-th frequency point. CWT coefficient: wherein, denotes the continuous-time signal after noise reduction; ψ(t) denotes the mother wavelet function; a denotes the scale parameter, which controls the frequency resolution; b denotes the translation parameter, which controls the time positioning; W(a, b) denotes the wavelet coefficient at scale a and time b. MFCC vector: where P(f) represents the power spectral density; H m (f) represents the frequency response of the mth filter in the mel filter bank; M represents the number of mel filter banks; N represents a normalization factor; q represents the index of the cepstral coefficient; Q represents the number of MFCC dimensions taken; C q represents the qth MFCC coefficient; Aligning the STFT spectrum, CWT coefficient and MFCC vector by time, and concatenating them into a tensor where T denotes the multi-resolution feature tensor; K STFT denotes the STFT dimension; K CWT denotes the CWT feature dimension; Q denotes the MFCC feature dimension; and M denotes the total number of frames. 5.The battery early failure identification method based on acoustic fingerprint according to claim 1, wherein, The sparse autoencoder in step (4) is represented as: where N is the number of samples; represents the i-th input feature vector; is the sample reconstructed by the autoencoder; W, b are the neural network weight matrix and bias; is the weight coefficient of the sparsity constraint; represents the Kullback-Leibler divergence, used to constrain the average activation probability; p is the target sparsity, p = 0.05; represents the average activation value of the j-th hidden unit; Encoding output: z i = f(x i ) = σ(Wx i + b) where σ(·) denotes a Sigmoid activation function, z i is a sparse fingerprint representation. 6.The battery early failure identification method based on acoustic fingerprint according to claim 1, wherein, In step (5), the Mahalanobis distance is represented as: wherein, represents a fingerprint vector to be tested; represents a mean vector of the healthy fingerprint library; represents a covariance matrix of the healthy fingerprint library;D M (z) represents the degree of deviation of the fingerprint vector z from the healthy distribution; represents a threshold value, subject to a chi-square distribution with m degrees of freedom and a confidence level of 1 - a; The one-class support vector machine is represented as: where w is the normal vector of the classification hyperplane; ξ i is the slack variable, allowing partial outliers; p is the offset, determining the position of the hyperplane; v e (0, 1] is the parameter, used to control the proportion of outliers; f (·) is the kernel function mapping (usually RBF kernel); N is the number of training samples.

7. An acoustic fingerprint-based early battery failure identification system, comprising: The system comprises: An acoustic signal acquisition module for collecting the output signal of the piezoelectric sensor on the surface of the battery at a high sampling rate; An adaptive noise suppression module for removing background noise by combining time-domain spectral subtraction and statistical thresholding, suppressing residual noise by adjusting the subtraction coefficient and statistical threshold, and dynamically updating the noise spectrum estimate to adapt to environmental changes; A multi-resolution feature tensor construction module for parallel computing short-time Fourier transform, continuous wavelet transform and mel-frequency cepstral coefficient, forming multi-resolution time-frequency features, aligning them by time, and concatenating them into a feature tensor; A fingerprint encoding module for learning a low-dimensional fingerprint vector through a sparse autoencoder, the loss function including a reconstruction error term and a sparse constraint term, and the target sparsity being set to a specific range to balance information retention and model simplification; An anomaly degree calculation module for using Mahalanobis distance or one-class support vector machine to measure the deviation of the fingerprint vector from the healthy fingerprint library, and outputting the fault confidence, which triggers early fault warning when the confidence exceeds a preset threshold.

8. An electronic device comprising a processor and a memory having a computer program stored therein, characterized in that, The processor executes the computer program to implement the acoustic fingerprint-based battery early fault identification method of any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the acoustic fingerprint-based battery early fault identification method of any one of claims 1 to 6.

10. A server, characterized by: The application discloses a method for identifying early battery failure based on an acoustic fingerprint, and a device thereof. The method comprises at least one processor and a memory connected with the processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the processor to enable the at least one processor to execute the method for identifying early battery failure based on an acoustic fingerprint according to any one of claims 1 to 6.

Images