Battery swelling force multi-source data extraction method and system

Abstract

The present application relates to the technical field of battery testing, and particularly relates to a battery expansion force multi-source data extraction method and system. The method comprises the following steps: obtaining a high-density time sequence data of a charge-discharge step time stamp and an expansion force sensor array; constructing a dynamic regulation window core to perform bidirectional nearest neighbor registration, eliminating time sequence mismatch between the step event and the force signal, and generating aligned data matrix under a unified time base; constructing an expansion-compression state memory chain framework to extract an implicit state sequence, reconstructing a stress rate transition matrix envelope, and constructing a tensor containing temperature segment compensation to peel off thermal expansion interference; constructing a diffusion graph model to perform spectral domain reconstruction on the missing data, and outputting a complete three-dimensional expansion feature sequence; and decomposing and extracting peak value, rate, rebound and mutation degree multi-dimensional characteristic parameters according to the step segment, and storing the multi-dimensional characteristic parameters into a database. The present application realizes high-precision registration and feature extraction of multi-source heterogeneous data, and is suitable for expansion force analysis and health state evaluation in battery high-low temperature cycle aging test.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, specifically to a method and system for extracting multi-source data on battery expansion force. Background Technology

[0002] During the charge-discharge cycle of a lithium-ion battery, the repeated volume changes in the electrode material caused by lithium-ion insertion and extraction exert alternating expansion stress on the external constraint structure. This stress signal is considered an important mechanical criterion for assessing the risk of lithium plating in the cell, the integrity of the electrode structure, and the overall health status throughout the battery's life cycle.

[0003] In engineering practice, charge-discharge testing systems record event timestamps on a minute-by-minute basis for each step, while thin-film pressure sensor arrays independently acquire expansion force data at high-density sampling rates on a second- or even millisecond-by-second basis. Existing technologies typically employ global timing offset alignment methods (such as fixed time difference correction) to address timing mismatches between the two, for example, by offsetting the entire step log by a preset constant and then aligning it with the expansion force signal. However, there are significant differences in stress response rates between different step segments: the stress response rates of the constant current charging segment and the constant voltage charging segment can differ by several times. This means that a fixed offset cannot simultaneously ensure the registration accuracy of each step segment, and the peak expansion force and rate of change at the step inflection points deviate significantly from the true physical values.

[0004] Meanwhile, the temperature in the high and low temperature cycling aging test fluctuates drastically within a wide range of -20℃ to 60℃. On the one hand, this induces thermal expansion and contraction of the fixtures and structural components, generating parasitic thermal expansion interference that superimposes on the electrochemical signal. On the other hand, the hardening of the binder at low temperatures alters the strain transfer efficiency, causing the expansion force signal of the same process step to exhibit drastically different constitutive characteristics in different temperature ranges. Existing temperature correction methods (such as global compensation based on a single linear expansion coefficient) lack the ability to provide differentiated compensation for different temperature ranges, making it difficult to extract the contribution of thermal expansion from the distorted signal and restore the true lithiation expansion force.

[0005] Furthermore, during long-term cyclic testing, occasional channel failures or storage anomalies in the multi-channel data acquisition system can lead to partial loss of expansion force time-series data in specific work steps. Existing methods for handling missing values ​​mostly rely on linear interpolation or global mean filling based on preceding and following time steps. These methods cannot utilize the inherent spatial correlation of strain fields between multiple channels under expansion and contraction conditions for information complementarity and adaptive recovery, thus compromising the integrity of the expansion force time series and the reliability of subsequent feature analysis. Summary of the Invention

[0006] The technical problem to be solved by this invention is that in high and low temperature cycling accelerated aging tests, due to the triple superposition of frequency mismatch between charge and discharge steps and expansion force timing, thermal expansion distortion coupling, and data loss, traditional methods cannot accurately register, decouple, and completely extract multidimensional mechanical characteristic parameters that reflect the true lithium expansion behavior of the battery.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for extracting multi-source data on battery expansion force includes the following steps: S1: Obtain the set of time stamps for the state transition of the charging and discharging test system and the high-density time-series data collected by the expansion force sensor array, wherein the expansion force sensor array contains multiple sensing units; take the arithmetic mean of the expansion force values ​​of each sensing unit at the same time to obtain the average expansion force signal, which is used as the registration reference signal. S2: Using the average expansion force signal as a reference, a weighted matching cost function for the dynamic reduction window kernel is constructed. The weighted matching cost function includes a time deviation penalty term and a local fluctuation penalty term. The local fluctuation penalty term is calculated based on the average of the absolute values ​​of the first-order differences of the average expansion force signal within the search window. Both the time deviation penalty term and the local fluctuation penalty term are dimensionless normalized. For each step migration point, the expansion force sampling time that minimizes the value of the weighted matching cost function within its search radius is selected as the optimal matching point, resulting in an optimal matching point set. A unified time base coordinate sequence is constructed using the optimal matching point set as anchor points. The original non-uniform time series data of each sensing unit of the expansion force sensor array are resampled onto the unified time base coordinate sequence through third-order Hermite piecewise interpolation to form an initial aligned data matrix. S3: Construct an expansion-compression state memory chain framework, which includes a long short-term memory network and an auxiliary fully connected layer. First, use the training set data in the initial aligned data matrix to perform sequence autoencoder reconstruction training on the long short-term memory network, fixing the encoder parameters. Then, use the hidden state sequence output by the encoder in the training set data to train the auxiliary fully connected layer in a linear regression manner, minimizing the mean square error between the predicted value and the actual value of the expansion force change rate output by the auxiliary fully connected layer. Fix all network parameters and perform forward calculation on all recurrent data to obtain the hidden state sequence at each time step. S4: Reconstruct the stress rate transition matrix envelope from the hidden state sequence and construct a fine-grained dimensional compensation tensor, wherein the fine-grained dimensional compensation tensor includes a temperature-related nonlinear influence factor, which is determined by a pre-calibrated piecewise function model and used to perform temperature compensation on the original aligned expansion force data, remove thermal expansion interference, and obtain a compensated expansion force sequence; the elements of the stress rate transition matrix envelope are the expansion force change rate of the sensing unit multiplied by the rate adjustment factor output by the hidden state and the auxiliary fully connected layer; S5: Construct a diffusion graph model based on the expansion-compression state transition map. Take each moment on the unified time base coordinate sequence as a node, define the edge weights by the similarity of the expansion force change rate features after compensation at the node, construct the graph Laplacian matrix, and use the iterative diffusion reconstruction algorithm to adaptively reconstruct and restore the missing expansion force data in the registration interval, and output a complete three-dimensional expansion feature sequence. S6: Based on the complete three-dimensional expansion feature sequence and combined with the step labels, the charge and discharge cycle is decomposed into step segments. Specific expansion force feature parameters are extracted in each step segment and stored in the feature database according to the sensing unit number, cycle number, and step segment.

[0008] Furthermore, in the above-mentioned method for extracting multi-source data on battery expansion force, the expression for the weighted matching cost function is: in, Indicates the migration point of the work step. With expansion force sampling time The matching cost between them; This is due to time deviation; The search radius is used as the normalized denominator for the time deviation. For The average value of the first-order difference absolute value of the average expansion force signal within the local time window centered on it; The average value of the absolute difference in the average expansion force between adjacent sampling points in the original training set is used as the normalized denominator for local fluctuations. This is the time deviation penalty coefficient. This is the penalty coefficient for local fluctuations.

[0009] Furthermore, in the above-mentioned method for extracting multi-source data on battery expansion force, the Long Short-Term Memory (LSTM) network calculates the hidden state at each time step according to the following update equation. : in, for The input feature vector at time t, This is the implicit state from the previous moment. Initialize to a zero vector; These are the input gate, forget gate, and output gate vectors, respectively. This is the weight matrix from the input layer to the hidden layer. This is the circular weight matrix from hidden layer to hidden layer. It is the bias vector; For the sigmoid function, It is the ReLU activation function. Represents the Hadamard product; the rate adjustment factor vector output by the auxiliary fully connected layer is... ,in and The weight matrix and bias of the auxiliary fully connected layer are respectively obtained through analytical linear regression. Confirmed, among which The hidden state matrix of the training set, For its Moore-Penrose pseudoreversal, To train the expansion force rate of change matrix, and These are the mean values ​​on the training set; the envelope of the stress rate transition matrix. The elements are: in For the first Each sensing unit in The original alignment expansion force value at time t. The time resolution of the unified time base coordinate sequence. The first of the rate adjustment factor vectors Each component.

[0010] Furthermore, in the above-mentioned method for extracting multi-source data on battery expansion force, the fine-grained dimensional compensation tensor... elements Calculated using the following compensation formula: in The start time of the test, For integration variables; For the first Each sensing unit in The original alignment expansion force value at time; For strain rate, through the first The first derivative of the expansion force of each sensing unit with respect to time divided by the calibration stiffness coefficient of the sensing unit get; This is the temperature value; The nonlinear influencing factor related to temperature is determined by a pre-calibrated piecewise function model, which introduces a low-temperature correction term when the temperature is below a set threshold. The coefficients are empirical fitting coefficients; the integrals are calculated point-by-point on the discrete grid of the unified time-based coordinate sequence using the numerical trapezoidal rule.

[0011] Furthermore, in the above-mentioned multi-source data extraction method for battery expansion force, in the diffusion map model, the nodes... and Edge weights between Calculated using the following formula: in and They are nodes and The expansion rate eigenvector at point , the expansion rate eigenvector of each node at point . Each component represents the rate of change of the expansion force after compensation. The Gaussian kernel bandwidth parameter controls the similarity decay rate. This is an indicator function; it takes the value 1 when the condition within the parentheses is met, and 0 otherwise. The diffusion window threshold determines whether a connection is established between node pairs whose time interval is less than this threshold; the graph Laplacian matrix... ,in This is the weight matrix. The degree is a diagonal matrix; the iteration step size of the iterative diffusion reconstruction algorithm satisfies , is the largest eigenvalue of the graph Laplacian matrix.

[0012] Furthermore, in the above-mentioned method for extracting multi-source data on battery expansion force, the specific expansion force feature parameters include: the maximum compensated expansion force of the sensing unit within each process step as the peak value of the segmented expansion force. The expansion rate of the charging section is obtained by fitting the slope of the expansion force over time after compensation within the work step segment using linear regression. and discharge segment contraction rate The difference in expansion force between the start and end times of the resting period after compensation is used as the springback deformation of the resting period. The degree of abrupt change in the rate of change of expansion force during the switching between adjacent work steps is used as the cross-segment differential variability. .

[0013] The battery expansion force multi-source data extraction system includes a multi-source data acquisition and parsing engine, a process step and expansion force data registration module, an expansion-compression state memory chain encoder decoding module, a diffusion map missing value correction module, and a structured output module. Each module executes the steps of the above method in sequence.

[0014] The beneficial effects of this invention are as follows: By constructing a bidirectional nearest neighbor interpolation strategy with a dynamic reduction window kernel and introducing normalization parameters calculated based on the original training data, the dependence on subsequent data during registration is eliminated, solving the technical problem of insufficient registration accuracy at the step transition point in traditional global time alignment methods. This reduces the registration error at the step switching time from the minute level to the millisecond level, achieving precise synchronization between the charge / discharge steps and the high-density time series of expansion force in high- and low-temperature cyclic accelerated aging test scenarios. Furthermore, by constructing an expansion-compression state memory chain framework and using a sequence autoencoder task to perform unsupervised pre-training on the long short-term memory network, combined with a linear regression auxiliary layer to adaptively adjust the weight contribution of each sensing unit in the stress rate transition matrix, the technical barrier to deep feature extraction of expansion force time series data is overcome. This not only automatically captures the deformation memory effect dependent on historical states during expansion-compression but also decouples the pure mechanical features corresponding to electrochemical expansion from the original signal. Finally, by introducing a pre-calibrated piecewise temperature compensation function… This method differentiates the thermal expansion distortion in the low-temperature range, resolving the technical contradiction that the expansion force signal is severely distorted due to temperature coupling under alternating high and low temperature conditions, making it impossible to reconstruct the true lithium-ionized expansion stress. This improves the accuracy of effective expansion force extraction in the low-temperature range. Furthermore, by constructing a sparse graph Laplacian matrix based on time window truncation and using the Gershgorin disk theorem to determine the upper bound of the iteration step size, efficient reconstruction of the missing interval signal in the spectral domain is achieved. This overcomes the engineering obstacles of time sequence breaks and the inability to continuously extract feature parameters due to data loss in long-term aging tests. Finally, by finely decomposing the process steps and extracting four-dimensional feature parameters (peak value, rate, rebound, and abrupt change), a multi-dimensional expansion feature database for battery lifecycle management is constructed. This enables the method to adaptively complete the entire chain of automated extraction from raw multi-source heterogeneous data to structured expansion and aging features without manual intervention, providing a reliable quantitative basis for engineering applications such as early lithium plating warning, fast charging strategy optimization, and battery life prediction. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a method for extracting multi-source data on battery expansion force according to a specific embodiment of the present invention. Detailed Implementation

[0016] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0017] This embodiment relates to a method for extracting multi-source data on battery expansion force; This embodiment uses a high-energy-density NCM ternary soft-pack lithium-ion battery (rated capacity 10Ah, rated voltage 3.65V) as the test object, and implements the method described in this invention on a high and low temperature cycling accelerated aging test platform. The test environment temperature is set to alternate between 45℃ (high temperature range) and 10℃ (low temperature range), with each temperature range maintained for 30 minutes and a temperature chamber switching time of approximately 5 minutes. The battery is placed in a special fixture, which uses a constant mechanical constraint force (approximately 50N) to keep the large surface of the battery in close contact with each sensing unit, ensuring the consistency of the expansion force transmission path. Eight thin-film pressure sensing units (model: FlexiForce A201, range 0~1000N, sampling rate 50Hz) are symmetrically arranged on both sides of the large surface of the battery to form an expansion force sensor array, and each sensing unit independently collects the expansion force timing signal. The charge-discharge test system (Arbin BT-5HC) synchronously records the time stamps of each step's state transition and current, voltage, and capacity data. The step sequence is as follows: 0.5C constant current charging to 4.2V, switching to constant voltage charging to the cutoff current of 0.1C, resting for 10 minutes, 1C constant current discharging to 2.8V, and resting for 10 minutes. This constitutes one complete cycle, which is executed continuously for 60 cycles. Temperature field data is synchronously acquired by a K-type thermocouple at the center of the battery surface at a sampling rate of 1Hz. All data acquisition and signal processing are performed on an industrial control computer, where the GPU is used for parallel acceleration of LSTM training, and the CPU is used for data preprocessing and feature extraction.

[0018] The method includes the following steps: S1: Obtain the set of time stamps for the state transition of the charging and discharging test system and the high-density time-series data collected by the expansion force sensor array, wherein the expansion force sensor array contains multiple sensing units; take the arithmetic mean of the expansion force values ​​of each sensing unit at the same time to obtain the average expansion force signal, which is used as the registration reference signal. In this embodiment, before the test begins, the step log file (CSV format) of the charge / discharge test system and the data file (SQLite database format) of the expansion force acquisition system are input into the parsing engine of the method described in this invention. The adaptive format discriminator identifies the column labels (timestamp, step_index, current, voltage, capacity, temperature, etc.) in the first row of the CSV file and the table header information (table_force contains sensor_id, time_stamp, force_value columns) of the SQLite file, and automatically calls the corresponding reader accordingly. After parsing, the set of step state transition timestamps is extracted from the step log. ,in This represents the total number of migration points, with each migration point corresponding to the state transition time of a work step. In this test sequence, each loop contains 4 migration points (charging start, charging end / rest start, discharging start, discharging end / rest start), and a total of 60 loops. One migration point. Independent time-series data from eight sensing units were extracted from the expansion force database. , Number the sensing unit. The set of sampling time points for expansion force is denoted as . At a sampling frequency of 50Hz, Approximately 4.5 × 10 6 Magnitude.

[0019] S2: Using the average expansion force signal as a reference, a weighted matching cost function for the dynamic reduction window kernel is constructed. The weighted matching cost function includes a time deviation penalty term and a local fluctuation penalty term. The local fluctuation penalty term is calculated based on the average of the absolute values ​​of the first-order differences of the average expansion force signal within the search window. Both the time deviation penalty term and the local fluctuation penalty term are dimensionless normalized. For each step migration point, the expansion force sampling time that minimizes the value of the weighted matching cost function within its search radius is selected as the optimal matching point, resulting in an optimal matching point set. A unified time base coordinate sequence is constructed using the optimal matching point set as anchor points. The original non-uniform time series data of each sensing unit of the expansion force sensor array are resampled onto the unified time base coordinate sequence through third-order Hermite piecewise interpolation to form an initial aligned data matrix. In this embodiment, considering the inter-channel differences in data from the eight sensing units at the same time, the arithmetic mean of the expansion force values ​​of each sensing unit at the same time is first taken to obtain the average expansion force signal. , as a reference signal for registration.

[0020] Before registration, normalization parameters are calculated from the raw sampled data of the first 10 cycles (i.e., the training set). This is done for all adjacent sampled point pairs within the first 10 cycles. Calculate the absolute value of the average expansion force difference. The average value is taken as the global fluctuation scale: in This is the total number of sampling points in the first 10 cycles, in this embodiment. This parameter does not depend on any registration results and can be calculated directly from the raw data.

[0021] For each work step migration point Set the search radius around it. (In this embodiment, a 50Hz sampling rate corresponds to an adjacent sampling interval of 20ms.) (Covering approximately 25 sampling points). Within the search interval Within this process, the migration point and the time of each expansion force sampling are calculated one by one. (satisfy Weighted matching cost between: In the above formula, Time deviation (unit: seconds). For local windows The average of the absolute values ​​of the first-order difference of the internal average expansion force signal, where Number of points on one side : The mean absolute difference (in N) of the original data in the aforementioned training set is used as the normalized denominator for the expansion force fluctuation. This is the time deviation penalty coefficient. Here, is the local fluctuation penalty coefficient; both are dimensionless quantities. We choose a coefficient that makes . Minimum value As the optimal matching point: For all Solve for each migration point one by one to obtain the optimal matching point set. .by Construct a unified time-based coordinate sequence for anchor points Equal time interval resolution Seconds to generate grid points For each sensing unit Its original non-uniform sampling points Resampling was performed using third-order Hermitian piecewise interpolation. The above yields the aligned sequence. The alignment sequences of all eight sensing units are stacked row by row to form the initial alignment data matrix. .

[0022] S3: Construct an expansion-compression state memory chain framework, which includes a long short-term memory network and an auxiliary fully connected layer. First, use the training set data in the initial aligned data matrix to perform sequence autoencoder reconstruction training on the long short-term memory network, fixing the encoder parameters. Then, use the hidden state sequence output by the encoder in the training set data to train the auxiliary fully connected layer in a linear regression manner, minimizing the mean square error between the predicted value and the actual value of the expansion force change rate output by the auxiliary fully connected layer. Fix all network parameters and perform forward calculation on all recurrent data to obtain the hidden state sequence at each time step. In this embodiment, temperature compensation parameters were calibrated separately before the formal test. A spare battery of the same model was fully discharged to 2.8V and placed in a temperature chamber. Temperature points were sequentially set as -5℃, 0℃, 5℃, 10℃, 15℃, 20℃, 25℃, 30℃, 35℃, 40℃, 45℃, and 50℃. After holding at each temperature point for 60 minutes, the stable expansion force readings of eight sensor units were recorded, and the average value was taken. Define the normalized thermal expansion factor. The Levenberg-Marquardt algorithm is used to fit the piecewise function model: fixed Segmentation threshold The selection criteria are as follows: this temperature is close to the freezing point of the electrolyte (typically, the freezing point of NCM system electrolytes is approximately -20℃ to -10℃, but the glass transition temperature of binders is usually in the range of 0℃ to -20℃). In this experimental system, preliminary experimental observations show that when the temperature is below 15℃, the slope of the expansion force changes with temperature shows a significant inflection point (deviating from the exponential decay law at room temperature). Therefore, this temperature point is set as the segmented threshold. For batteries with different chemical systems, those skilled in the art can use the same method, i.e., plotting... Analyze the curve and observe its curvature changes to determine the applicable curve. Values. Initial values ​​of the fitted parameters. Maximum number of iterations: 100; convergence tolerance: Calibration results Coefficient of determination The calibration was repeated three times, and the relative standard deviation of each parameter was less than 5%, indicating that the calibration results have good repeatability. This represents the low-temperature correction factor, used to quantify the situation when the temperature is below a set threshold. When the strain transfer efficiency of materials such as battery binders or electrolytes changes due to hardening / solidification, this is an additional enhancing factor. The above calibration temperature range covers the temperature range of the test conditions in this embodiment (10°C to 45°C), and the segmented thresholds... Within the calibration range, it ensures that the compensation model has experimental data to support it within the operating temperature range.

[0023] For each moment Constructing an 11-dimensional input feature vector The expansion force amplitudes of the eight sensing units are shown in sequence. The overall first-order difference mean of expansion force (Boundary point one-sided difference), two components of temperature state feature code.

[0024] The dataset is divided as follows: 60 loops are arranged in chronological order, with the first 10 loops ( to , ) as the training set for learning the parameters of the LSTM network; loops 11-12 ( to , The training set, validation set, and test set are used as the validation set for monitoring the early stopping strategy; loops 13-60 serve as the test set for final feature extraction and performance evaluation. The training set, validation set, and test set are strictly non-overlapping. All features are Z-score standardized before being fed into the network; the mean and standard deviation are calculated only from the training set data, i.e., for the first 60 loop... each feature component The validation and test sets use the same set. Standardize the process to ensure no data leaks. This is the value of a scalar component of the input feature vector. Representing the The mean of each input feature component on the training set (i.e., the arithmetic mean of the training set samples). Representing the The standard deviation of each input feature on the training set.

[0025] LSTM network hidden layer dimension Input layer to hidden layer weight matrix Hidden layer cyclic weight matrix Bias vector All weight matrices are initialized using Xavier uniform initialization, and biases are initialized to zero vectors. Training employs a sequence autoencoder reconstruction task with a linear decoder. Hidden state Reconstructed The training data batches are constructed using a sliding window approach: sliding along the time axis with a step size of 1, each batch contains a subsequence of 20 consecutive time steps (i.e., the BPTT unfolded length), with an overlap of 19 time steps between adjacent batches. The loss function at each time step within each batch is used in gradient calculation. Loss function: Adam optimizer, learning rate The batch size is 32, the BPTT unrolling time step is 20, and the maximum number of training epochs is 200. The early stopping strategy is implemented as follows: after each complete training epoch, the average loss on the validation set is calculated. If the current validation set loss is lower than the historical best value, the current model parameters are saved; if the validation set loss is not lower than the historical best value for 15 consecutive epochs, training is stopped and the saved best model parameters are restored. In this embodiment, training stops at epoch 147 (the optimal validation set loss of 0.031 is reached at epoch 132), with a training set loss of 0.023 and a validation set loss of 0.031.

[0026] After training, fix the encoder parameters (i.e.) For each time step in the training set. Forward computation of hidden states Then, using the hidden states as input, a linear regression model is trained to predict the rate of change of expansion force at the next time step. The linear regression model takes the form of: ,in . Represents the weight matrix of the auxiliary fully connected layer. The bias vector representing the auxiliary fully connected layer is obtained by minimizing the mean squared error on the training set. Solve using ordinary least squares. and Its analytical solution is ,in The hidden state matrix (number of rows) (The number of columns is 64) For its Moore-Penrose pseudoreversal, and These are the means of the training set. The values ​​obtained in this embodiment are... and Save as a binary file for later use.

[0027] At this point, all network parameters (encoder parameters, ...) are... , All parameters have been determined. These parameters will remain fixed during subsequent feature extraction in all loops.

[0028] S4: Reconstruct the stress rate transition matrix envelope from the hidden state sequence and construct a fine-grained dimensional compensation tensor, wherein the fine-grained dimensional compensation tensor includes a temperature-related nonlinear influence factor, which is determined by a pre-calibrated piecewise function model and used to perform temperature compensation on the original aligned expansion force data, remove thermal expansion interference, and obtain a compensated expansion force sequence; the elements of the stress rate transition matrix envelope are the expansion force change rate of the sensing unit multiplied by the rate adjustment factor output by the hidden state and the auxiliary fully connected layer; In this embodiment, for each time point The hidden state is calculated using the LSTM update equation: Initialize as a zero vector. , , For Hadamard products.

[0029] Reconstructing the stress-rate transition matrix envelope from the hidden state sequence Define the rate adjustment factor vector. , its first Each component is ,but: in For the first Each sensing unit in The original alignment expansion force value at time (Note: This is the value before compensation, because...) (For subsequent compensation).

[0030] Constructing fine-grained dimensional compensation tensors The compensation formula is: in , N / mm is the calibration stiffness coefficient of the sensing unit. These are the empirical fitting coefficients. The piecewise function calibrated in S4 is used. Integration is performed point-by-point on the discrete grid using the numerical trapezoidal rule. Compensation tensor. The element is Finally, a three-dimensional hierarchical feature tensor is extracted. The dimensions are sensing unit, time, and feature type (3: original force, stress rate change envelope, and compensated force).

[0031] To verify whether the hidden states obtained during training effectively represent the physical states of the expansion-compression process, the following method was used: For the validation set data, the hidden states at each time step were calculated. The mutual information between the hidden state and the direction of the expansion force change (expansion or contraction) at the next moment is calculated. The mutual information is calculated using an equal-width discretization method, dividing each dimension of the hidden state into 10 equal intervals. The direction of the expansion force change is binarized as +1 (expansion) and -1 (contraction). In this embodiment, the mutual information value is 0.47 (significantly higher than the random baseline of 0.02), indicating that the hidden state carries effective information that helps determine the expansion direction. Simultaneously, the average Euclidean distance of the hidden state during the charge-discharge cycle is calculated. The distance between the centers of the hidden states in the charging and discharging segments is 2.3 times the standard deviation within the cell, proving that the hidden state has good discriminative power for different stages. The above verification methods provide an objective evaluation basis for the network training effect.

[0032] S5: Construct a diffusion graph model based on the expansion-compression state transition map. Take each moment on the unified time base coordinate sequence as a node, define the edge weights by the similarity of the expansion force change rate features after compensation at the node, construct the graph Laplacian matrix, and use the iterative diffusion reconstruction algorithm to adaptively reconstruct and restore the missing expansion force data in the registration interval, and output a complete three-dimensional expansion feature sequence. In this embodiment, the simulation of the third sensing unit during the 12th to 18th cycles ( The data was completely lost. A diffusion graph model was used to reconstruct the missing data for this region.

[0033] Define node set ,node Corresponding time .node and Edge weights between: For an 8-dimensional expansion rate eigenvector, the first... Each component is One-sided difference at boundary points. For Gaussian kernel bandwidth, This is the diffusion window threshold.

[0034] Construct a symmetric weight matrix , matrix number Line 1 Column elements Construct a degree diagonal matrix The first one on its diagonal element Graph Laplace matrix .

[0035] This embodiment is for To address the storage and computation issues, the following strategy is adopted: Due to exist When it is zero, It is a banded sparse matrix with a bandwidth of approximately The data is stored using a compressed sparse row format, storing only non-zero elements and their row and column indices. When Lanczos iterates to find the first 20 smallest eigenvalues, the convergence criterion is set to a residual norm less than... The computational cost of each matrix-vector multiplication is In this embodiment, the iterations are approximately 200 times to converge to the specified accuracy, with a total computation time of approximately 30 minutes.

[0036] Iteration step size According to Gershgorin's disk theorem , , Meeting convergence conditions . Represents the Graph Laplace matrix The largest eigenvalue. For the th Each sensing unit will transmit the compensated expansion force sequence. (Including missing values) are considered as graph signals. Missing positions are set to 0. The Lanczos algorithm is used for calculation. The former The smallest eigenvectors Iterative Refactoring: After each iteration, the known position signal value is forcibly assigned back to the true observation value, iterating until the L2 norm of the changes between two consecutive iterations is less than 1. The process stops at a certain point (convergence occurs after approximately 1500 iterations in this embodiment), yielding the reconstructed sequence. All eight sensing units were reconstructed individually to obtain a complete three-dimensional dilation feature sequence. .

[0037] S6: Based on the complete three-dimensional expansion feature sequence and combined with the step labels, the charge and discharge cycle is decomposed into step segments. Specific expansion force feature parameters are extracted in each step segment and stored in the feature database according to the sensing unit number, cycle number, and step segment.

[0038] In this embodiment, based on Combined with work step labels (mapped to) The charge / discharge cycle is decomposed into a constant current charging segment, a constant voltage charging segment, a resting segment (after charging), a constant current discharging segment, and a resting segment (after discharging). Within each segment, the following is extracted: Peak segmented expansion force For example, if the constant current charging segment of the 5th cycle starts from... arrive ,but . represent Take the maximum value and iterate through all time step indices within that work step segment. Find the largest value.

[0039] expansion rate and contraction rate The first The average rate of change (slope) of the compensated expansion force over time for each sensing unit during the charging and discharging phases is obtained by least-squares fitting using linear regression. The unified calculation formula is: ; Springback deformation of the stationary segment (the first) (The springback deformation of each sensing unit during the resting period) ; The starting time of the static section is at Index in; The end time of the settling period is at The index in.

[0040] Cross-segment differential variability (the magnitude of change in expansion rate when switching between adjacent steps) . The expansion rate (positive value) of the subsequent charging stage represents the rate of increase of the expansion force during the charging process of the next step. The contraction rate of the preceding discharge segment (in absolute value) represents the rate (in absolute value) of decrease in expansion force during the previous discharge step.

[0041] The above characteristic parameters are numbered according to the sensing unit. The features are categorized by cycle number and work step segment, and stored in a time-series database (TimescaleDB) format. The feature database can be retrieved using a composite primary key (cycle number, work step segment, sensor unit ID) for precise queries, or by conditions (cycle number range, sensor unit ID) for range queries. It supports aggregated statistics and time-series trend analysis by any dimension. This database is used for subsequent anomaly detection (such as categorizing a specific segment of a cycle). The analysis includes: comparison with the historical mean ± 3 standard deviations (exceeding which triggers an alert); work condition retrospection (recalling the corresponding stored characteristic curves for playback based on the specified cycle number and work step); and full life cycle expansion force change trend analysis (plotting the same work step in each cycle). The curve showing the change over the number of cycles is used to assess the aging process.

[0042] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for extracting multi-source data on battery expansion force, characterized in that, Includes the following steps: S1: Obtain the set of time stamps for the state transition of the charging and discharging test system and the high-density time-series data collected by the expansion force sensor array, wherein the expansion force sensor array contains multiple sensing units; take the arithmetic mean of the expansion force values ​​of each sensing unit at the same time to obtain the average expansion force signal, which is used as the registration reference signal. S2: Using the average expansion force signal as a reference, a weighted matching cost function for the dynamic reduction window kernel is constructed. The weighted matching cost function includes a time deviation penalty term and a local fluctuation penalty term. The local fluctuation penalty term is calculated based on the average of the absolute values ​​of the first-order differences of the average expansion force signal within the search window. Both the time deviation penalty term and the local fluctuation penalty term are dimensionless normalized. For each step migration point, the expansion force sampling time that minimizes the value of the weighted matching cost function within its search radius is selected as the optimal matching point, resulting in an optimal matching point set. A unified time base coordinate sequence is constructed using the optimal matching point set as anchor points. The original non-uniform time series data of each sensing unit of the expansion force sensor array are resampled onto the unified time base coordinate sequence through third-order Hermite piecewise interpolation to form an initial aligned data matrix. S3: Construct an expansion-compression state memory chain framework, which includes a long short-term memory network and an auxiliary fully connected layer. First, use the training set data in the initial aligned data matrix to perform sequence autoencoder reconstruction training on the long short-term memory network, fixing the encoder parameters. Then, use the hidden state sequence output by the encoder in the training set data to train the auxiliary fully connected layer in a linear regression manner, minimizing the mean square error between the predicted value and the actual value of the expansion force change rate output by the auxiliary fully connected layer. Fix all network parameters and perform forward calculation on all recurrent data to obtain the hidden state sequence at each time step. S4: Reconstruct the stress rate transition matrix envelope from the hidden state sequence and construct a fine-grained dimensional compensation tensor, wherein the fine-grained dimensional compensation tensor includes a temperature-related nonlinear influence factor, which is determined by a pre-calibrated piecewise function model and used to perform temperature compensation on the original aligned expansion force data, remove thermal expansion interference, and obtain a compensated expansion force sequence; the elements of the stress rate transition matrix envelope are the expansion force change rate of the sensing unit multiplied by the rate adjustment factor output by the hidden state and the auxiliary fully connected layer; S5: Construct a diffusion graph model based on the expansion-compression state transition map. Take each moment on the unified time base coordinate sequence as a node, define the edge weights by the similarity of the expansion force change rate features after compensation at the node, construct the graph Laplacian matrix, and use the iterative diffusion reconstruction algorithm to adaptively reconstruct and restore the missing expansion force data in the registration interval, and output a complete three-dimensional expansion feature sequence. S6: Based on the complete three-dimensional expansion feature sequence and combined with the step labels, the charge and discharge cycle is decomposed into step segments. Specific expansion force feature parameters are extracted in each step segment and stored in the feature database according to the sensing unit number, cycle number, and step segment.

2. The method for extracting multi-source data on battery expansion force according to claim 1, characterized in that, The expression for the weighted matching cost function is: in, Indicates the migration point of the work step. With expansion force sampling time The matching cost between them; This is due to time deviation; The search radius is used as the normalized denominator for the time deviation. For The average value of the first-order difference absolute value of the average expansion force signal within the local time window centered on it; The average value of the absolute difference in the average expansion force between adjacent sampling points in the original training set is used as the normalized denominator for local fluctuations. This is the time deviation penalty coefficient. This is the penalty coefficient for local fluctuations.

3. The method for extracting multi-source data on battery expansion force according to claim 1, characterized in that, The Long Short-Term Memory network calculates the hidden state at each time step according to the following update equation. : in, for The input feature vector at time t, This is the implicit state from the previous moment. Initialize to a zero vector; These are the input gate, forget gate, and output gate vectors, respectively. This is the weight matrix from the input layer to the hidden layer. This is the circular weight matrix from hidden layer to hidden layer. It is the bias vector; For the sigmoid function, It is the ReLU activation function. Represents the Hadamard product; the rate adjustment factor vector output by the auxiliary fully connected layer is... ,in and The weight matrix and bias of the auxiliary fully connected layer are respectively obtained through analytical linear regression. Confirmed, among which The hidden state matrix of the training set, For its Moore-Penrose pseudoreversal, To train the expansion force rate of change matrix, and These are the mean values ​​on the training set; the envelope of the stress rate transition matrix. The elements are: in For the first Each sensing unit in The original alignment expansion force value at time t. The time resolution of the unified time base coordinate sequence. The first of the rate adjustment factor vectors Each component.

4. The method for extracting multi-source data on battery expansion force according to claim 1, characterized in that, The fine-grained dimensional compensation tensor elements Calculated using the following compensation formula: in The start time of the test, For integration variables; For the first Each sensing unit in The original alignment expansion force value at time; For strain rate, through the first The first derivative of the expansion force of each sensing unit with respect to time divided by the calibration stiffness coefficient of the sensing unit get; This is the temperature value; The nonlinear influencing factor related to temperature is determined by a pre-calibrated piecewise function model, which introduces a low-temperature correction term when the temperature is below a set threshold. The coefficients are empirical fitting coefficients; the integrals are calculated point-by-point on the discrete grid of the unified time-based coordinate sequence using the numerical trapezoidal rule.

5. The method for extracting multi-source data on battery expansion force according to claim 1, characterized in that, In the diffusion graph model, nodes and Edge weights between Calculated using the following formula: in and They are nodes and The expansion rate eigenvector at point , the expansion rate eigenvector of each node at point . Each component represents the rate of change of the expansion force after compensation. The Gaussian kernel bandwidth parameter controls the similarity decay rate. This is an indicator function; it takes the value 1 when the condition within the parentheses is met, and 0 otherwise. The diffusion window threshold determines whether a connection is established between node pairs whose time interval is less than this threshold; the graph Laplacian matrix... ,in This is the weight matrix. The degree is a diagonal matrix; the iteration step size of the iterative diffusion reconstruction algorithm satisfies , is the largest eigenvalue of the graph Laplacian matrix.

6. The method for extracting multi-source data on battery expansion force according to claim 1, characterized in that, The specific expansion force characteristic parameter includes: the maximum compensated expansion force of the sensing unit in each work step segment as the peak value of the segmented expansion force. The expansion rate of the charging section is obtained by fitting the slope of the expansion force over time after compensation within the work step segment using linear regression. and discharge segment contraction rate The difference in expansion force between the start and end times of the resting period after compensation is used as the springback deformation of the resting period. The degree of abrupt change in the rate of change of expansion force during the switching between adjacent work steps is used as the cross-segment differential variability. .

7. A multi-source data extraction system for battery expansion force, characterized in that, It includes a multi-source data acquisition and parsing engine, a process step and expansion force data registration module, an expansion-compression state memory chain encoder decoding module, a diffusion map missing value correction module, and a structured output module, each module sequentially executing the steps of the method described in any one of claims 1 to 6.

Images