Neural network driven battery acoustic pressure monitoring and early warning method and system

Abstract

The invention discloses a neural network-driven battery acoustic pressure monitoring and early warning method and system, and aims to realize omnibearing analysis and fault early warning of battery pressure signals by constructing a multi-stage deep learning framework. According to the method, a neural network coding technology is adopted, and causal features of pressure data are extracted through reverse time evolution and chaotic attractor projection; identifying a pressure periodic rhythm by using a time convolutional network, establishing a rhythm reference model and generating an abnormal index; a mutual information resonance mechanism is introduced, and the information correlation degree between data is determined through phase slip analysis; a structured feature map is generated through variational echo reconstruction, and the evolutionary process of the features is captured through a diffusion model; predicting a phase change critical point based on a reformed group theory, converting early warning parameters into an oscillator network, and analyzing synchronization characteristics; and finally, graded early warning is realized through singular value decomposition and activation intensity calculation, so that early abnormal symptoms of the battery can be accurately identified, and reliable technical support is provided for safety management of the battery.

Description

Technical Field

[0001] The present invention relates to the technical field of battery safety monitoring, and in particular to a neural network-driven battery acoustic pressure monitoring and early warning method and system. Background Art

[0002] With the rapid development of the new energy industry, battery system safety has become a key factor restricting its large-scale application. Batteries generate complex pressure fluctuations during charging and discharging, and these pressure signals contain important information about the battery's internal state. However, traditional pressure monitoring methods rely primarily on fixed thresholds and simple statistical analysis, making it difficult to capture subtle patterns and early signs of failure in pressure signals. Consequently, the accuracy and timeliness of early warnings fall short of meeting practical needs.

[0003] Current battery monitoring technology has obvious deficiencies when processing high-dimensional, nonlinear pressure data. Existing methods usually use a single analysis method in the time domain or frequency domain, lack comprehensive consideration of the multi-scale characteristics of the pressure signal, and cannot effectively identify abnormal evolution processes on different time scales. At the same time, traditional technologies have difficulty establishing a complex mapping relationship between pressure changes and fault development, and lack the ability to predict system state transitions, so that fault warnings often lag behind the occurrence of actual risks. In addition, existing monitoring systems mostly use independent sensor data processing, ignoring the correlation between multiple monitoring points, and cannot form a comprehensive understanding of the overall status of the battery. Therefore, there is an urgent need to develop an intelligent battery pressure monitoring method to achieve early warning and accurate identification of battery faults. Summary of the Invention

[0004] This invention discloses a neural network-driven battery acoustic pressure monitoring and early warning method and system, designed to achieve intelligent analysis of battery pressure signals and fault warning through deep learning technology. This method extracts deep features from pressure data through neural network encoding, identifies abnormal patterns in pressure cycle rhythms, exploits mutual information maximization to explore intrinsic correlations between data at different times, employs variational echo reconstruction technology to generate feature representations rich in structural information, and predicts the system's critical state based on phase transition theory. Ultimately, a multi-level early warning decision-making network is constructed to achieve hierarchical early warning and intelligent decision-making for battery failures.

[0005] A first aspect of the present invention provides a neural network-driven battery acoustic pressure monitoring and early warning method, comprising the following steps: Obtaining pressure monitoring data during battery operation, and performing neural network encoding on the pressure monitoring data to generate a training data set; Constructing a temporal convolutional neural network based on the training data set, training the temporal convolutional neural network to identify pressure cycle rhythms, extracting rhythm parameters from the pressure cycle rhythms to generate a rhythm reference model, and performing deviation analysis on the rhythm reference model to generate a rhythm abnormality indicator; Performing mutual information maximization processing on the training data set to establish a resonance learning network, determining information correlation of data at different times through the resonance learning network, identifying a resonance frequency pattern using the information correlation, and amplifying the rhythm abnormality indicator based on the resonance frequency pattern to generate a mutual information resonance vector; performing variational echo reconstruction on the mutual information resonance vector to generate a variational coding feature, performing probability modeling on the variational coding feature to determine a probability distribution of echo propagation, sampling and reconstructing a propagation path from the probability distribution, and extracting structural change features based on the propagation path to obtain a variational echo feature map; Inputting the rhythm abnormality index and the mutual information resonance vector into a deep classification network to generate a nonlinear conversion pattern, predicting a phase change critical point from the nonlinear conversion pattern, determining a phase change warning parameter based on the phase change critical point, and constructing a warning decision network using the phase change warning parameter; The early warning decision network integrates the variational echo characteristic graph to output a graded early warning signal, thereby completing the neural network-driven battery acoustic pressure monitoring and early warning.

[0006] A second aspect of the present invention provides a neural network-driven battery acoustic pressure monitoring and early warning system, comprising: A data encoding module is used to obtain pressure monitoring data during battery operation and perform neural network encoding on the pressure monitoring data to generate a training data set; a rhythm analysis module, configured to construct a temporal convolutional neural network based on the training data set, train the temporal convolutional neural network to identify pressure cycle rhythms, extract rhythm parameters from the pressure cycle rhythms to generate a rhythm reference model, and perform deviation analysis on the rhythm reference model to generate a rhythm abnormality indicator; a resonance learning module, configured to perform mutual information maximization processing on the training data set to establish a resonance learning network, determine information correlation of data at different times through the resonance learning network, identify a resonance frequency pattern using the information correlation, and amplify the rhythm abnormality indicator based on the resonance frequency pattern to generate a mutual information resonance vector; a variational reconstruction module, configured to perform variational echo reconstruction on the mutual information resonance vector to generate a variational coding feature, perform probability modeling on the variational coding feature to determine a probability distribution of echo propagation, sample and reconstruct a propagation path from the probability distribution, and extract structural change features based on the propagation path to obtain a variational echo feature map; a phase change prediction module, configured to input the rhythm abnormality index and the mutual information resonance vector into a deep classification network to generate a nonlinear conversion pattern, predict a phase change critical point from the nonlinear conversion pattern, determine a phase change warning parameter based on the phase change critical point, and construct a warning decision network using the phase change warning parameter; The early warning output module is used to output a graded early warning signal by integrating the variational echo characteristic graph through the early warning decision network, thereby completing the battery acoustic pressure monitoring and early warning driven by the neural network.

[0007] The beneficial effects of the present invention are reflected in the following aspects: 1. By combining neural network encoding with chaotic dynamics analysis, it can capture weak perturbations and nonlinear evolution characteristics in the pressure signal. The deviation analysis mechanism of the rhythmic baseline model can identify subtle changes in the normal periodic rhythm, detecting abnormal signs in the early stages of fault formation. This multi-level feature extraction method enables comprehensive monitoring of small pressure fluctuations, slow trend changes, and sudden anomalies, advancing the anomaly detection window from the onset of the fault to the incipient stage. 2. A data association network established through the mutual information resonance mechanism can track the propagation path of anomalies between different monitoring points and predict the development trend of the fault. Variational echo reconstruction technology enhances the expressiveness of features through multiple information echoes, and the probabilistic modeling of the diffusion model provides a quantitative assessment of uncertainty. The introduction of phase transition theory enables the system to accurately predict the arrival of critical states, achieving a transition from passive response to active prevention. 3. Synchronous analysis of the oscillator network reveals the collective dynamic behavior of the system, making early warning decisions clearly physical and interpretable. The five-level hierarchical early warning system can provide differentiated treatment recommendations based on the type and severity of the anomaly, achieving precise policy implementation. The graph structure design of the early warning decision network retains the topological characteristics of the failure mode, enabling the system to adaptively handle new failure modes and improving the stability of long-term operation.

[0008] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The accompanying drawings herein illustrate specific examples of the technical solutions described in the present invention, and together with the specific implementation methods constitute a part of the specification, and are used to explain the technical solutions, principles and effects of the present invention.

[0010] Unless otherwise specified, the same reference numerals in different drawings represent the same or similar technical features, and the same or similar technical features may also be represented by different reference numerals.

[0011] Figure 1 It is a flow chart of a neural network driven battery acoustic pressure monitoring and early warning method of the present invention.

[0012] Figure 2 This is a structural block diagram of a neural network-driven battery acoustic pressure monitoring and early warning system of the present invention. DETAILED DESCRIPTION

[0013] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0014] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0015] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0016] The technical solutions of the embodiments of this application are introduced below.

[0017] like Figure 1 As shown, an embodiment of the present invention provides a neural network driven battery acoustic pressure monitoring and early warning method, comprising the following steps S110 to S160: Step S110 , obtaining pressure monitoring data during battery operation, and performing neural network encoding on the pressure monitoring data to generate a training data set.

[0018] Specifically, pressure monitoring data is obtained during battery operation. Ultrasonic detection sensors are placed on the surface of the battery casing, using non-invasive ultrasonic technology to monitor internal pressure changes in real time. The ultrasonic detection system utilizes high-frequency ultrasonic probes, with one ultrasonic sensor positioned at the center of each of the top, bottom, and four side surfaces of the battery cell, forming a stereo acoustic monitoring network. The ultrasonic probes are made of piezoelectric ceramic material and operate within a frequency range of 5-10 MHz, ensuring that ultrasonic waves can penetrate the battery casing and achieve sufficient detection accuracy. Ultrasonic detection is based on the acoustoelastic effect. Changes in internal battery pressure alter the density and elastic modulus of the battery material, which in turn affects the propagation speed, attenuation, and frequency characteristics of ultrasonic waves within the material. The detection system transmits ultrasonic pulses and receives reflected echoes, measuring the ultrasonic wave's propagation time, amplitude change, and frequency offset. These acoustic parameters are converted into pressure values ​​using acoustoelastic calibration curves. The sensors maintain good acoustic contact with the battery casing via a coupling agent. The ultrasonic signal is processed by a preamplifier and a bandpass filter before being digitized using a high-speed ADC. The acquisition system continuously records acoustic parameters at high frequencies in the kilohertz range, ensuring the capture of transient pressure fluctuations and subtle pressure changes. The pressure monitoring data includes information in four dimensions: timestamp, sensor location identifier, acoustic time difference, and amplitude ratio. A temperature compensation algorithm is implemented during data acquisition to eliminate the effects of temperature changes on ultrasonic propagation velocity. The collected raw acoustic data undergoes preprocessing, including outlier removal, missing value interpolation, and noise filtering, before being converted to pressure values ​​using an acoustoelastic calibration curve to ensure data quality. It should be noted that the acoustoelastic calibration curve is a mapping of acoustic parameters to pressure established through experiments or theoretical models. It converts measured acoustic parameters (including acoustic time difference, amplitude attenuation, and frequency offset) into actual pressure values. Specifically, the experimental calibration steps for the acoustoelastic calibration curve include applying a known pressure gradient to the battery in a controlled environment (such as a constant temperature chamber); measuring the corresponding acoustic parameters (including acoustic time difference, amplitude attenuation, and frequency offset) at each pressure point; and fitting the data to obtain a curve equation (polynomial), namely the acoustoelastic calibration curve.

[0019] In some embodiments, the neural network encoding of the pressure monitoring data to generate a training data set includes: performing reverse time evolution on the pressure monitoring data to obtain a causal precursor signal; applying chaotic attractor projection to the causal precursor signal to extract singular trajectories; using the singular trajectories as neuron dead zone activation patterns to generate sparse coding; and constructing a training data set based on the sparse coding.

[0020] The pressure monitoring data is subjected to reverse time evolution to obtain the causal precursor signal. A time reversal operator is constructed to perform a reverse integration operation on the pressure monitoring data P(t): P_pre(τ)=∫[t→t-Δt]P(t)·exp(-λ(t-τ))dt, where P_pre(τ) is the causal precursor signal, P(t) is the pressure data at time t, λ is the decay coefficient, τ is the reverse time variable, t is the current time, and Δt is the time window length. The reverse time evolution is calculated using a high-order numerical integration method, with adaptive step-size control ensuring numerical stability and accuracy. An exponential decay factor is introduced during the integration process to give greater weight to the influence of recent historical data on the current state, while the influence of more distant data gradually decreases. By reverse time evolution, the current pressure anomaly is traced back to the time of its cause, obtaining a precursor signal sequence that contains causal relationships. The precursor signal not only preserves the historical dependence of pressure changes but also reveals the complete dynamic process of gas generation, accumulation, and release within the battery. A sliding window mechanism is used in the calculation process, and appropriate overlap is maintained between windows to ensure that continuous causal evolution features are captured and information loss is avoided.

[0021] A chaotic attractor projection is applied to the causal precursor signal to extract singular trajectories. The causal precursor signal is mapped into the Lorenz attractor phase space using a classical dynamical system consisting of three coupled differential equations. This system describes the interaction of three state variables: the first equation states that the rate of change of x is proportional to the difference between y and x; the second equation states that the rate of change of y is composed of the nonlinear coupling term between x and z and the attenuation term of y; and the third equation states that the rate of change of z is determined by the product term of x and y and the attenuation term of z. Standard Lorenz parameter values ​​are used for the system parameters. The normalized precursor signal and its first- and second-order derivatives are used as the initial conditions in the phase space, and the complete phase space trajectory is generated through iterative calculation. The trajectory S(t) is calculated using the fourth-order Runge-Kutta method to ensure the numerical accuracy of the dynamical system evolution. In the generated phase space trajectory, the deviation D = ||r(t)-r_center|| / ||r_center|| is calculated for each trajectory point, where r(t) is the position vector of the current trajectory point and r_center is the attractor center position. When the deviation exceeds a dynamic threshold, the trajectory segment is marked as a singular trajectory. Singular trajectory identification uses a local density adaptive method, raising the threshold in dense trajectory areas and lowering it in sparse areas to ensure the sensitivity of anomaly detection. The extracted singular trajectory sequence fully preserves the nonlinear dynamic characteristics of the pressure anomaly evolution, which reflect early signs of potential faults such as imbalanced chemical reactions within the battery and abnormal gas generation.

[0022] Sparse codes are generated by using singular trajectories as neuronal dead-zone activation patterns. An activation function with a dynamic dead zone is designed: f(x)=max(0,x-θ). The dead-zone threshold θ is dynamically determined by the statistical properties of the singular trajectories: θ=mean(S(t))+2×std(S(t)), where S(t) is the sequence of singular trajectories. This activation function only responds to abnormal signals that exceed the normal fluctuation range, thus automatically screening for abnormal pressure features. After processing the singular trajectories through the activation function, the resulting sparse activation vectors are fed into a multi-layer sparse autoencoder. The encoder employs a layer-by-layer decreasing network structure, with the number of neurons in each layer decreasing according to a specific ratio, achieving gradual feature compression and refinement. L1 norm regularization is used during the encoding process to strictly control sparsity, ensuring that only a few key features are activated in the final encoded vector. Each activated non-zero element corresponds to a specific pressure anomaly pattern, such as a transient pressure spike, a slow pressure climb, or periodic pressure oscillations. Sparse coding not only significantly reduces the feature dimension, but also enhances the interpretability of abnormal features. Each coding dimension has a clear physical meaning.

[0023] A training dataset is constructed based on sparse coding. The generated sparse coding vectors are precisely matched with the corresponding battery state labels to form a complete training sample. The battery state labeling system covers detailed state categories, including normal operation, mild overcharge, severe overcharge, mild overdischarge, deep overdischarge, initial stage of internal short circuit, developing stage of internal short circuit, and precursor to thermal runaway. State labels are determined using a multi-parameter joint judgment method that comprehensively considers multi-dimensional monitoring information such as voltage, current, temperature, and internal resistance. An accurate labeling system is established based on expert rules and historical failure cases. High-density sampling is performed for each complete charge and discharge cycle, focusing on data during state transitions to ensure that key transition features are captured. The sampling strategy uses an adaptive method, reducing the sampling density during periods of stable pressure fluctuations and increasing it during periods of anomalies. Data augmentation techniques such as timeline stretching, compression, and translation are used to simulate pressure variation patterns under different operating conditions, expanding the original sample size several times. The augmented dataset is divided into training, validation, and test sets in a strict ratio to ensure scientific research in model training and evaluation.

[0024] Step S120, constructing a temporal convolutional neural network based on the training data set, training the temporal convolutional neural network to identify the pressure cycle rhythm, extracting rhythm parameters from the pressure cycle rhythm to generate a rhythm benchmark model, and performing deviation analysis on the rhythm benchmark model to generate a rhythm abnormality indicator.

[0025] Specifically, a temporal convolutional neural network (TCNN) was constructed based on the training dataset. Using the training dataset as input, a specially designed one-dimensional temporal convolutional architecture was employed to process the pressure sequence data. The network input layer received a 256-dimensional sparsely coded vector sequence with a length of 128 consecutive time steps, covering the complete pressure variation cycle. The first temporal convolution layer was configured with 64 filters and a kernel size of 7. A causal convolution mechanism was used to ensure temporal causality and avoid the use of future information. Dilated convolutions were employed to expand the receptive field layer by layer, with the dilation rate increasing exponentially through 1, 2, 4, 8, and 16, enabling the network to capture pressure variation patterns at different time scales. A batch normalization layer was added after each convolutional layer to stabilize the training process. Nonlinearity was introduced using the ReLU activation function, and residual connections were used to mitigate gradient issues in deep networks. The network consists of six temporal convolutional blocks, each of which employs a gated linear unit to regulate information flow and selectively transmit important temporal features. The gating mechanism uses a sigmoid function to control the opening and closing of information channels, enabling adaptive filtering of different frequency components. Finally, the time series features are aggregated through global average pooling and connected to a two-layer fully connected network to output the pressure period feature vector. The total number of network parameters is kept to around two million, achieving a balance between model capacity and computational efficiency.

[0026] A temporal convolutional neural network was trained to identify periodic pressure rhythms. The constructed network was trained end-to-end using the preprocessed training dataset. A composite loss function, L = L_ce + α·L_period + β·L_reg, was designed. L_ce represents the multi-class cross-entropy loss, L_period represents the periodic consistency constraint loss, and L_reg represents the weight regularization term. α and β are balancing coefficients set to 0.3 and 0.01, respectively. The periodic consistency loss constrains the network to learn stable periodic features by calculating the phase difference between the predicted and true rhythms. Training was performed using the Adam optimizer with an initial learning rate of 0.001 and momentum parameters β1 = 0.9 and β2 = 0.999. The learning rate was adjusted using a cosine annealing strategy, with each annealing cycle completed every 50 epochs. The batch size was set to 256 samples, and the total number of training epochs was 200. Model performance was evaluated on the validation set every five epochs, monitoring the loss curve and accuracy. Early stopping was triggered when the validation loss stopped decreasing for 10 consecutive epochs to prevent overfitting. Through sufficient training, the network learns to automatically identify and extract multiple pressure cycle patterns from sparse coding sequences, including multi-scale rhythmic features such as the main cycle synchronized with charging and discharging, the secondary cycle caused by the breathing effect of the electrode material, and the high-frequency microcycle caused by electrolyte convection.

[0027] Rhythm parameters were extracted from periodic pressure rhythms to generate a rhythmic benchmark model. Time-frequency analysis of the periodic pressure rhythm was performed using a short-time Fourier transform (SFT) with a window length of 32 time steps and a 50% overlap rate to obtain the time-varying spectrum of the pressure signal. Frequency components with concentrated energy were identified from the spectrum, and the dominant rhythm frequency f_i and its harmonic components were determined. The main rhythm types identified included the basic charge-discharge rhythm f_base (period 2-4 hours), the electrode breathing rhythm f_breath (period 10-30 minutes), the electrolyte convection rhythm f_conv (period 1-5 minutes), and the interface reaction rhythm f_interface (period 10-60 seconds). A complete set of characteristic parameters was extracted for each rhythm: frequency f, amplitude A, initial phase φ, duty cycle D, modulation depth m, and frequency stability σ_f. Rhythm parameter extraction employed a combination of adaptive bandpass filtering and Hilbert transform to ensure accurate parameter estimation. All rhythm parameters were organized into a structured parameter matrix P∈R^(n×6), where n is the number of identified rhythms. Through statistical analysis of tens of thousands of samples under normal operating conditions, kernel density estimation was used to establish probability distribution models for each rhythm parameter. This distribution model employs a Gaussian mixture form, with the number of mixture components automatically determined using the Bayesian Information Criterion. By integrating the parameter distribution model, normal range thresholds, and abnormality determination rules, a complete rhythm baseline model was constructed.

[0028] In some embodiments, the deviation analysis performed on the rhythm reference model to generate a rhythm abnormality indicator includes: activating the self-diagnosis mode of the rhythm reference model to obtain an internal parameter distribution map; identifying weight imbalance areas on the parameter distribution map to establish an imbalance pattern classification, wherein the imbalance pattern classification includes gradient explosion type, gradient vanishing type, and oscillation divergence type; performing severity assessment processing on the imbalance pattern classification to generate a rhythm abnormality indicator.

[0029] Activate the self-diagnosis mode of the rhythm benchmark model to obtain an internal parameter distribution map. Switch the rhythm benchmark model from conventional inference mode to self-diagnosis mode. In this mode, the model not only outputs rhythm analysis results but also monitors and records detailed information about the internal computational process in real time. By registering forward hook functions at each network layer, we intercept intermediate results such as the input tensor, weight matrix, and activation output of each layer. Comprehensive statistical analysis is performed on the intercepted tensor data, calculating statistics such as mean μ, standard deviation σ, maximum, minimum, 25th and 75th percentiles, kurtosis, and skewness. Statistical indicators for each layer are organized into a multidimensional array, and a heatmap visualization technique is used to generate a panoramic view of the parameter distribution. The distribution map uses a logarithmic color scale to enhance visual sensitivity to abnormal parameter changes. The horizontal axis arranges the layers according to network depth, while the vertical axis displays different statistical indicators, with light and dark colors indicating numerical magnitude. Simultaneously, activation value histograms are plotted to analyze the activation rate distribution of neurons in each layer and identify the proportion of dead neurons (with a constant activation value of 0) and saturated neurons (with a constant activation value of 0). The high-dimensional weight space is projected onto a two-dimensional plane through the t-SNE dimensionality reduction algorithm to observe the evolution trajectory and clustering pattern of the parameters during the training process.

[0030] The parameter distribution graph is used to identify regions of weight imbalance and establish a classification of imbalance patterns. Computer vision algorithms are applied to the generated parameter distribution graph to automatically detect anomalous regions. A 5×5 sliding window is used to traverse the entire distribution graph, and the statistical anomaly score within each window is calculated. The imbalance metric B=σ² / μ² is defined, where σ is the standard deviation of the parameter within the window and μ is the parameter mean. Detected anomalous regions are classified into three patterns based on the imbalance metric. Exploding gradient imbalance is characterized by B > 10, with exponential growth of parameter values, an extremely long-tailed distribution, and a large number of extremely large values ​​in the histogram. Vanishing gradient imbalance is characterized by B < 0.01, with parameter values ​​rapidly decaying toward zero, a highly concentrated distribution near zero, and severe obstruction of information transfer between network layers. Oscillatory divergent imbalance is characterized by multiple, isolated peaks in the parameter distribution, with inter-peak distances exceeding two standard deviations, and parameter values ​​repeatedly jumping between different orders of magnitude. Feature vectors are extracted for each imbalance pattern, including attributes such as the area of ​​the imbalance region, center location, diffusion rate, and impact depth. An imbalance pattern library is established to store characteristic templates of typical imbalance cases. The current detection results are matched with the template library through cosine similarity to automatically determine the imbalance type.

[0031] Imbalance pattern classification is processed for severity assessment to generate rhythm abnormality indicators. A quantitative severity assessment system is designed for each type of identified imbalance pattern. The severity calculation formula for the exploding gradient type is: S_exp = log(max(|W|) / threshold_exp), where |W| is the maximum absolute value of the weight, and threshold_exp = 100 is the preset explosion threshold. When the maximum weight exceeds the threshold, the severity increases logarithmically. The severity calculation formula for the vanishing gradient type is: S_van = -log(mean(|W|) / threshold_van), where mean(|W|) is the mean of the absolute values ​​of the weights, and threshold_van = 0.001 is the vanishing threshold. A smaller mean weight indicates a higher severity. The severity calculation formula for the diverging oscillation type is: S_osc = N_peaks × D_peaks / D_ref, where N_peaks is the number of detected peaks, D_peaks is the average inter-peak distance, and D_ref is the reference distance. Considering the varying impact of different imbalance patterns on rhythm recognition accuracy, weight coefficients w1 = 0.5, w2 = 0.3, and w3 = 0.2 were set. The composite severity score was calculated as: S_total = w1 × S_exp + w2 × S_van + w3 × S_osc. The severity score was mapped to the standard interval [0, 1] using the sigmoid function: RAI = 1 / (1 + exp(-k × (S_total - S_mid))), where k = 2 is the steepness parameter and S_mid = 5 is the midpoint parameter. The resulting rhythm abnormality index, RAI, not only quantifies the internal parameter health of the model but also reflects the reliability of pressure rhythm recognition.

[0032] Step S130 , performing mutual information maximization processing on the training data set to establish a resonance learning network, determining the information correlation of data at different times through the resonance learning network, identifying the resonance frequency pattern using the information correlation, and amplifying the rhythm abnormality index based on the resonance frequency pattern to generate a mutual information resonance vector.

[0033] Specifically, a resonant learning network is established by maximizing mutual information on the training dataset. A variational mutual information estimation method is used to process time series sample pairs in the training dataset, constructing a resonant learning network with a two-tower architecture. The left tower receives data x(t) at time t, and the right tower receives data x(t+τ) at time t+τ, where τ is the time delay parameter. Each tower contains three encoding layers with 128, 64, and 32 neurons, respectively, and uses the LeakyReLU activation function to maintain gradient flow. The feature vectors output by the two towers are passed through a bilinear layer to calculate the mutual information estimate: I(x,y)=E[log(f(x,y) / f(x)f(y))], where f is the ratio of the joint distribution to the marginal distribution. The training objective is to maximize the mutual information under different time delays, and the loss function adopts a noise contrastive estimation form. By comparing the score differences between positive sample pairs (real time series pairs) and negative sample pairs (randomly shuffled time series pairs), the network learns to capture the inherent correlation patterns between time series data. During the resonance learning process, an adaptive temperature parameter was introduced to adjust the difficulty of contrastive learning. The initial temperature value was set to 0.07 and dynamically adjusted as training progressed. The network was trained using the SGD optimizer with a learning rate of 0.01, momentum of 0.9, and a batch size of 512. After sufficient training, the resonance learning network was able to accurately quantify the degree of information dependence between data at any two moments in time.

[0034] In some embodiments, determining the information correlation of data at different times through the resonance learning network includes: mapping the training data set to a complex domain to construct an analytical signal; performing a Hilbert-Huang transform on the analytical signal to obtain an instantaneous phase; calculating a phase slip rate based on the instantaneous phase; and using a stability index of the phase slip rate as the information correlation.

[0035] The training dataset is mapped to the complex domain to construct an analytical signal. A Hilbert transform is applied to the real-valued pressure signal x(t) in the training dataset to generate the corresponding imaginary signal: y(t) = H[x(t)] = (1 / π)·PV∫[-∞,+∞]x(τ) / (t-τ)dτ, where PV represents the Cauchy principal value integral. The real and imaginary components are combined to form a complex analytical signal z(t) = x(t) + j·y(t), where j is the imaginary unit. The analytical signal is constructed using the fast Hilbert transform algorithm, with the computational process accelerated by the FFT. The signal is preprocessed, including detrending and endpoint extension, to reduce boundary effects. The extension method uses mirror-symmetric extension, with an extension length of 10% of the signal length. The constructed analytical signal retains all information of the original signal while providing direct access to the instantaneous amplitude A(t) = |z(t)| and the instantaneous phase φ(t) = arg(z(t)). Analytical signal representation makes subsequent phase analysis and frequency estimation more accurate and stable.

[0036] The instantaneous phase is obtained by performing Hilbert-Huang transform on the analytical signal. The complex analytical signal z(t) is input into the Hilbert-Huang transform algorithm, and the signal is decomposed into multiple intrinsic mode functions (IMFs) through empirical mode decomposition. The decomposition process uses the ensemble empirical mode decomposition method, adding white noise for multiple decompositions and taking the average. The standard deviation of the noise is set to 0.2 times the original signal, and the number of ensembles is 100. The analytical form of each IMF component is calculated to extract the instantaneous phase. Phase unwrapping uses the Itoh algorithm to ensure phase continuity: φ_unwrap(t)=φ(t)+2πk, where k is an integer chosen to minimize phase jumps. The instantaneous frequency is calculated using the time derivative of the phase: f(t)=(1 / 2π)·dφ / dt. The instantaneous frequency is smoothed, and a 5-point moving average filter is used to remove high-frequency noise. The resulting instantaneous phase sequence fully describes the phase evolution of the signal, laying the foundation for subsequent phase dynamics analysis.

[0037] Exemplarily, the calculation of the phase slip rate based on the instantaneous phase includes: performing time differentiation based on the instantaneous phase to generate a phase change rate and a phase acceleration; using the phase acceleration to perform sudden change detection on the phase change rate to form a slip event; performing time interval statistics on the slip events to generate a slip period; and performing a reciprocal operation on the slip period to obtain the phase slip rate.

[0038] Time differentiation of the instantaneous phase generates the phase rate of change and phase acceleration. An improved numerical differentiation method is used to process the instantaneous phase sequence φ(t), employing a Savitzky-Golay filter for simultaneous smoothing and differentiation. The filter window length is set to 11 points, and the polynomial order is 3, preserving signal detail while suppressing noise amplification. The first-order derivative yields the phase rate of change ω(t) = dφ / dt, representing the instantaneous angular frequency. The second-order derivative yields the phase acceleration α(t) = d²φ / dt², reflecting the rate of frequency change. Outlier detection is performed on the calculated results, using the median absolute deviation method to identify outliers. Data points exceeding five times the MAD are marked as outliers and corrected using linear interpolation. The differential results are post-processed using a zero-phase filter with a cutoff frequency set to 0.8 times the Nyquist frequency. The resulting phase dynamics parameter sequence accurately reflects the instantaneous frequency characteristics and frequency modulation pattern of the signal.

[0039] Phase acceleration is used to detect sudden changes in the phase change rate, forming slip events. A sudden change detection algorithm is constructed to identify anomalous change points in phase dynamics. The sudden change index J(t) is defined as |α(t)|·|Δω(t)|, where Δω(t) = ω(t) - ω(t-1) represents the difference in the phase change rate. A sudden change point is identified when J(t) exceeds an adaptive threshold J_th(t). The adaptive threshold is determined using a local statistical method: J_th(t) = μ_local(t) + k·σ_local(t), where μ_local and σ_local are the local mean and standard deviation within a 100-point window, and k = 3 is the threshold coefficient. The DBSCAN algorithm is used to cluster sudden changes, classifying sudden changes within a temporal distance of less than five sampling points as the same slip event. The characteristics of each slip event include its onset time, duration, peak intensity, and cumulative phase change. A slip event library is established to record all detected events and their characteristic parameters. Through event pattern analysis, different types of slip such as periodic slip, random slip and trend slip can be identified.

[0040] The time interval statistics of the slip events are used to generate the slip period. i}Perform time interval analysis and calculate the interval between adjacent events . The statistical analysis of the interval series adopts a robust estimation method, and the HuberM estimator is used to calculate the location parameters and scale parameters. An interval histogram is constructed, and the interval width is automatically determined using the Freedman-Diaconis rule. The kernel density estimation of the histogram is performed using a Gaussian kernel function, and the bandwidth is selected by cross-validation. The main peak position is identified from the density distribution, which corresponds to the most likely slip period T_mode. Multiple statistics of the slip period are calculated: arithmetic mean T_mean, geometric mean T_geo, and harmonic mean T_harm. The coefficient of variation CV and interquartile range IQR are used to evaluate the stability of the period. When CV<0.3 and IQR / T_median<0.2, the slip process is considered to have good periodicity. For multimodal distributions, the periodic components corresponding to each peak are identified, and the possible physical mechanisms are analyzed.

[0041] The phase slip rate is obtained by performing the inverse operation on the slip period. The statistically obtained slip period T is converted into a slip rate v = 1 / T, representing the number of slips per unit time. The rate calculation takes into account the uncertainty of the period, using the error propagation formula: σ_v = σ_T / T², where σ_T is the standard deviation of the period. A time-varying model of the slip rate, v(t), is established using a local polynomial regression method with adaptive window length adjustment. The model order is selected using the Akaike Information Criterion to balance goodness of fit and model complexity. The spectral characteristics of the slip rate are calculated to identify the main frequency components of the rate variation. The multi-scale characteristics of the rate are analyzed using the wavelet transform, with the Morlet wavelet being the mother wavelet. Features such as energy distribution, ridges, and modulus maxima are extracted from the wavelet coefficients. The stability of the slip rate is assessed using the Lyapunov exponent, with a positive exponent indicating chaotic behavior and a negative exponent indicating stable convergence. The resulting phase slip rate sequence and its statistical characteristics serve as core indicators of information correlation.

[0042] The stability index of the phase slip rate is used as the information correlation. The stability index S is defined as exp(-CV_v) / (1+σ_v / μ_v), where CV_v is the coefficient of variation of the slip rate, and σ_v and μ_v are the standard deviation and mean, respectively. The stability index lies in the interval [0,1]. Larger values ​​indicate more stable phase evolution and more reliable information transmission between moments. The stability index S(τ) is calculated for different time delays τ, and an information correlation curve is constructed. The decay characteristics of the curve reflect the time scale of information transmission, and characteristic time constants are extracted through exponential fitting. Fourier analysis is performed on the correlation curve to identify periodic correlation patterns. The stability index is extended to two dimensions, and the correlation S(t1, t2) between any two moments t1 and t2 is calculated. The constructed correlation matrix is ​​symmetricized and normalized to ensure the mathematical validity of its properties. Eigenvalue decomposition is used to extract the main correlation patterns, with the first few eigenvectors corresponding to the dominant information transmission channels. The spatial distribution of the information correlation intuitively displays the temporal correlation structure and causal network of the pressure data.

[0043] Resonant frequency patterns are identified using information correlation. A two-dimensional Fourier transform is performed on the information correlation matrix to convert the time-domain correlation patterns to the frequency domain for analysis. The transformed spectrum displays the resonance strength between different frequency components, with the spectrum amplitude |F(u,v)| indicating the degree of resonance between frequencies u and v. A resonance detection threshold is set at three times the average spectrum amplitude, and frequency pairs exceeding the threshold are marked as resonant modes. Typical resonant modes identified include fundamental and second harmonic resonance (f0-2f0), cross-frequency resonance (f1-f2), sum frequency resonance (f1+f2), and difference frequency resonance (|f1-f2|). For each resonant mode, a quality factor (Q = f_center / Δf) is calculated, where f_center is the center frequency and Δf is the 3dB bandwidth. The quality factor reflects the selectivity and energy concentration of the resonance; higher Q values ​​indicate sharper resonances. Phase spectrum analysis is used to determine the phase relationship between the resonant modes. In-phase resonances enhance the signal, while anti-phase resonances attenuate it. A resonance frequency spectrum is created, with the horizontal and vertical axes representing the two frequency components involved in the resonance, and the color depth representing the resonance intensity. The bright spots and lines in the spectrum reveal the energy transfer paths and frequency coupling mechanisms in the pressure signal.

[0044] A mutual information resonance vector is generated by amplifying rhythm abnormality indicators based on resonant frequency patterns. A resonance amplification operator A(f) is designed to selectively enhance rhythm abnormality indicators at specific frequency components based on the identified resonant patterns. The amplification coefficient is calculated as: A(f) = 1 + β·R(f)·Q(f), where R(f) is the resonance strength at frequency f, Q(f) is the quality factor, and β = 0.5 is the amplification gain parameter. The original rhythm abnormality indicator RAI is decomposed into N frequency band components RAI_i using a frequency domain filter bank, with each band corresponding to a resonant pattern. A corresponding amplification coefficient, RAI_i' = A(f_i)·RAI_i, is applied to each frequency band component, achieving resonance-driven selective amplification. The amplified components are reconstructed through an inverse transform to generate an enhanced abnormality indicator sequence. The enhanced indicator sequence is concatenated with the resonant pattern features to form a mutual information resonance vector V_mir∈R^D, where D = 256 is the vector dimension. The first 128 dimensions of the vector encode the time series characteristics of the amplified abnormality indicator, and the second 128 dimensions encode the frequency domain characteristics of the resonant pattern. Each vector element is batch normalized to ensure numerical stability. The mutual information resonance vector combines the time domain anomaly information and frequency domain resonance features.

[0045] Step S140 , performing variational echo reconstruction on the mutual information resonance vector to generate variational coding features, performing probability modeling on the variational coding features to determine the probability distribution of echo propagation, sampling and reconstructing the propagation path from the probability distribution, and extracting structural change features based on the propagation path to obtain a variational echo feature map.

[0046] Specifically, variational echo reconstruction is performed on the mutual information resonance vector to generate variational coding features. The mutual information resonance vector V_mir is input into a variational autoencoder architecture. The encoder uses a multilayer perceptron structure with three hidden layers and a decreasing number of neurons at each layer. The encoder outputs a mean vector μ and a logarithmic variance vector log(σ²), both of dimension 32. A reparameterization technique is used to sample a latent variable from a Gaussian distribution: z = μ + σ·ε, where z is the sampled latent variable, μ is the mean vector output by the encoder, σ is the standard deviation vector, and ε is a noise vector sampled from a standard normal distribution. An echo mechanism is introduced to pass the latent variable through a time delay network to generate multiple echo versions, each of which is implemented using exponential decay and cosine modulation. Echo sequences are generated with different delay times and decay strengths, simulating multiple reflections of a signal in a complex medium. The original latent variable is concatenated with the multiple echo versions to form an extended variational coding feature. The variational loss function is: L_VAE = ||x - x'||² + β·KL(q(z|x)||p(z)), where x is the original input vector, x' is the decoder's reconstructed output, ||·||² represents the square of the Euclidean distance, β is the balancing coefficient controlling the weight of the KL term, KL(·||·) is the KL divergence, q(z|x) is the posterior distribution generated by the encoder, and p(z) is the standard normal prior distribution. During training, the encoder learns to extract the essential features of the data, and the decoder learns to reconstruct the original data from the latent representation.

[0047] In some embodiments, the probabilistic modeling of the variational coding features to determine the probability distribution of echo propagation includes: inputting the variational coding features into a diffusion model to obtain a forward trajectory; performing time reversal on the forward trajectory to generate a reverse path; sampling on the reverse path to obtain an intermediate state distribution; and aggregating the intermediate state distribution into a probability distribution of echo propagation.

[0048] The variational coded features are input into the diffusion model to obtain the forward trajectory. A denoising diffusion probability model is constructed to process the variational coded features. The forward process transforms the data into pure noise by gradually adding Gaussian noise. The conditional distribution of the added noise is: q(z_t|z_{t-1})=N(z_t|√(1-β_t)z_{t-1},β_tI), where z_t is the state vector at time step t, z_{t-1} is the state at the previous time step, β_t is the noise intensity parameter at time t that controls the amount of added noise, √(1-β_t) is the signal preservation coefficient, I is the identity matrix, and N(·|μ,Σ) represents a Gaussian distribution with mean μ and covariance Σ. The noise intensity varies according to a cosine scheduling strategy to ensure smooth diffusion. The total number of diffusion steps is set to 1000, which is sufficient to completely randomize the structured data. The state distribution at any time step can be directly calculated through cumulative products, avoiding the computational overhead of step-by-step simulation. The forward process records the state at each time step, forming a complete diffusion trajectory. The trajectory shows the dynamic process of data structure gradually disappearing and information being lost.

[0049] The forward trajectory is time-reversed to generate a reverse path. The state evolution information recorded in the forward trajectory is used as training data to train the neural network to learn the reverse denoising process. The network input includes the noise state at each moment in the forward trajectory and the corresponding time step information. The network architecture adopts U-Net design, which maintains the flow of multi-scale information through jump connections. Time information is embedded in the network through sinusoidal coding, allowing the model to adapt to different noise levels. The training goal is to minimize the noise prediction error and grasp the data distribution law by learning the noise added at each step in the forward process. After training, the learned reverse sampling formula is used: , where μ_θ is the predicted denoised mean, α_t=1-β_t is the signal retention rate, is the cumulative retention rate from the beginning to time t, and ε_θ(z_t,t) is the noise predicted by the neural network. Starting from pure Gaussian noise, reverse sampling is performed and denoising is gradually performed to generate clear data.

[0050] The intermediate state distributions were sampled along the reverse path. Dense sampling was performed at key time points during the reverse denoising process, specifically at time points t∈{250,500,750}, corresponding to positions 0.25, 0.5, and 0.75 of a total of 1000 steps. These sampling points cover the early, middle, and late stages of the reconstruction process, capturing the complete evolution of structure formation. The denoising process was run independently 100 times at each sampling point to generate the state set at that time, ensuring statistical reliability. Kernel density estimation was used to fit the probability density at each time point, using a Gaussian kernel as the kernel function and an adaptive bandwidth determined by the Silverman rule. The 2-Wasserstein distance between the distributions at different times was calculated, a distance metric that accounts for differences in the geometric structure of the distributions. Analysis of the distance matrix revealed the heterogeneity of the reconstruction process: the early stages (0-250 steps) saw slow changes, primarily removing high-frequency noise; the middle stages (250-500 steps) saw rapid formation of the main structure; and the late stages (500-750 steps) saw fine-tuning of local details. We visualized the intermediate states in a reduced-dimensional space, projecting the high-dimensional distribution onto a two-dimensional plane using the t-SNE algorithm. The visualization clearly demonstrated how the data points gradually aggregated from a random distribution into meaningful clusters.

[0051] The intermediate state distributions are aggregated into a probability distribution for echo propagation. An intelligent weighting strategy is designed to integrate the intermediate state distributions at three key moments to avoid the blurring effect caused by simple averaging. Weights are calculated based on the information gain at each moment relative to the initial pure noise state, measured by relative entropy. A soft maximization function is used to ensure smooth weight distribution and numerical stability. The temperature parameter is set to 0.1 to avoid extreme weight distribution while maintaining variability. The aggregation formula is: p_echo(z)=Σ_tw_t·pt_t(z), where p_echo(z) is the final probability distribution for echo propagation, w_t is the normalized weight at time t such that Σw_t=1, and p_t(z) is the intermediate state distribution at time t, summed over the three sampling moments. The Student's t-distribution family is used as the parameterized form of the aggregated distribution. Compared to the Gaussian distribution, its thick tails better capture rare events and anomalous patterns. The three core parameters of the t-distribution: the position vector μ, the scale matrix Σ, and the degrees of freedom ν, are iteratively estimated using the expectation-maximization algorithm. Adaptive estimation of the degrees of freedom parameter ν is particularly critical, as it controls the thickness of the distribution's tail and reflects the degree of data heterogeneity. After the algorithm converges, validation sampling is performed on the aggregate distribution to ensure that the generated samples cover the typical patterns of each intermediate state. The moments and information entropy of the aggregate distribution are calculated to evaluate its statistical properties and expressive power. The final echo propagation probability distribution p_echo fully integrates the characteristics of different stages of the diffusion-reconstruction process, providing a rich probabilistic representation for anomaly detection.

[0052] Propagation paths are reconstructed by sampling from a probability distribution. A Markov Chain Monte Carlo method is used to generate representative samples from the learned probability distribution. The Hamiltonian Monte Carlo sampling algorithm is chosen, which utilizes gradient information to guide the sampling process, improving efficiency in high-dimensional spaces. A potential energy function is defined as: U(z) = -logp(z), where U(z) is the potential energy function and p(z) is the target probability density. A negative logarithmic transformation transforms the probability maximization problem into an energy minimization problem. A Hamiltonian system containing potential and kinetic energy is constructed, achieving efficient sampling by simulating the dynamic evolution of the physical system. A leapfrog integrator is used to numerically solve the Hamiltonian equations, preserving the system's symplectic structure and energy conservation. The integration step size is adjusted using an adaptive algorithm to monitor numerical errors and maintain an appropriate acceptance rate. Each sampled trajectory records the complete state evolution sequence, and the trajectory length is set to be long enough to explore various modes in the distribution. Statistical analysis is performed on the generated trajectory set to calculate similarities and differences between them. Cluster analysis is used to group the trajectories, with each group representing a typical propagation path. Discrete trajectory points are smoothly interpolated to generate a continuously differentiable path function.

[0053] The variational echo signature is obtained by extracting structural variation features based on the propagation path. Differential geometric features are calculated for each propagation path, with a focus on curvature. The curvature calculation formula is: κ(s) = ||dz / ds×d²z / ds²|| / ||dz / ds||³, where κ(s) is the curvature value at the arc length parameter s, z(s) is a parameterized path vector function, dz / ds is the first-order derivative of the path, i.e., the tangent vector, and d²z / ds² is the second-order derivative. × represents the vector cross product, and ||·|| represents the Euclidean norm of the vector. Curvature reflects the local curvature of the path, with regions of high curvature corresponding to sharp changes in the propagation process. Other geometric quantities, such as the path torsion, are also calculated to describe the three-dimensional torsional characteristics of the path. Feature points on the path are identified, including extreme curvature points, points of zero curvature, and regions of high torsion. A graph structure is constructed to represent the topological relationships between feature points, with node attributes containing local geometric information. The edge weights of the graph are determined using a Gaussian kernel function, ensuring that spatially adjacent nodes have stronger connections. A graph convolutional network is applied to extract multi-scale structural features. The network learns high-order representations by aggregating neighborhood information. The extracted features are organized into a spatial feature map with a resolution of 64×64 and 16 feature channels.

[0054] Step S150: Input the rhythm abnormality index and the mutual information resonance vector into the deep classification network to generate a nonlinear conversion pattern, predict the phase change critical point from the nonlinear conversion pattern, determine the phase change warning parameter based on the phase change critical point, and use the phase change warning parameter to build a warning decision network.

[0055] Specifically, the rhythm abnormality indicator and the mutual information resonance vector are input into a deep classification network to generate a nonlinear transformation pattern. A deep neural network architecture specifically designed to handle multimodal input is constructed. The first input branch receives the rhythm abnormality indicator RAI in scalar form, and the second input branch receives the 256-dimensional mutual information resonance vector V_mir. The two branches use different feature extraction strategies. The scalar branch is expanded to a high-dimensional representation through an embedding layer, while the vector branch extracts local correlations through a convolutional layer. The first branch consists of an embedding layer and a three-layer fully connected network with 64, 128, and 128 neurons, respectively. Batch normalization and a dropout of 0.3 are used to prevent overfitting. The second branch adopts a one-dimensional convolutional structure. The three convolutional layers have 32, 64, and 128 filters, respectively, with a convolution kernel size of 5 and a Reluctant Unit (ReLU) activation function. The features of the two branches are weighted and combined at the fusion layer using a learnable attention mechanism: F_fused = α·F_RAI + (1-α)·F_mir, where α is the attention weight of the gated unit output, and F_RAI and F_mir represent the 128-dimensional feature representations extracted by the two branches, respectively. The fused features are input to a backbone network consisting of six residual blocks, each of which contains two convolutional layers, batch normalization, ReLU activation, and skip connections. The network output layer generates a 64-dimensional nonlinear transition pattern vector through a fully connected projection, with each dimension encoding a specific system state transition path. Training uses a combination of triplet loss and cross-entropy loss, enabling the network to learn to distinguish different nonlinear evolution patterns while maintaining feature discriminability.

[0056] In some embodiments, predicting the phase transition critical point from the nonlinear conversion pattern includes: constructing a renormalization group flow for the nonlinear conversion pattern; tracking fixed point trajectories in the renormalization group flow; identifying bifurcation values ​​where the fixed point trajectory turns from stable to unstable; and marking parameter points corresponding to the bifurcation values ​​as phase transition critical points.

[0057] A renormalized group flow is constructed for the nonlinear conversion mode. The 64-dimensional nonlinear conversion mode is regarded as the field variable φ(x) in statistical field theory, and the momentum shell renormalization transformation is defined to gradually eliminate high-frequency fluctuations. The renormalization transformation consists of two steps: first, the mode elimination in momentum space is performed, and the degrees of freedom within the momentum shell Λ / b<|k|<Λ are integrated, where Λ is the ultraviolet cutoff and b>1 is the scale factor; then the remaining field variables and spatial coordinates are rescaled. The specific transformation form is: φ'(k)=b^(d / 2-[φ])φ(k / b), where d is the spatial dimension and [φ] is the scale dimension of the field determined by dimensional analysis. The renormalization of the coupling constant follows: g'_i=b^(y_i)g_i+Σ_jkB_ijkg_jg_k, where y_i is the scale dimension of the coupling constant g_i and B_ijk is the beta function coefficient. A continuous RG flow is constructed by iteratively applying the renormalization transformation T: , where l = ln(b) is the flow parameter and β_i is the beta function. The numerical implementation employs the Monte Carlo renormalization group, with real-space renormalization achieved via block variable techniques. The evolution of the effective action is calculated using the truncated Dyson-Schwinger equation in momentum space. While maintaining the Helmholtz free energy under renormalization, the scaling relationship between the critical exponents is determined. The constructed RG flow fully describes the effective theory of the system at different energy scales, revealing the scaling invariance of the phase transition.

[0058] Track fixed point trajectories in the renormalization group flow. Determine the fixed points g* of the renormalization group by solving for the zeros of the beta function β(g*)=0. These points correspond to scale-invariant critical theories. Solve the nonlinear system of equations using the multidimensional Newton method, and approximate the Jacobian matrix using finite differences. Perform stability analysis on each fixed point found and calculate the linearized RG transformation matrix. The eigenvalue spectrum of . The sign of the eigenvalue λ_α determines the correlation of the corresponding eigendirection: λ_α>0 indicates a correlated direction (increasing), λ_α<0 indicates an uncorrelated direction (decaying), and λ_α=0 indicates an edge direction. The correlation exponent is calculated as y_α=ln|λ_α| / ln(b) and is directly related to the critical exponent of the physical quantity. RG trajectories starting from different initial conditions are tracked, and the flow equation is integrated using the fourth-order Runge-Kutta method. The domain of attraction of a stable fixed point corresponds to all systems of the same universality class, with the boundary being the critical surface. Unstable fixed points act as saddle points on the critical surface, separating different phase regions. The trajectories of fixed points are tracked as a function of external parameters (such as temperature and pressure) through parametric extension. When two fixed points collide and annihilate, a topological phase transition occurs in the system. A complete RG flow diagram is drawn, with arrows indicating the direction of flow and color coding the flow velocity.

[0059] Identify the bifurcation value where the fixed-point trajectory transitions from stable to unstable. Systematically analyze the stability of the fixed point as a function of the control parameters, focusing on the zero crossings of the stability matrix eigenvalues. The fixed point loses stability when the maximum eigenvalue λ_max(p) changes from negative to positive as the parameter p changes. A bifurcation occurs at λ_max(p_c)=0, where p_c is the value of the bifurcation parameter. Classify bifurcations by the eigenvalue crossing: a single real eigenvalue crossing zero corresponds to a saddle-node bifurcation, while a pair of conjugate complex eigenvalues ​​crossing zero corresponds to a Hopf bifurcation. Calculate the normal form of the bifurcation, determine the bifurcation expansion coefficient, and the effects of nonlinear terms. For a saddle-node bifurcation: dx / dt=r+x², where r=p-p_c is the deviation parameter; for a Hopf bifurcation, consider the complex amplitude equation. Use the central manifold theorem to reduce the high-dimensional system to the bifurcation subspace, simplifying the analysis. Construct the bifurcation equation using the Lyapunov-Schmidt method to systematically handle high-codimensional bifurcations. Pseudo-arc-length extension is used numerically to track bifurcation curves and process turning points and branching points. For each identified bifurcation point, its homoclinic and heteroclinic orbits are calculated. These special orbits determine the dynamical path of the phase transition.

[0060] The parameter points corresponding to the bifurcation values ​​are marked as critical points of the phase transition. A quantitative mapping relationship between mathematical bifurcation parameters and physical control parameters is established, and the conversion coefficients are determined through dimensional analysis and experimental calibration. For each bifurcation value p_c, the corresponding physical quantities are calculated: temperature T_c, pressure P_c, concentration C_c, etc., taking into account the measurement units and scaling factors. The thermodynamic characteristics of the phase transition are verified, and the behavior of the free energy and its derivatives at the critical point is calculated. The scaling behavior of the order parameter η near the critical point is: η~|T_c|^β, where β is the critical exponent. Response functions such as specific heat, compressibility, and magnetic susceptibility exhibit power-law divergence or finite jumps at the critical point. Finite-scale scaling theory is used to address the effects of finite systems and extract the critical behavior of the thermodynamic limit. Data collapse analysis verifies the scaling assumption: F(L^{1 / ν}t)=L^{-β / ν}η(t,L), where L is the system scale and t=(T-T_c) / T_c. Verify various scaling relationships and verify the internal consistency of critical exponents. For first-order phase transitions, determine coexistence curves, metastable regions, and hysteresis effects. Construct complete phase diagrams, annotating all phase transition lines, triple points, and critical endpoints. A database of critical points for phase transitions includes precise parameter values, error bounds, applicable conditions, and relevant physical quantities.

[0061] Phase transition warning parameters are determined based on the critical points of phase transitions. Precursor characteristics are calculated for each critical point, including the decay time of the autocorrelation function, the abnormal increase in variance, and the enhancement of low-frequency components in the power spectrum. A comprehensive warning parameter is defined as: P_w = (σ² / μ²) exp(τ / τ_0) (1 + λ·S), where σ² is the variance of the state variable within a sliding window, μ is the mean, τ is the autocorrelation time obtained by integrating the autocorrelation function, τ_0 is the system's characteristic timescale, S is the skewness quantifying the asymmetry of the distribution, and λ is the skewness influence coefficient. An adaptive sliding window algorithm is designed, with the window length W = W_0 · (1 + κ · |dP / dt|) dynamically adjusted, where W_0 is the baseline window length, κ is the adjustment coefficient, and |dP / dt| is the parameter change rate. A multi-scale warning parameter set is constructed, consisting of instantaneous parameters (fast response but high noise), short-term parameters (balancing timeliness and stability), and long-term parameters (stable but with high latency). By integrating multi-scale information through Bayesian inference, the posterior probability P(phase transition | parameter) ∝ P(parameter | phase transition) · P(phase transition) is determined. Four warning thresholds are set: normal (P_w < 0.3), concern (0.3 ≤ P_w < 0.5), warning (0.5 ≤ P_w < 0.7), and danger (P_w ≥ 0.7). Response strategies and intervention measures are developed for each warning level.

[0062] In some embodiments, the use of the phase change warning parameters to construct a warning decision network includes: converting the phase change warning parameters into oscillator coupling strength; constructing an oscillator network based on the oscillator coupling strength; performing synchronization analysis on the oscillator network to obtain a critical topology; and solidifying the critical topology into a warning decision network.

[0063] The phase change warning parameters are converted into oscillator coupling strengths. A nonlinear mapping function is designed to convert the multidimensional warning parameter P_w into the coupling matrix of the oscillator network. The coupling strength is calculated using the Gaussian kernel function: K_ij = K_0 · exp(-||P_i - P_j||² / (2σ²)) · H(r_c - ||P_i - P_j||), where K_ij is the coupling strength between nodes i and j, K_0 is the baseline strength, ||P_i - P_j|| is the Euclidean distance between warning parameter vectors, σ is the interaction range parameter, H is a step function to implement distance truncation, and r_c is the truncation radius. A directional factor is introduced to reflect causal relationships: K_ij^{dir} = K_ij · (1 + α · TE_ij), where TE_ij is the transfer entropy from node j to i, and α is the directional strength parameter. Time-varying coupling is updated using exponential smoothing: K_ij(t) = γ·K_ij(t-Δt) + (1-γ)·K_ij^{new}, where γ is the forgetting factor, Δt is the update interval, and K_ij^{new} is the newly calculated coupling strength. The coupling matrix is ​​spectrally normalized to ensure that the maximum eigenvalue is within a reasonable range and to avoid numerical instability. A multi-layer coupling network is constructed, with different layers corresponding to different time scales such as seconds, minutes, and hours. Inter-layer coupling is determined by a scale matching function to maintain the coordination of multiscale dynamics.

[0064] An oscillator network is constructed based on the coupling strength of oscillators. A generalized Kuramoto model is used to describe the collective dynamics of N coupled oscillators, each representing a monitoring node. The dynamical equation is: dθ_i / dt = ω_i + Σ_jK_ij · f(θ_j - θ_i) + ξ_i(t), where θ_i is the phase of the i-th oscillator, ranging from [0 to 2π], ω_i is the natural frequency, f(Δθ) = sin(Δθ) + (ε / 2)sin(2Δθ) is the coupling function including the second harmonic, ε controls the nonlinear strength, and ξ_i(t) is a Gaussian white noise term. The natural frequency is sampled from a Lorentzian distribution, with the center frequency and half-width set according to the system characteristics. The network topology is obtained by thresholding the coupling matrix, retaining connections with strength exceeding a threshold. A community structure is introduced to reflect the functional modules, with strong coupling within the same community and weak coupling between communities. Numerical integration is performed using a stochastic differential equation solver with an adaptive step size. Monitor the macroscopic order parameter: r(t)e^{iψ(t)}=(1 / N)Σ_je^{iθ_j(t)}, where r is the degree of synchronization, ranging from 0 to 1, ψ is the average phase, and i is the imaginary unit. Record local order parameters to characterize the synchrony within the community and identify partially synchronized states. Power spectrum analysis is used to identify collective oscillation modes and frequency locking.

[0065] Synchronicity analysis of oscillator networks is performed to obtain critical topologies. The stability conditions for synchronization in oscillator networks are determined through linear stability analysis of the system, and the master stability function method is used to handle complex oscillator coupling networks. The transverse Lyapunov exponent of the oscillator network is calculated, and the network synchronization is stable when the maximum value is less than zero. The critical value K_c for synchronization transition is determined by sweeping the coupling strength between oscillators, at which desynchronization bifurcation occurs in the oscillator network. Finite-time Lyapunov exponents are used to identify Lagrangian coherent structures in the oscillator phase space. Synchronous cliques in the oscillator network are detected using a modularity optimization algorithm to maximize the network's modularity function. The phase difference distribution between different cliques in the oscillator network is calculated to identify phase locking and phase slip phenomena. The spectral properties of the oscillator network are analyzed. The eigenvalue spectrum of the Laplacian matrix determines the timescale of energy diffusion in the network, and the spectral radius of the adjacency matrix affects the synchronization ability of the oscillators. The eigenmodes that contribute most to synchronization in the oscillator network are identified. The eigenvectors corresponding to the smallest non-zero eigenvalues ​​are also known as Federer vectors. A targeted attack strategy is used to identify key nodes and edges in the oscillator network. Removing these nodes and edges significantly reduces the network's synchronization performance. The synchronization modes of the oscillator network in different parameter regions are recorded, including complete synchronization, group synchronization, traveling waves, spiral waves, and singular states. These synchronization modes and key structures together constitute the critical topology of the oscillator network.

[0066] The critical topology is solidified into a warning decision network. The critical topology features obtained from oscillator network analysis are converted into a graph neural network architecture design. Key nodes in the critical topology are mapped to the network's core processing units, and key connections are transformed into information transmission paths. The network adopts a hierarchical structure: the input layer receives a real-time warning parameter stream, the hidden layer simulates the oscillator's coupling dynamics based on the critical topology, and the output layer produces the decision result. Nodes are designed using gated recurrent units, controlling information retention and updating through a gating mechanism. Edge information transmission uses a message passing framework, with nodes updating their own states by aggregating neighbor messages. A multi-head graph attention mechanism is introduced, with different attention heads focusing on different coupling patterns. The network consists of four layers of graph convolution, each followed by layer normalization and residual connections to maintain gradient flow and training stability. Global information is aggregated through virtual nodes or readout functions, and attention weights are adaptively learned. The output layer uses a hybrid expert system, with different experts responsible for different types of fault modes. The decision output includes a probability distribution of fault type, an estimated time of occurrence with uncertainty, a severity score (0 to 10), and a ranking of recommended actions. Knowledge distillation is used to encode the dynamic behavior of the oscillator network into static network parameters. Reinforcement learning is used to fine-tune the decision-making strategy. The reward function is: R = α·TPR - β·FPR - γ·delay, where TPR is the true positive rate, FPR is the false positive rate, delay is the detection delay, and α, β, and γ are weight coefficients. During deployment, the network structure is fixed, and only the hidden states of the nodes are updated. Inference time is kept within 10 milliseconds.

[0067] Step S160: Outputting a graded warning signal through the early warning decision network integrated variational echo feature map to complete the neural network driven battery acoustic pressure monitoring and early warning.

[0068] In some embodiments, the output of a graded warning signal by synthesizing the variational echo feature map through the warning decision network includes: performing multi-resolution decomposition on the variational echo feature map to extract a singular value spectrum; inputting the singular value spectrum into the warning decision network to calculate the activation intensity; and determining the graded warning signal based on the distribution range of the activation intensity.

[0069] The variational echo feature map undergoes multiresolution decomposition to extract singular value spectra. Singular value decomposition is performed channel-by-channel on the input variational echo feature map, treating each channel as a 64×64 matrix. Singular value decomposition is performed on the cth channel: F_c=U_c·Σ_c·V_c^T, where F_c is the feature matrix of the cth channel, U_c is the left singular vector matrix containing the spatial basis functions, Σ_c is a diagonal matrix whose diagonal elements are singular values ​​representing the importance of each basis function, and V_c is the right singular vector matrix encoding the combination pattern of features. The superscript T denotes the matrix transpose. The singular value sequence of each channel is extracted and arranged in descending order to form a feature representation with decreasing energy. A multiresolution pyramid is constructed, reconstructing features using different numbers of principal singular values. 4, 8, 16, and 32 principal components are selected to correspond to coarse to fine resolution levels. The energy retention rate is calculated for each resolution as the ratio of the sum of squares of the first k singular values ​​to the sum of squares of all singular values. This metric quantifies the information integrity at different resolutions. The singular value sequence is logarithmically transformed and normalized to enhance the discrimination of small singular values ​​and make the features more numerically stable. The singular value spectra of the 16 channels are combined into a singular value spectrum matrix S∈R^(16×64) according to the channel order, with each row corresponding to the complete singular value sequence of a channel. High-order statistical features of the singular value spectrum are calculated, including spectral entropy to measure complexity, effective rank to assess information dimensionality, condition number to reflect numerical stability, and spectral gaps to identify the boundaries between primary and secondary components. The singular value sequence is subjected to multi-scale analysis using discrete wavelet transform. The wavelet coefficients reveal the energy distribution of different frequency components, and mutation points correspond to the emergence of abnormal patterns.

[0070] The singular value spectrum is input into the early warning decision network to calculate the activation intensity. The singular value spectrum matrix S and the extracted statistical feature vector are concatenated and fed into the specially designed spectral analysis module of the early warning decision network. This module first processes the singular value sequence of each channel through a parallel one-dimensional convolution group. Convolution kernels of different sizes (3, 5, and 7) capture spectral patterns at different scales. After batch normalization and activation function, the convolution output uses a channel-by-channel attention mechanism to adaptively adjust the importance of each channel. The multi-scale features are fused and input into the self-attention layer to calculate the long-range dependencies within the singular value sequence. The attention mechanism learns which singular value positions are most critical for anomaly detection through a query-key-value transformation. The attention-weighted features are mapped to the anomaly pattern space through a two-layer fully connected network with a dropout layer to prevent overfitting. The activation intensity is calculated for M predefined anomaly patterns, each corresponding to a specific fault mechanism, such as thermal runaway precursors, internal short circuit development, and electrolyte leakage. The activation intensity is calculated using a normalized exponential function to ensure numerical stability and interpretability. A temperature parameter, T, is introduced to adjust the sharpness of activations. The value of T is optimized through grid search on the validation set. The activation intensity sequence is filtered in the time domain, using an exponential moving average to eliminate transient disturbances while retaining persistent abnormal patterns.

[0071] A graded warning signal is determined based on the distribution range of activation intensity. The distribution characteristics of the M-dimensional activation intensity vector A = [A_1, A_2, ..., A_M] are analyzed, and a multi-level decision logic is designed to map continuous activation values ​​to discrete warning levels. Five warning levels are defined: normal state, minor abnormality, moderate risk, high risk, and emergency state. The activation intensity interval boundaries corresponding to each level are determined by quantiles of historical data to ensure that the probability of occurrence of each level conforms to the actual distribution. A comprehensive warning index is calculated: W = Σ_mw_m·A_m, where W is the weighted warning index, w_m is the hazard weight of the mth abnormal pattern, reflecting the probability of it causing a serious failure, and A_m is the normalized activation intensity. The sum is traversed over all M abnormal patterns. The hazard weight is determined through fault tree analysis and expert knowledge, taking into account the severity, development speed, and controllability of the fault. Fuzzy logic is used to handle cases where the activation intensity approaches the level boundaries. A trapezoidal membership function μ_L(W) is defined to calculate the membership of the warning index W for each level L. When the membership of multiple levels is similar, evidence theory is used to integrate judgments from different information sources, including historical trends, environmental factors, and system status. A dynamic threshold mechanism is introduced to adjust the criteria for determining the warning level based on the battery's stage of use, environmental conditions, and historical performance. A temporal consistency constraint is implemented, requiring that changes in the warning level must persist for multiple sampling cycles before confirmation, to prevent false alarms caused by transient interference. The warning change rate indicator is calculated: R = |W(t) - W(t - Δt)| / Δt, where R is the rate of change and Δt is the time interval. This indicator is used to assess the rapidity of the abnormal development. A structured warning output is generated, including the primary warning level, secondary risk factors, development trend forecasts, recommended response time windows, and graded disposal plans, ultimately completing neural network-driven battery acoustic pressure monitoring and warning.

[0072] In order to implement a neural network driven battery acoustic pressure monitoring and early warning method corresponding to the above method embodiment, to achieve the corresponding functions and technical effects. Figure 2 , Figure 2 The following is a block diagram of a neural network-driven battery acoustic pressure monitoring and warning system 200 provided in an embodiment of the present application. For ease of illustration, only the parts relevant to this embodiment are shown. The neural network-driven battery acoustic pressure monitoring and warning system 200 provided in an embodiment of the present application includes: The data encoding module 201 is used to obtain pressure monitoring data during battery operation and perform neural network encoding on the pressure monitoring data to generate a training data set; a rhythm analysis module 202 for constructing a temporal convolutional neural network based on the training data set, training the temporal convolutional neural network to identify pressure cycle rhythms, extracting rhythm parameters from the pressure cycle rhythms to generate a rhythm reference model, and performing deviation analysis on the rhythm reference model to generate a rhythm abnormality indicator; A resonance learning module 203 is configured to perform mutual information maximization processing on the training data set to establish a resonance learning network, determine information correlation of data at different times through the resonance learning network, identify a resonance frequency pattern using the information correlation, and amplify the rhythm abnormality indicator based on the resonance frequency pattern to generate a mutual information resonance vector; a variational reconstruction module 204 configured to perform variational echo reconstruction on the mutual information resonance vector to generate a variational coding feature, perform probability modeling on the variational coding feature to determine a probability distribution of echo propagation, sample and reconstruct a propagation path from the probability distribution, and extract structural change features based on the propagation path to obtain a variational echo feature map; Phase change prediction module 205, configured to input the rhythm abnormality index and the mutual information resonance vector into a deep classification network to generate a nonlinear conversion pattern, predict a phase change critical point from the nonlinear conversion pattern, determine a phase change warning parameter based on the phase change critical point, and construct a warning decision network using the phase change warning parameter; The warning output module 206 is used to output a graded warning signal by integrating the variational echo characteristic graph through the warning decision network, thereby completing the battery acoustic pressure monitoring and warning driven by the neural network.

[0073] The above-mentioned neural network-driven battery acoustic pressure monitoring and early warning system 200 can implement the neural network-driven battery acoustic pressure monitoring and early warning method of the above-mentioned method embodiment. The optional options in the above-mentioned method embodiment are also applicable to this embodiment and will not be described in detail here. The remaining contents of the embodiment of this application can refer to the contents of the above-mentioned method embodiment and will not be repeated in this embodiment.

[0074] The purpose of the above embodiments is to exemplify and deduce the technical solution of the present invention, and to fully describe the technical solution, purpose and effect of the present invention. Its purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosed content of the present invention, and it does not limit the scope of protection of the present invention.

[0075] The above embodiments are not exhaustive and may include many other embodiments not listed above. Any replacements and improvements made without violating the concept of the present invention are within the scope of protection of the present invention.

Claims

1. A neural network driven battery acoustic pressure monitoring and early warning method, characterized in that: include: Obtaining pressure monitoring data during battery operation, and performing neural network encoding on the pressure monitoring data to generate a training data set; Constructing a temporal convolutional neural network based on the training data set, training the temporal convolutional neural network to identify pressure cycle rhythms, extracting rhythm parameters from the pressure cycle rhythms to generate a rhythm reference model, and performing deviation analysis on the rhythm reference model to generate a rhythm abnormality indicator; Performing mutual information maximization processing on the training data set to establish a resonance learning network, determining information correlation of data at different times through the resonance learning network, identifying a resonance frequency pattern using the information correlation, and amplifying the rhythm abnormality indicator based on the resonance frequency pattern to generate a mutual information resonance vector; performing variational echo reconstruction on the mutual information resonance vector to generate a variational coding feature, performing probability modeling on the variational coding feature to determine a probability distribution of echo propagation, sampling and reconstructing a propagation path from the probability distribution, and extracting structural change features based on the propagation path to obtain a variational echo feature map; Inputting the rhythm abnormality index and the mutual information resonance vector into a deep classification network to generate a nonlinear conversion pattern, predicting a phase change critical point from the nonlinear conversion pattern, determining a phase change warning parameter based on the phase change critical point, and constructing a warning decision network using the phase change warning parameter; The early warning decision network integrates the variational echo characteristic graph to output a graded early warning signal, thereby completing the neural network-driven battery acoustic pressure monitoring and early warning.

2. The method according to claim 1, characterized in that The step of performing neural network encoding on the pressure monitoring data to generate a training data set includes: Performing reverse time evolution on the pressure monitoring data to obtain a causal precursor signal; Applying chaotic attractor projection to the causal precursor signal to extract singular trajectories; Using the singular trajectory as a neuron dead zone activation pattern to generate a sparse code; A training dataset is constructed based on the sparse coding.

3. The method according to claim 1, characterized in that The generating of a rhythm abnormality index by performing deviation analysis on the rhythm reference model includes: activating a self-diagnosis mode of the rhythm reference model to obtain an internal parameter distribution diagram; Identifying weight imbalance regions on the parameter distribution graph and establishing imbalance pattern classification, wherein the imbalance pattern classification includes gradient explosion type, gradient vanishing type, and oscillation divergence type; A severity assessment process is performed on the imbalance pattern classification to generate a rhythm abnormality index.

4. The method according to claim 1, wherein Determining the information correlation of data at different times through the resonance learning network includes: Mapping the training data set to a complex domain to construct an analytical signal; Performing Hilbert-Huang transform on the analytical signal to obtain an instantaneous phase; calculating a phase slip rate based on the instantaneous phase; The stability index of the phase slip rate is used as the information correlation degree.

5. The method according to claim 1, wherein The performing probability modeling on the variational coding feature to determine the probability distribution of echo propagation includes: Inputting the variational coding features into a diffusion model to obtain a forward trajectory; performing time reversal on the forward trajectory to generate a reverse path; Sampling on the reverse path to obtain an intermediate state distribution; The intermediate state distribution is aggregated into a probability distribution of echo propagation.

6. The method according to claim 1, characterized in that The predicting of the phase transition critical point from the nonlinear conversion mode comprises: constructing a renormalization group flow for the nonlinear conversion mode; tracking fixed point trajectories in the renormalized group flow; Identifying a bifurcation value where the fixed point trajectory changes from stable to unstable; The parameter point corresponding to the bifurcation value is marked as the phase transition critical point.

7. The method according to claim 1, characterized in that The method of constructing a warning decision network using the phase change warning parameters includes: converting the phase change warning parameter into an oscillator coupling strength; constructing an oscillator network based on the oscillator coupling strength; Performing synchronization analysis on the oscillator network to obtain a critical topology; The critical topology is solidified into an early warning decision network.

8. The method according to claim 1, characterized in that The step of outputting a graded warning signal by synthesizing the variational echo feature map through the warning decision network includes: Performing multi-resolution decomposition on the variational echo feature map to extract a singular value spectrum; Inputting the singular value spectrum into the early warning decision network to calculate the activation intensity; A graded warning signal is determined according to the distribution range of the activation intensity.

9. The method according to claim 4, characterized in that The calculating the phase slip rate based on the instantaneous phase includes: Performing time differentiation based on the instantaneous phase to generate a phase change rate and a phase acceleration; Performing a sudden change detection on the phase change rate using the phase acceleration to form a slip event; Performing time interval statistics on the slip events to generate a slip cycle; A reciprocal operation is performed on the slip period to obtain a phase slip rate.

10. A neural network driven battery acoustic pressure monitoring and early warning system, characterized in that: include: A data encoding module is used to obtain pressure monitoring data during battery operation and perform neural network encoding on the pressure monitoring data to generate a training data set; a rhythm analysis module, configured to construct a temporal convolutional neural network based on the training data set, train the temporal convolutional neural network to identify pressure cycle rhythms, extract rhythm parameters from the pressure cycle rhythms to generate a rhythm reference model, and perform deviation analysis on the rhythm reference model to generate a rhythm abnormality indicator; a resonance learning module, configured to perform mutual information maximization processing on the training data set to establish a resonance learning network, determine information correlation of data at different times through the resonance learning network, identify a resonance frequency pattern using the information correlation, and amplify the rhythm abnormality indicator based on the resonance frequency pattern to generate a mutual information resonance vector; a variational reconstruction module, configured to perform variational echo reconstruction on the mutual information resonance vector to generate a variational coding feature, perform probability modeling on the variational coding feature to determine a probability distribution of echo propagation, sample and reconstruct a propagation path from the probability distribution, and extract structural change features based on the propagation path to obtain a variational echo feature map; a phase change prediction module, configured to input the rhythm abnormality index and the mutual information resonance vector into a deep classification network to generate a nonlinear conversion pattern, predict a phase change critical point from the nonlinear conversion pattern, determine a phase change warning parameter based on the phase change critical point, and construct a warning decision network using the phase change warning parameter; The early warning output module is used to output a graded early warning signal by integrating the variational echo characteristic graph through the early warning decision network, thereby completing the battery acoustic pressure monitoring and early warning driven by the neural network.

Images