Lithium battery residual life prediction method based on random degradation state space model

Abstract

This invention discloses a method for predicting the remaining life of lithium batteries based on a stochastic degradation state-space model, belonging to the field of battery life prediction technology. The method includes: extracting degradation feature vectors and auxiliary feature vectors for each charge-discharge cycle based on the full life-cycle data of the lithium battery; outputting the composite health state of the corresponding charge-discharge cycle using a first neural network model; obtaining the predicted remaining life value after the end of the corresponding charge-discharge cycle through a degradation process model; obtaining the true value of the degradation capacity for each charge-discharge cycle and constructing a composite health state degradation observation equation using a second neural network model; calculating the total loss function to update the model parameters until the optimal parameter set of the corresponding model is output; constructing a stochastic degradation state-space model; and using particle filtering to update the model a preset number of times to obtain the predicted remaining life of the target lithium battery. This method enables online prediction of the remaining life of lithium batteries and has high accuracy and strong robustness.

Description

Technical Field

[0001] This invention belongs to the field of battery life prediction technology, specifically relating to a method for predicting the remaining life of lithium batteries based on a stochastic degradation state-space model. Background Technology

[0002] Predicting the remaining life of lithium-ion batteries is a crucial step in ensuring the safe operation and maintenance of new energy vehicle power systems and grid-scale energy storage devices. The core bottleneck lies in clarifying the complex coupling relationship between the battery's internal multi-scale failure mechanisms and external time-varying operating conditions. From the perspective of material degradation mechanisms, capacity decay is mainly attributed to the continuous evolution at the microscopic level, including phase transitions in the electrode active material structure, electrolyte oxidation consumption, and interfacial side reactions. However, under actual dynamic operating conditions, electrochemical polarization, concentration polarization, and thermo-mechanical-electrochemical multi-physics fields are deeply intertwined, forming a strongly nonlinear synergistic degradation mechanism. This deep coupling not only rapidly accelerates the progress of microscopic failure paths such as the dynamic evolution of the solid electrolyte interfacial film and lithium metal deposition, but also causes the material aging behavior to exhibit distinct time-varying characteristics, masking early performance degradation signs with strong system noise. Therefore, traditional linear extrapolation methods based on fixed criteria are insufficient to accurately identify and track potential degradation trajectories.

[0003] To address this challenge, current RUL (Remaining Useful Life) prediction methods are mainly developing along three paths: mechanistic models, purely data-driven models, and a combined mathematical-model framework integrating both. Physical mechanism models attempt to quantify the degradation process through electrochemical equations, but their accuracy is limited by the completeness of modeling complex underlying mechanisms. Data-driven methods, leveraging the powerful data mining capabilities of machine learning and deep learning, have become a research hotspot: recurrent neural networks construct end-to-end prediction frameworks to capture temporal dependencies; kernel methods utilize nonlinear mappings to characterize degradation trajectories and provide probabilistic uncertainty assessments.

[0004] The advantage of the mathematical-model linkage framework lies in its effective integration of mathematical theoretical insights and data intelligence. This framework significantly improves the model's ability to identify true degradation features by introducing stochastic process theory or signal processing mechanisms to address the non-stationarity of the original capacity sequence. Simultaneously, it combines intelligent optimization algorithms to efficiently search and optimize key model parameters, accelerating model convergence and significantly improving prediction accuracy and stability. Furthermore, some studies have further enhanced the model's ability to focus on key degradation features by introducing strategies such as attention mechanisms.

[0005] While purely data-driven methods bypass complex mechanistic modeling, their performance bottlenecks lie in their heavy reliance on high-quality, large-scale labeled data and the challenges of feature engineering. The mathematical-model linkage framework is demonstrating its unique potential: by integrating mathematically model-guided signal decomposition techniques with deep learning and probabilistic models, it achieves efficient extraction and probabilistic prediction of multi-scale decay signals; it significantly improves the efficiency of model parameter tuning using nature-inspired intelligent optimization algorithms; and it alleviates the interpretability challenges posed by the "black box" nature of deep neural networks to some extent by utilizing physical constraints or simulated data. However, while advancements in sensors and computing power provide more data, insufficient model interpretability, challenges in adapting to different battery types and operating conditions, and the high cost of acquiring data throughout the entire battery lifecycle remain obstacles to large-scale application. The key direction for future breakthroughs lies in deepening the mathematical-model linkage, exploring combinations such as generative pre-trained models and self-supervised learning, and overcoming existing bottlenecks through cross-domain knowledge transfer and multi-source data fusion. Summary of the Invention

[0006] The purpose of this invention is to address the above-mentioned problems by proposing a method for predicting the remaining life of lithium batteries based on a stochastic degradation state-space model, which enables online prediction of the remaining life of lithium batteries and has the advantages of high accuracy and strong robustness.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] The present invention proposes a method for predicting the remaining lifetime of lithium batteries based on a stochastic degradation state-space model, comprising the following steps:

[0009] S1. Based on the collected full life cycle data of the lithium battery, extract the degradation feature vector and auxiliary feature vector corresponding to each charge-discharge cycle;

[0010] S2. Based on the degradation feature vector of each charge-discharge cycle, the first neural network model is used to output the composite health status of the corresponding charge-discharge cycle.

[0011] S3. Based on the composite health status of charge-discharge cycles, a nonlinear Wiener process model is used to construct a degradation process model, and the remaining lifetime prediction value after the end of the corresponding charge-discharge cycle is obtained through the degradation process model.

[0012] S4. Obtain the true value of the degradation capacity for each charge-discharge cycle, and construct the composite health state degradation observation equation using the second neural network model based on the composite health state and auxiliary feature vector of the corresponding charge-discharge cycle.

[0013] S5. Calculate the total loss function to update the model parameters of the first neural network model, the degradation process model and the second neural network model until the training ends, and output the optimal parameter set of the corresponding model.

[0014] S6. Calculate the variance of the model parameters of the nonlinear Wiener process model based on the trained degradation process model, obtain the nonlinear Wiener process model that follows a normal distribution, and construct a stochastic degradation state space model by combining the composite health state degradation observation equation of the trained second neural network model.

[0015] S7. Obtain the composite health state of the target lithium battery in the 0th charge-discharge cycle based on the trained first neural network model, and use particle filtering to update the composite health state a preset number of times based on the random degradation state space model. Calculate the remaining lifetime prediction value corresponding to the updated composite health state as the remaining lifetime prediction result of the target lithium battery.

[0016] Preferably, the full life cycle data of the lithium battery includes the voltage sequence of the constant current charging stage, the current sequence of the constant voltage charging stage, and the voltage sequence of the constant current discharging stage in each charge-discharge cycle; the degradation feature vector is a vector composed of the normalized spectral entropy and Hilbert spectral energy of the voltage sequence of the constant current charging stage, the spectral entropy and Hilbert spectral energy of the current sequence of the constant voltage charging stage, the spectral entropy and Hilbert spectral energy of the voltage sequence of the constant current discharging stage, the time of the constant current charging stage, the time of the constant voltage charging stage, the time of the constant current discharging stage, the root mean square value of the voltage in the voltage sequence of the constant current charging stage, the root mean square value of the current in the current sequence of the constant voltage charging stage, and the root mean square value of the voltage in the voltage sequence of the constant current discharging stage; and the auxiliary feature vector is a vector composed of at least some elements of the degradation feature vector.

[0017] Preferably, a degradation process model is constructed using a nonlinear Wiener process model based on the combined health state of charge-discharge cycles, as detailed below:

[0018] S31. Construct a nonlinear Wiener process model based on the combined health state of charge-discharge cycles;

[0019] S32. Obtain the probability density function of the composite health state increment based on the nonlinear Wiener process model. The composite health state increment is the difference between the composite health states of two adjacent charge-discharge cycles.

[0020] S33. Calculate the remaining lifetime value after each charge-discharge cycle based on the nonlinear Wiener process model to obtain the first-failure time, and calculate the probability density function of the first-failure time based on the probability density function of the composite health state increment.

[0021] S34. The expected value of the remaining lifetime after the end of the charge-discharge cycle is calculated as the degradation process model based on the potential value of the remaining lifetime after the end of the charge-discharge cycle and the probability density function of the first failure time.

[0022] Preferably, the total loss function is obtained by weighted summation of the remaining lifetime prediction error term, the nonlinear Wiener process constraint term, and the degradation capacity prediction error term, and the gradient descent method is used to minimize the total loss function to update the model parameters of the first neural network model, the degradation process model, and the second neural network model.

[0023] Preferably, the remaining life prediction error term is the mean square error of the remaining life calculated based on the predicted remaining life value and the actual remaining life value after each charge-discharge cycle in the entire life cycle of each lithium battery.

[0024] The constraint terms for the nonlinear Wiener process are obtained as follows:

[0025] The difference in the composite health status between two adjacent charge-discharge cycles throughout the entire life cycle of each lithium battery is calculated to form the corresponding composite health status increment.

[0026] Construct the negative log-likelihood function of all composite health state increments based on the probability density function of the composite health state increments;

[0027] The negative log-likelihood function after removing the constant is used as a constraint term for the nonlinear Wiener process;

[0028] The degradation capacity prediction error term is the mean square error of the degradation capacity calculated based on the predicted degradation capacity value and the actual degradation capacity value for each charge-discharge cycle throughout the entire life cycle of each lithium battery.

[0029] Preferably, the first neural network model is a convolutional neural network, and the second neural network model is a channel aggregation network.

[0030] Preferably, the variance of the model parameters of the nonlinear Wiener process model is calculated based on the trained degradation process model, as follows:

[0031] S61. The mean square error of the predicted remaining lifetime and the actual remaining lifetime output by the trained degradation process model is used as the objective function.

[0032] S62. Calculate the Hessian matrix of the objective function. Use the diagonal elements of the inverse of the Hessian matrix as the variance of the model parameters of the nonlinear Wiener process model. That is, take the top left and bottom right elements of the inverse of the Hessian matrix as the variance of the degradation rate and the variance of the set of kernel function parameters, respectively.

[0033] Preferably, particle filtering is used to update the state space model a preset number of times, as follows:

[0034] S71. Calculate the observation likelihood of each particle in the current charge-discharge cycle;

[0035] S72. Update the weights of the corresponding particles based on the observation likelihood of the current charge-discharge cycle and Bayes' theorem.

[0036] S73. Normalize the weights of each particle;

[0037] S74. Perform a weighted update on the composite health status of all particles to obtain a weighted composite health status.

[0038] S75. Based on the weighted composite health status accumulated from the 0th charge-discharge cycle to the current charge-discharge cycle, the remaining lifetime prediction value after the end of the current charge-discharge cycle is calculated using the full probability formula.

[0039] S76. Set the current charge / discharge cycle to the next charge / discharge cycle, and return to step S71 until the preset number of times has been updated.

[0040] Preferably, the lithium battery remaining lifetime prediction method based on a stochastic degradation state-space model further includes the following steps:

[0041] S8. Using the constraint that the target lithium battery remains unchanged throughout its entire life cycle, and combining the current number of charge-discharge cycles and the corresponding updated composite health state, calculate the final predicted value of the target lithium battery's remaining life. Update the final predicted value of the target lithium battery's remaining life to the predicted result of the target lithium battery's remaining life.

[0042] Preferably, the remaining lifespan prediction of the final target lithium battery is calculated by utilizing the constraint that the lithium battery remains unchanged throughout its entire lifespan, combined with the current number of charge-discharge cycles and the corresponding updated composite health state-of-the-art remaining lifespan prediction value, as follows:

[0043] S81. Add the historical remaining life prediction value of the current charge-discharge cycle to the corresponding number of charge-discharge cycles to obtain the full life cycle of the target lithium battery at the end of the corresponding charge-discharge cycle.

[0044] S82. Calculate the average value of the target lithium battery over its entire life cycle at the end of the historical charge-discharge cycles;

[0045] S83. Subtract the current charge / discharge cycle number from the average value of the target lithium battery's total lifespan at the end of the historical charge / discharge cycles to obtain the predicted remaining lifespan of the target lithium battery at the end of the current charge / discharge cycle.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] This invention collects lithium battery lifecycle data through multi-sensor collaboration, extracts spectral entropy, Hilbert spectral energy, and multi-stage time-domain statistical features (time and root mean square value at each stage) using Fourier transform and Hilbert transform, and constructs degradation feature vectors and auxiliary feature vectors. A convolutional neural network is used to extract composite health states, and a degradation process model is established based on a nonlinear Wiener process to quantify the stochastic evolution characteristics of composite health state increments. A joint optimization objective function is fused with remaining lifetime prediction error terms, nonlinear Wiener process constraint terms, and degradation capacity prediction error terms, simultaneously updating the convolutional neural network, channel aggregation network, and nonlinear Wiener process degradation rate parameters. This achieves efficient joint optimization of data-driven convolutional neural network composite health state extraction, observation equation construction based on the neural network model, and model-driven nonlinear Wiener process degradation modeling, completing a two-way collaboration between data-driven and mechanistic models. Unlike traditional lifespan prediction methods that rely on a single health indicator or a fixed degradation model, this invention starts with in-depth mining of multi-source sensor data throughout the entire lifespan of lithium batteries. It utilizes convolutional neural networks to adaptively extract and fuse key features characterizing battery degradation to estimate the composite health state. By constructing a nonlinear Wiener process as a state equation and combining it with a channel aggregation network to output the composite health state degradation observation equation, a stochastic degradation state space model is formed. Particle filtering is used to achieve fine-grained recursive updates of the composite health state, dynamically updating the state of the non-stationary stochastic process in the stochastic degradation state space model, accurately characterizing the randomness and evolution of battery capacity decay. Furthermore, it introduces constant physical constraints throughout the entire lifespan of lithium batteries and constructs a coarse-grained lifespan benchmark using historical prediction averages. This effectively solves the problems of insufficient interpretability of traditional data-driven methods and limited accuracy of single-model-driven methods, significantly improving the prediction robustness under cross-operating conditions. It realizes online prediction and state update of the remaining lifespan of lithium batteries, combining the advantages of high accuracy, strong robustness, and engineering practicality. Attached Figure Description

[0048] Figure 1 The flowchart shows the lithium battery remaining lifetime prediction method based on the stochastic degradation state-space model of the present invention.

[0049] Figure 2 This is a graph showing the change of the composite health status of the present invention with the number of charge-discharge cycles;

[0050] Figure 3 This is a comparison chart of the predicted and actual remaining lifespan of the lithium battery of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0052] It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application.

[0053] like Figures 1-3 As shown, a method for predicting the remaining lifetime of a lithium battery based on a stochastic degradation state-space model includes the following steps:

[0054] S1. Based on the collected full life cycle data of the lithium battery, extract the degradation feature vector and auxiliary feature vector corresponding to each charge-discharge cycle.

[0055] In one embodiment, the full life cycle data of the lithium battery includes the voltage sequence of the constant current charging stage, the current sequence of the constant voltage charging stage, and the voltage sequence of the constant current discharging stage in each charge-discharge cycle; the degradation feature vector is a vector composed of the normalized spectral entropy and Hilbert spectral energy of the voltage sequence of the constant current charging stage, the spectral entropy and Hilbert spectral energy of the current sequence of the constant voltage charging stage, the spectral entropy and Hilbert spectral energy of the voltage sequence of the constant current discharging stage, the time of the constant current charging stage, the time of the constant voltage charging stage, the time of the constant current discharging stage, the root mean square value of the voltage in the voltage sequence of the constant current charging stage, the root mean square value of the current in the current sequence of the constant voltage charging stage, and the root mean square value of the voltage in the voltage sequence of the constant current discharging stage; and the auxiliary feature vector is a vector composed of at least some elements of the degradation feature vector.

[0056] This embodiment uses multiple sensors (such as voltage sensors and current sensors) to collect the voltage and current of the lithium battery throughout its entire life cycle. Charging is divided into constant current mode (constant current charging stage) and constant voltage mode (constant voltage charging stage). First, constant current charging is performed until the voltage rises to a first set threshold, then the voltage is maintained constant and charging continues until the current drops to a second set threshold. After a first preset time, constant current discharging mode (constant current discharging stage) is adopted. As the battery capacity decreases, the voltage drops to a third set threshold and the charging stops. Then, the battery is left to rest for a second preset time, thus forming a complete charge and discharge cycle.

[0057] The full lifecycle data of a lithium battery includes the voltage sequence during the constant current charging phase, the current sequence during the constant voltage charging phase, and the voltage sequence during the constant current discharging phase in each charge-discharge cycle. Based on the full lifecycle data, Fast Fourier Transform and Hilbert Transform are performed to calculate the corresponding spectral entropy and Hilbert spectral energy, and other characteristics are obtained, as detailed below:

[0058] Taking a voltage sensor as an example, the data collected... The voltage sequence during the constant current discharge phase of the next charge-discharge cycle is subjected to a discrete Fourier transform, as shown in the following formula:

[0059]

[0060] in, Frequency Index The 1st after the lower discrete Fourier transform Voltage sequence during the constant current discharge phase of the next charge-discharge cycle. , For the first The voltage sequence of the constant current discharge phase in the next charge-discharge cycle is as follows: The voltage at each sampling point , The length of the sampling sequence, The imaginary unit;

[0061] According to frequency index The 1st after the lower discrete Fourier transform Voltage sequence during the constant current discharge phase of the second charge-discharge cycle Calculate the frequency index Next Power spectral density of the second charge-discharge cycle The formula is as follows:

[0062]

[0063] Frequency index Next Power spectral density of the second charge-discharge cycle Normalization is performed to obtain the frequency index. Next Normalized power spectrum of the second charge-discharge cycle The formula is as follows:

[0064]

[0065] in, This represents the floor function. .

[0066] According to frequency index Next Normalized power spectrum of the second charge-discharge cycle Calculate the first Spectral entropy of the voltage sequence during the constant current discharge phase of the second charge-discharge cycle The formula is as follows:

[0067]

[0068] Building a frequency index Hilbert filter under :

[0069]

[0070] According to frequency index The 1st after the lower discrete Fourier transform Voltage sequence during the constant current discharge phase of the second charge-discharge cycle and frequency index Hilbert filter under Calculate the frequency domain representation of the corresponding Hilbert change as the frequency index. Next Frequency domain sequence of the next charge-discharge cycle :

[0071]

[0072] According to frequency index Next Frequency domain sequence of the next charge-discharge cycle Calculate the first Hilbert spectral energy of the voltage sequence during the constant current discharge phase of the secondary charge-discharge cycle :

[0073]

[0074] in,

[0075]

[0076]

[0077] in, The first step after the inverse discrete Fourier transform The voltage sequence of the constant current discharge phase in the next charge-discharge cycle is as follows: The voltage at each sampling point For the first The voltage sequence of the constant current discharge phase in the next charge-discharge cycle is as follows: The analytical signal of each sampling point.

[0078] Based on the above, we can also obtain the spectral entropy and Hilbert spectral energy corresponding to the voltage sequence of the constant current charging stage and the current sequence of the constant voltage charging stage in a charge-discharge cycle.

[0079] The spectral entropy and Hilbert spectral energy of the voltage sequence during the constant current charging phase, the spectral entropy and Hilbert spectral energy of the current sequence during the constant voltage charging phase, and the spectral entropy and Hilbert spectral energy of the voltage sequence during the constant current discharging phase are included as part of the features. Other features include the time of the constant current charging phase, the time of the constant voltage charging phase, the time of the constant current discharging phase, the root mean square value of the voltage in the voltage sequence during the constant current charging phase, the root mean square value of the current in the current sequence during the constant voltage charging phase, and the root mean square value of the voltage in the voltage sequence during the constant current discharging phase. Then the... Other characteristics of the next charge-discharge cycle include the first The duration of the constant current charging phase in the next charge-discharge cycle , No. The duration of the constant voltage charging phase in the next charge-discharge cycle , No. The duration of the constant current discharge phase in the next charge-discharge cycle , No. The root mean square value of the voltage in the voltage sequence during the constant current charging phase of the next charge-discharge cycle. , No. The root mean square value of the current in the current sequence during the constant voltage charging phase of the second charge-discharge cycle. , No. The root mean square value of the voltage in the voltage sequence during the constant current discharge phase of the next charge-discharge cycle. ,in:

[0080]

[0081]

[0082]

[0083] in, Indicates the first The current sequence during the constant voltage charging phase of the next charge-discharge cycle is as follows: The current at each sampling point Indicates the first The voltage sequence of the constant current charging phase in the next charge-discharge cycle is as follows: The voltage at each sampling point.

[0084] Normalized features are obtained by normalizing the spectral entropy, Hilbert spectral energy, and other features. Specifically, for any one of the features, the spectral entropy, Hilbert spectral energy, and other features are normalized using the formula... Taking [the data] as an example, by normalizing it to the same range, we can reduce differences in data units and noise interference. The normalization formula is as follows:

[0085]

[0086] in, Indicates the first The first charge-discharge cycle The normalization results of each feature Indicates the first The first charge-discharge cycle One characteristic, express The maximum value, express The minimum value. Features are those in the set consisting of all spectral entropy, Hilbert spectral energy, and other features.

[0087] The first The vector representation of the normalized results of all characteristics of the next charge-discharge cycle is the first... Degradation eigenvectors of the second charge-discharge cycle Then the auxiliary feature vector Degenerative feature vector A vector consisting of some or all of its elements can be adjusted according to actual needs.

[0088] S2. Based on the degradation feature vector of each charge-discharge cycle, the first neural network model is used to output the composite health status of the corresponding charge-discharge cycle.

[0089] Specifically, according to the first Degradation eigenvectors of the second charge-discharge cycle The first neural network model is used to output the lithium battery's first... Composite health status after one charge-discharge cycle In this embodiment, a convolutional neural network is used as the first neural network model to obtain the first... Composite health status after one charge-discharge cycle The formula is as follows:

[0090]

[0091] in, Represents a convolutional neural network. Represents the weights of a convolutional neural network. Indicates the first The degradation feature vector of the next charge-discharge cycle.

[0092] S3. Based on the composite health status of charge-discharge cycles, a nonlinear Wiener process model is used to construct a degradation process model, and the remaining lifetime prediction value after the end of the corresponding charge-discharge cycle is obtained through the degradation process model.

[0093] In one embodiment, a degradation process model is constructed using a nonlinear Wiener process model based on the combined health state of charge-discharge cycles, as detailed below:

[0094] S31. Construct a nonlinear Wiener process model based on the combined health state of charge-discharge cycles;

[0095] S32. Obtain the probability density function of the composite health state increment based on the nonlinear Wiener process model. The composite health state increment is the difference between the composite health states of two adjacent charge-discharge cycles.

[0096] S33. Calculate the remaining lifetime value after each charge-discharge cycle based on the nonlinear Wiener process model to obtain the first-failure time, and calculate the probability density function of the first-failure time based on the probability density function of the composite health state increment.

[0097] S34. The expected value of the remaining lifetime after the end of the charge-discharge cycle is calculated as the degradation process model based on the potential value of the remaining lifetime after the end of the charge-discharge cycle and the probability density function of the first failure time.

[0098] In this embodiment, a nonlinear Wiener process model (NWPM) is used to construct a degradation process model of the composite health state throughout the entire life cycle of a lithium battery, as detailed below:

[0099] 3.1) According to the first A nonlinear Wiener process model is constructed based on the combined health state of the second charge-discharge cycle, and the formula is as follows:

[0100]

[0101] in, Indicates the first Composite health status of one charge-discharge cycle. Indicates the first The composite health state of the next charge-discharge cycle, i.e., the initial composite health state. Indicates the rate of degradation, characterizing individual differences. Indicates condition Kernel function under exist Integral over an interval, Indicates to Find the differential. , Represents a time variable. express and The set of parameters for the kernel function. Indicates the parameters of the first kernel function. This represents the parameters of the second kernel function. Indicates the diffusion coefficient. Indicates the first The standard Brownian motion of one charge-discharge cycle.

[0102] 3.2) Obtain the probability density function of the composite health state increment based on the nonlinear Wiener process model. The formula is as follows:

[0103]

[0104] in, This represents the increment of the composite health state for the corresponding charge-discharge cycle, which is the difference between the composite health states of two adjacent charge-discharge cycles, and is a random variable. This represents an exponential function with the natural constant e as its base.

[0105] 3.3) Based on the degradation path of the dynamic health score conforming to a nonlinear Wiener process, the first... Remaining lifetime value after the end of the next charge-discharge cycle The formula is as follows:

[0106]

[0107] in, Let be a random variable. The first-delivery failure time is considered to have been reached when the condition in parentheses on the right side of the equation is met. For the upper bound, Indicates the first The potential value of the remaining lifetime after the end of the next charge-discharge cycle. Indicates from the first The next charge / discharge cycle to the [number]th Composite health status of one charge-discharge cycle. Indicates the failure threshold. This indicates the time from the 0th charge / discharge cycle to the 1st charge / discharge cycle. The composite health status sequence of the next charge-discharge cycle. Indicates in Under the condition of first satisfaction Less than time The value of . It should be noted that the remaining lifespan of a lithium battery is essentially a random variable that follows a specific probability distribution, rather than a single, exact value. Therefore, It refers to the remaining lifetime independent variable in the sample space of the random variable, that is, from the current time. Initially, all possible time spans of candidate values ​​that may cause the composite health state to cross the failure threshold for the first time are called 'potential values' to emphasize that they are integral variables in the probability density function of the remaining lifetime, covering all theoretically possible failure moments. The predicted remaining lifetime value after the end of the corresponding charge-discharge cycle is obtained through subsequent expectation calculations.

[0108] The degradation path is a nonlinear Wiener process, and the probability density function of the first-reach failure time is... The formula is as follows:

[0109]

[0110] in, Indicates condition The kernel function under the prediction failure time (the first time) The degradation intensity (after one charge-discharge cycle).

[0111] 3.4) By selecting The mathematical expectation as Point estimates yield the degradation process model for the remaining lifespan of lithium batteries:

[0112]

[0113] in, Indicates the first Potential value of remaining lifetime after the end of the next charge-discharge cycle Find the differential. Indicates the first The predicted (expected) remaining lifetime after the end of the next charge-discharge cycle.

[0114] This step employs a nonlinear Wiener process model to construct a degradation process model for the dynamic health score of lithium batteries throughout their entire lifecycle. After obtaining the composite health state of the current charge-discharge cycle, it is substituted into the degradation process model to obtain the predicted remaining lifespan of the lithium battery. Compared to traditional time-domain features, which can only statically describe the degradation process of lithium batteries through statistical quantities such as mean and variance, the composite health state of lithium batteries, by introducing a nonlinear dynamic modeling method (first neural network model and nonlinear Wiener process model), overcomes the limitations of traditional methods that rely solely on manually preset, static, and shallow statistical indicators such as mean and variance to form a fixed statistical dimension. The latter's ability to represent complex nonlinear degradation is far lower than the deep dynamic features adaptively extracted by the neural network in this application, thus limiting its ability to represent complex degradation patterns.

[0115] S4. Obtain the true value of the degradation capacity for each charge-discharge cycle, and construct the composite health state degradation observation equation using a second neural network model based on the composite health state and auxiliary feature vector of the corresponding charge-discharge cycle.

[0116] In one embodiment, the first neural network model is a convolutional neural network, and the second neural network model is a channel aggregation network.

[0117] In this embodiment, a Channel Aggregation Network (CAN) is used as the second neural network model to construct the composite health state degradation observation equation, as detailed below:

[0118] 4.1) Obtain the first True value of degradation capacity after one charge-discharge cycle The formula is as follows:

[0119]

[0120] in, Indicates the first The current sequence during the constant voltage charging phase of the next charge-discharge cycle is as follows: The current at each sampling point.

[0121] 4.2) Using a channel aggregation network as the prediction model, the first... Auxiliary eigenvectors of the next charge-discharge cycle With the Composite health status after one charge-discharge cycle As input, the composite health state degradation observation equation is obtained:

[0122]

[0123] in, For the first Predicted degradation capacity after one charge-discharge cycle The weights of the channel aggregation network, This indicates a channel aggregation network.

[0124] This step uses a channel aggregation network to construct the observation equations during the composite health state degradation process. After obtaining the composite health state of the current charge-discharge cycle, the predicted degradation capacity of the lithium battery for the current charge-discharge cycle can be obtained by combining the auxiliary feature vectors.

[0125] It should be noted that the second neural network model is not limited to the Channel Aggregation Network. Any neural network architecture with multidimensional feature extraction and nonlinear mapping capabilities can be applied, such as multilayer perceptron (MLP), attention network, residual network (ResNet), etc. As long as it can achieve the purpose of mapping the composite health state and auxiliary feature vector to predict degradation capacity, it falls within the scope of application of this method.

[0126] S5. Calculate the total loss function to update the model parameters of the first neural network model, the degradation process model, and the second neural network model until training ends, and output the optimal parameter set of the corresponding model.

[0127] In one embodiment, the total loss function is obtained by weighted summation of the remaining lifetime prediction error term, the nonlinear Wiener process constraint term, and the degradation capacity prediction error term, and the gradient descent method is used to minimize the total loss function to update the model parameters of the first neural network model, the degradation process model, and the second neural network model.

[0128] In one embodiment, the remaining lifetime prediction error term is the mean square error of the remaining lifetime calculated based on the predicted remaining lifetime value and the actual remaining lifetime value after each charge-discharge cycle in the entire life cycle of each lithium battery.

[0129] The constraint terms for the nonlinear Wiener process are obtained as follows:

[0130] The difference in the composite health status between two adjacent charge-discharge cycles throughout the entire life cycle of each lithium battery is calculated to form the corresponding composite health status increment.

[0131] Construct the negative log-likelihood function of all composite health state increments based on the probability density function of the composite health state increments;

[0132] The negative log-likelihood function after removing the constant is used as a constraint term for the nonlinear Wiener process;

[0133] The degradation capacity prediction error term is the mean square error of the degradation capacity calculated based on the predicted degradation capacity value and the actual degradation capacity value for each charge-discharge cycle throughout the entire life cycle of each lithium battery.

[0134] In this embodiment, for a scenario involving collaborative training of multiple lithium batteries, a total loss function is calculated to update the model parameters of the convolutional neural network, the degradation process model, and the channel aggregation network. This is achieved by minimizing the total loss function. Specifically, this is achieved by iteratively adjusting the weights of the convolutional neural network, the degradation rate in the nonlinear Wiener process, the diffusion coefficient, and the weight parameters in the total loss function. and Factors such as failure thresholds and weights in the channel aggregation network are used to balance data-driven features and physical statistical properties, avoiding overfitting or underfitting caused by a single optimization. Specifically:

[0135] 5.1) Based on the predicted and actual remaining lifetime values ​​after each charge-discharge cycle throughout the entire life cycle of each lithium battery, the mean square error of the remaining lifetime is calculated as the remaining lifetime prediction error term. :

[0136]

[0137] in, This indicates the actual remaining lifespan of the lithium battery. Indicates the first The entire life cycle of a lithium-ion battery. This represents the total number of lithium batteries. The remaining lifespan prediction can be obtained by combining the composite health status output by the convolutional neural network with step S3.

[0138] 5.2) Calculate the constraint terms for the nonlinear Wiener process, as follows:

[0139] 5.2.1) Obtain the negative log-likelihood function of the total composite health state increment for each lithium battery, then the... The negative log-likelihood function of the total compound health state increment of a block lithium battery :

[0140]

[0141] in, This indicates the time from the 0th charge / discharge cycle to the 1st charge / discharge cycle. The complete composite health status increment sequence of the next charge-discharge cycle. Indicates the first The composite health status increment of each charge-discharge cycle.

[0142] 5.2.2) The summation of the negative log-likelihood functions of all composite health state increments of each lithium battery after removing the constant is used as a constraint term for the nonlinear Wiener process. The formula is as follows:

[0143]

[0144] 5.3) Calculate the mean square error of the degradation capacity. The formula is as follows:

[0145]

[0146] 5.4) Incorporate the remaining lifetime prediction error term Constraints of nonlinear Wiener processes Mean square error of degradation capacity The total loss function is obtained by performing a weighted summation. ,Right now:

[0147]

[0148] in, The weights of the constraint terms in the nonlinear Wiener process are... The weight of the mean square error of the degradation capacity.

[0149] 5.5) Minimize the total loss function using gradient descent. The system obtains a trained convolutional neural network, a degradation process model, and a channel aggregation network, and outputs the optimal parameter set. , This represents the weights of the trained convolutional neural network. This represents the degradation rate after training. This represents the diffusion coefficient after training. This represents the optimal failure threshold after training. This represents the set of parameters for the trained kernel function. This represents the weights of the trained channel aggregation network. For a trained degradation process model, the nonlinear Wiener process model within that process is also trained.

[0150] This embodiment uses gradient descent to minimize the total loss function and obtain the optimal parameter set. Finally, feedback between the composite health state and the nonlinear Wiener process model is achieved, thus forming a closed-loop feedback mechanism between the composite health state and the nonlinear Wiener process model.

[0151] It should be noted that in the scenario of collaborative training of multiple lithium batteries, the full life cycle data of all collected lithium batteries can be divided into training set and validation set according to a preset ratio. The training set is used for modeling the degradation process of composite health state, and the validation set is used for quantifying parameter uncertainty.

[0152] S6. Calculate the variance of the model parameters of the nonlinear Wiener process model based on the trained degradation process model, obtain the nonlinear Wiener process model that follows a normal distribution, and construct a stochastic degradation state-space model by combining the composite health state degradation observation equation of the trained second neural network model.

[0153] In one embodiment, the variance of the model parameters of the nonlinear Wiener process model is calculated based on the trained degradation process model, as follows:

[0154] S61. The mean square error of the predicted remaining lifetime and the actual remaining lifetime output by the trained degradation process model is used as the objective function.

[0155] S62. Calculate the Hessian matrix of the objective function. Use the diagonal elements of the inverse of the Hessian matrix as the variance of the model parameters of the nonlinear Wiener process model. That is, take the top left and bottom right elements of the inverse of the Hessian matrix as the variance of the degradation rate and the variance of the set of kernel function parameters, respectively.

[0156] To evaluate the reliability of the model parameters, an objective function is constructed based on the mean squared error between the predicted and actual remaining lifetime values ​​output by the trained degradation process model. The Hessian matrix of the objective function is then calculated. The inverse of the Hessian matrix is ​​used to approximate the posterior covariance matrix of the model parameters of the nonlinear Wiener process model. The diagonal elements are extracted as the parameter variances, thereby quantifying the uncertainty of the nonlinear Wiener process parameters (such as degradation rate and diffusion coefficient). The details are as follows:

[0157] 6.1) Construct the mean square error between the predicted remaining lifetime and the actual remaining lifetime output of the trained degradation process model as the objective function. :

[0158]

[0159] in, This represents the total number of samples. The output of the trained degradation process model represents the first... The remaining lifetime prediction value for a sample, which is the observation data of a lithium battery during one charge-discharge cycle. Indicates the first The true remaining lifetime value of each sample. Represents the size of the validation set, from The sum of observational data (sampling point data) extracted from different charge-discharge cycles throughout the entire lifecycle of a lithium-ion battery. Since each lithium-ion battery contributes multiple sampling point data points as samples throughout its entire lifecycle, therefore... Usually much larger .

[0160] 6.2) Based on the objective function, use the Hessian matrix to pair... and Estimate the variance of the Hessian matrix, and the inverse matrix of the Hessian matrix. , means as follows:

[0161]

[0162] in, This indicates finding the objective function. about The second-order partial derivative, This indicates finding the objective function. about The second-order partial derivatives of the objective function, in a physical sense, reflect the objective function's position on the parameters. and Curvature in direction, This indicates finding the objective function. about and The mixed partial derivatives, This indicates finding the objective function. about and The mixed partial derivatives reflect and The mutual influence between the two parameters.

[0163] Based on parameters All random fluctuations satisfy the assumption of a normal distribution. diagonal elements , Corresponding to each other in turn variance and variance .

[0164] This embodiment estimates the model parameters of a trained nonlinear Wiener process model, quantifies its randomness, and lays the foundation for state-space modeling that considers process uncertainties.

[0165] 6.3) According to variance and variance A nonlinear Wiener process model following a normal distribution is obtained. A stochastic degradation state-space model is constructed by combining the composite health state degradation observation equation of the nonlinear Wiener process model following a normal distribution and the trained channel aggregation network model, as detailed below:

[0166] 6.3.1) Use the trained nonlinear Wiener process model as the state equation:

[0167]

[0168] in, This represents Gaussian white noise. Indicates condition The kernel function below, Indicates the first The composite health state of the next charge-discharge cycle can be calculated based on the state equation.

[0169] 6.3.2) Combining the composite health state degradation observation equation of the trained channel aggregation network model, and considering the uncertainty of the process, the stochastic degradation state space model is constructed as follows:

[0170]

[0171] in, Indicates condition The kernel function below, Indicates the first Auxiliary feature vector of each charge-discharge cycle Indicates the first Predicted degradation capacity after one charge-discharge cycle Indicates the observed degradation rate Follow the first mean and first variance Normal distribution under the following conditions Represents the set of parameters of the observation kernel function Follow the second mean Second variance Normal distribution under the following conditions The mean of the degradation rate that follows a normal distribution is represented by the parameters obtained by gradient descent optimization. Let represent the variance of the degradation rate that follows a normal distribution, and be the element on the top-left diagonal of the inverse of the Hessian matrix. Let represent the mean of the set of kernel function parameters that satisfy a normal distribution, obtained by optimization using the gradient descent method. The variance of the set of kernel function parameters that satisfy a normal distribution is represented by the lower right diagonal element of the kernel function parameter set in the inverse of the Hessian matrix. In this embodiment, the trained nonlinear Wiener process model is used as the state equation, and it is combined with the trained channel aggregation network model to obtain the state space equation, laying the foundation for the next step of using particle filtering for state updates.

[0172] S7. Obtain the composite health state of the target lithium battery in the 0th charge-discharge cycle based on the trained first neural network model, and use particle filtering to update the composite health state a preset number of times based on the random degradation state space model. Calculate the remaining life prediction value corresponding to the updated composite health state as the remaining life prediction result of the lithium battery.

[0173] In one embodiment, particle filtering is used to update the state space model a preset number of times, as detailed below:

[0174] S71. Calculate the observation likelihood of each particle in the current charge-discharge cycle;

[0175] S72. Update the weights of the corresponding particles based on the observation likelihood of the current charge-discharge cycle and Bayes' theorem.

[0176] S73. Normalize the weights of each particle;

[0177] S74. Perform a weighted update on the composite health status of all particles to obtain a weighted composite health status.

[0178] S75. Based on the weighted composite health status accumulated from the 0th charge-discharge cycle to the current charge-discharge cycle, the remaining lifetime prediction value after the end of the current charge-discharge cycle is calculated using the full probability formula.

[0179] S76. Set the current charge / discharge cycle to the next charge / discharge cycle, and return to step S71 until the preset number of times has been updated.

[0180] Specifically, for the target lithium battery, the full lifecycle data of the lithium battery obtained by multiple sensors during the current charge-discharge cycle is used, and the following operations are performed:

[0181] 7.1) Calculate the degradation feature vector and auxiliary feature vector for the 0th charge-discharge cycle, as well as the auxiliary feature vector for subsequent charge-discharge cycles.

[0182] 7.2) The degradation feature vector of the 0th charge-discharge cycle is input into the trained convolutional neural network to obtain the composite health state of the target lithium battery in the 0th charge-discharge cycle, and particle filtering is used to update the data. The steps are as follows:

[0183] 7.2.1) Calculate the first... Observation likelihood of each particle :

[0184]

[0185] in, Indicates the first The true degradation capacity value of the next charge-discharge cycle, obtained in real-time by a current sensor, is used to correct the particle weights. , This indicates the total number of updates, which is the preset number of times. Indicates the first The predicted number of particles Composite health status of one charge-discharge cycle. Indicates the first The first charge-discharge cycle Predicted degradation capacity of individual particles, Indicates the first Auxiliary degradation characteristics of the next charge-discharge cycle This represents an exponential function with the natural constant e as its base. The composite health state at the 0th charge-discharge cycle is the initial composite health state.

[0186] 7.2.2) Update the particle weights based on the observation likelihood and Bayes' theorem, as shown in the following formula:

[0187]

[0188] in, Indicates the first The particle in the first Weighting of each charge-discharge cycle Indicates the first The particle in the first Weighting of each charge-discharge cycle.

[0189] 7.2.3) Subsequently, the particle weights are normalized:

[0190]

[0191] in, Indicates the normalized i-th The first charge-discharge cycle The weight of each particle, where E is the total number of particles.

[0192] 7.2.4) Subsequently, the composite health status of all particles is updated with weights to obtain the... Weighted composite health status of the next charge-discharge cycle :

[0193]

[0194] 7.2.5) Based on the first Weighted composite health status of the next charge-discharge cycle From the 0th charge-discharge cycle to the 1st The composite health status accumulated over the next charge-discharge cycle is calculated using the full probability formula. Predicted remaining lifetime after the end of the next charge-discharge cycle :

[0195]

[0196] in, The probability density function representing the remaining lifetime. The posterior probability density representing the observed degradation rate. This represents the posterior probability density of the set of parameters of the observed kernel function. Indicates to The differential, Indicates to The differential, This indicates the time from the 0th charge / discharge cycle to the 1st charge / discharge cycle. The updated composite health state sequence after each charge-discharge cycle yields the predicted remaining lifetime corresponding to the updated composite health state.

[0197] This step uses particle filtering to analyze the complex health state. Step-by-step update, implemented at the initial point (0th charge / discharge cycle). The fine-grained update strategy effectively avoids abrupt changes in state estimation and significantly reduces the interference of single measurement errors on the overall prediction, thereby improving the stability and reliability of the state trajectory. At the same time, the particle filtering mechanism endows the local dynamic evolution capability of the random degenerate state space model with the ability to evolve.

[0198] In one embodiment, the lithium battery remaining lifetime prediction method based on a stochastic degradation state-space model further includes the following steps:

[0199] S8. Using the constraint that the target lithium battery remains unchanged throughout its entire life cycle, and combining the current number of charge-discharge cycles and the corresponding updated composite health state, calculate the final predicted value of the target lithium battery's remaining life. Update the final predicted value of the target lithium battery's remaining life to the predicted result of the target lithium battery's remaining life.

[0200] In one embodiment, the remaining lifetime prediction value of the final target lithium battery is calculated by utilizing the constraint that the lithium battery remains unchanged throughout its entire life cycle, combined with the current number of charge-discharge cycles and the remaining lifetime prediction value corresponding to the updated composite health state, as follows:

[0201] S81. Add the historical remaining life prediction value of the current charge-discharge cycle to the corresponding number of charge-discharge cycles to obtain the full life cycle of the target lithium battery at the end of the corresponding charge-discharge cycle.

[0202] S82. Calculate the average value of the target lithium battery over its entire life cycle at the end of the historical charge-discharge cycles;

[0203] S83. Subtract the current charge / discharge cycle number from the average value of the target lithium battery's total lifespan at the end of the historical charge / discharge cycles to obtain the predicted remaining lifespan of the target lithium battery at the end of the current charge / discharge cycle.

[0204] Among them, taking into account the constraint that the target lithium battery remains unchanged throughout its entire life cycle, the remaining life prediction value is obtained based on the principle of consistency of the total lifespan of the historical prediction, according to the current number of charge-discharge cycles and the historical prediction results of the target lithium battery. The historical prediction results of the target lithium battery are the remaining life prediction value of the lithium battery before the current number of charge-discharge cycles.

[0205] 8.1) Based on the historical target lithium battery remaining life prediction sequence of the current charge-discharge cycle count. Calculate the entire life cycle of the target lithium battery, then the first... The full life cycle of the target lithium battery at the end of the second charge-discharge cycle The formula is as follows:

[0206]

[0207] in, For the first Predicted remaining lifetime after the end of the next charge-discharge cycle. To improve the rationality of the constraints and their adaptability to data distribution characteristics, a z-score standardization method is introduced to optimize the constraint logic. The mean of the historical predicted total lifetime is then calculated. variance of historical predicted total lifespan :

[0208]

[0209]

[0210] Based on the historical average predicted total lifespan variance of historical predicted total lifespan Based on the z-score standardization method, the historical predicted total lifespan is converted into a standardized score. The entire lifespan of the target lithium battery within a preset distribution range is selected as a constraint value. After setting a confidence interval (e.g., 95%) and eliminating extreme outliers, the mean is calculated to obtain the expected value of the standardized distribution fit of the historical predicted total lifespan. The expected value is fitted based on the standardized distribution of historical predicted total lifetime. As the total lifetime benchmark for satisfying consistency constraints, the first... Predicted remaining lifetime of the target lithium battery at the end of the next charge-discharge cycle:

[0211]

[0212] in, For the first The predicted remaining life of the target lithium battery at the end of the next charge-discharge cycle.

[0213] This embodiment utilizes a stochastic degradation state-space model constructed based on a nonlinear Wiener process and a channel aggregation network to achieve online real-time prediction of the remaining lifespan of lithium batteries. It should be noted that the model training process, corresponding to steps S1-S5, typically requires multiple iterations until the training termination condition is met. The prediction application process, corresponding to steps S6-S7, is executed the corresponding number of times as needed. These two processes can be executed independently or synchronously in an online learning scenario to achieve continuous optimization of model parameters.

[0214] To facilitate understanding, the following specific experiments will be used to verify and illustrate this:

[0215] This embodiment is verified using the Xi'an Jiaotong University lithium battery public dataset Batch-1, as detailed below:

[0216] The initial calibration cycle is set to constant current mode (CC, 1A) during the initial calibration charging phase (t=0). When the cutoff voltage reaches 4.2V, it switches to constant voltage mode (CV, 4.2V). The charging continues until the current decays to 0.04A, followed by a 5-minute rest period. Then, the discharge phase uses constant current mode (CC, 0.4A) until the voltage decays to 2.5V, followed by another 5-minute rest period. This is the initial calibration cycle. The accelerated degradation cycle phase (t≥1 until failure) uses constant current mode (CC, 4A). When the cutoff voltage reaches 4.2V, it switches to constant voltage mode (CV, 4.2V). The charging continues until the current decays to 0.1A, followed by a 5-minute rest period. Then, the discharge phase uses constant current mode (CC, 2A) until the voltage decays to 2.5V, followed by another 5-minute rest period. This accelerated degradation cycle phase continues until the battery ages. Nine auxiliary features for charge-discharge cycles were extracted from the full life-cycle data of lithium batteries. These features include the spectral entropy of the constant current charging voltage sequence, the spectral entropy of the constant voltage charging voltage sequence, the spectral entropy of the constant current discharging voltage sequence, the duration of the constant current charging phase, the duration of the constant voltage charging phase, the duration of the constant current discharging phase, the root mean square (RMS) value of the voltage in the voltage sequence during the constant current charging phase, the RMS value of the current in the current sequence during the constant voltage charging phase, and the RMS value of the voltage in the voltage sequence during the constant current discharging phase. These nine features, along with the following three features, form a degradation feature vector: the Hilbert spectral energy of the constant current charging voltage sequence, the constant voltage charging voltage sequence, and the constant voltage discharging voltage sequence. The Hilbert spectral energies of the electric voltage sequence and the constant current discharge voltage sequence were used. Additionally, the actual values ​​of the degradation capacity from each charge-discharge cycle were extracted as observations. 70% of the 2C_battery-1, 2C_battery-2, and 2C_battery-4 datasets were used as the training set to model the composite health state degradation process. During model training, the Adam optimizer was used for iterative parameter updates to adaptively adjust the learning rate. The initial learning rate was set to 0.0005, the training epochs were set to 400, and the batch size was set to 128. After training, the optimal model parameters of the convolutional neural network, the degradation process model, and the channel aggregation network were obtained. 30% of the 2C_battery-1, 2C_battery-2, and 2C_battery-4 datasets were used as a validation set for quantifying parameter uncertainty. Based on this, a stochastic degradation state-space model is obtained, and particle filtering is used to update the composite health state in 6 steps on the 2C_battery-5 dataset. Combined with the principle of consistency between historical predicted total lifetime, online prediction of the remaining lifespan of lithium batteries is achieved. Experimental results show that the proposed method has excellent prediction performance on the test set: the mean absolute error (MAE) is only 3.4641, the root mean square error (RMSE) is 3.8445, and the coefficient of determination (R²) is as high as 0.9985. These indicators demonstrate that the model of this invention has extremely high prediction accuracy and fitting ability, and can accurately track the degradation trajectory of the battery. Figure 2 This is a graph showing the change in composite health status over the number of charge-discharge cycles. Figure 3 This is a comparison chart of the predicted and actual remaining lifespan of lithium batteries.

[0217] This invention collects lithium battery lifecycle data through multi-sensor collaboration, extracts spectral entropy, Hilbert spectral energy, and multi-stage time-domain statistical features (time and root mean square value at each stage) using Fourier transform and Hilbert transform, and constructs degradation feature vectors and auxiliary feature vectors. A convolutional neural network is used to extract composite health states, and a degradation process model is established based on a nonlinear Wiener process to quantify the stochastic evolution characteristics of composite health state increments. An innovative joint optimization objective function is fused with remaining lifetime prediction error terms, nonlinear Wiener process constraint terms, and degradation capacity prediction error terms, simultaneously updating the convolutional neural network, channel aggregation network, and nonlinear Wiener process degradation rate parameters, achieving bidirectional synergy between data-driven and mechanistic models. By constructing the nonlinear Wiener process as a state equation and combining it with the observation equation output by the channel aggregation network to form a stochastic degradation state-space model, particle filtering is used to achieve fine-grained recursive updates of composite health states. Furthermore, a constant physical constraint throughout the lithium battery lifecycle is introduced, and a coarse-grained lifetime benchmark is constructed using historical prediction averages, effectively solving the problems of insufficient interpretability in traditional data-driven methods and limited accuracy in single-model-driven methods, significantly improving prediction robustness under cross-operating conditions. Specifically, fine-grained particle filtering handles dynamic tracking, processing local fluctuations and uncertainties. Coarse-grained filtering handles global calibration, using historical average lifetime constraints to prevent predictions from deviating from physical realities, achieving digital-analog linkage, and ultimately outputting a remaining lifetime prediction value that possesses both random tracking capability and global robustness. This achieves a high-precision, highly interpretable online prediction scheme for the remaining lifetime of lithium batteries.

[0218] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0219] The embodiments described above are merely specific and detailed examples of the embodiments described in this application, and should not be construed as limiting the scope of the application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A method for predicting the remaining lifetime of a lithium battery based on a stochastic degradation state-space model, characterized in that: Includes the following steps: S1. Based on the collected full life cycle data of the lithium battery, extract the degradation feature vector and auxiliary feature vector corresponding to each charge-discharge cycle; S2. Based on the degradation feature vector of each charge-discharge cycle, the first neural network model is used to output the composite health status of the corresponding charge-discharge cycle. S3. Based on the composite health status of charge-discharge cycles, a nonlinear Wiener process model is used to construct a degradation process model, and the remaining lifetime prediction value after the end of the corresponding charge-discharge cycle is obtained through the degradation process model. S4. Obtain the true value of the degradation capacity for each charge-discharge cycle, and construct the composite health state degradation observation equation using the second neural network model based on the composite health state and auxiliary feature vector of the corresponding charge-discharge cycle, so as to obtain the predicted value of the degradation capacity for each charge-discharge cycle. S5. Calculate the total loss function to update the model parameters of the first neural network model, the degradation process model and the second neural network model until the training ends, and output the optimal parameter set of the corresponding model. S6. Calculate the variance of the model parameters of the nonlinear Wiener process model based on the trained degradation process model to obtain a nonlinear Wiener process model that follows a normal distribution. Combine this with the composite health state degradation observation equation of the trained second neural network model to construct a stochastic degradation state-space model. The stochastic degradation state-space model is constructed as follows: in, Indicates the first Composite health status of one charge-discharge cycle. Indicates the first Composite health status of one charge-discharge cycle. Indicates to Find the differential. Represents a time variable. This represents the diffusion coefficient after training. This represents Gaussian white noise. Indicates a channel aggregation network. This represents the weights of the trained channel aggregation network. Indicates condition The kernel function below, Indicates the first Auxiliary feature vector of each charge-discharge cycle Indicates the first Predicted degradation capacity after one charge-discharge cycle Indicates the observed degradation rate Follow the first mean and first variance Normal distribution under the following conditions Represents the set of parameters of the observation kernel function. Follow the second mean Second variance Normal distribution under the following conditions This represents the mean of the degradation rates that follow a normal distribution. The variance represents the degradation rate that follows a normal distribution. Let represent the mean of the set of parameters of the kernel function that satisfy a normal distribution. This represents the variance of the set of parameters of a kernel function that satisfies a normal distribution. S7. Obtain the composite health state of the target lithium battery in the 0th charge-discharge cycle based on the trained first neural network model, and use particle filtering to update the composite health state a preset number of times based on the random degradation state space model. Calculate the remaining lifetime prediction value corresponding to the updated composite health state as the remaining lifetime prediction result of the target lithium battery.

2. The lithium battery remaining lifetime prediction method based on a stochastic degradation state-space model as described in claim 1, characterized in that: The full lifecycle data of the lithium battery includes the voltage sequence of the constant current charging stage, the current sequence of the constant voltage charging stage, and the voltage sequence of the constant current discharging stage in each charge-discharge cycle. The degradation feature vector is a vector composed of the normalized spectral entropy and Hilbert spectral energy of the voltage sequence of the constant current charging stage, the spectral entropy and Hilbert spectral energy of the current sequence of the constant voltage charging stage, the spectral entropy and Hilbert spectral energy of the voltage sequence of the constant current discharging stage, the time of the constant current charging stage, the time of the constant voltage charging stage, the time of the constant current discharging stage, the root mean square value of the voltage in the voltage sequence of the constant current charging stage, the root mean square value of the current in the current sequence of the constant voltage charging stage, and the root mean square value of the voltage in the voltage sequence of the constant current discharging stage. The auxiliary feature vector is a vector composed of at least some elements of the degradation feature vector.

3. The lithium battery remaining lifetime prediction method based on a stochastic degradation state-space model as described in claim 1, characterized in that: The degradation process model is constructed using a nonlinear Wiener process model based on the combined health state of charge-discharge cycles, as detailed below: S31. Construct a nonlinear Wiener process model based on the combined health state of charge-discharge cycles; S32. Obtain the probability density function of the composite health state increment according to the nonlinear Wiener process model. The composite health state increment is the difference between the composite health states of two adjacent charge-discharge cycles. S33. Calculate the remaining lifetime value after each charge-discharge cycle based on the nonlinear Wiener process model to obtain the first-failure time, and calculate the probability density function of the first-failure time based on the probability density function of the composite health state increment. S34. The expected value of the remaining lifetime after the end of the charge-discharge cycle is calculated as the degradation process model based on the potential value of the remaining lifetime after the end of the charge-discharge cycle and the probability density function of the first failure time.

4. The lithium battery remaining lifetime prediction method based on a stochastic degradation state-space model as described in claim 1, characterized in that: The total loss function is obtained by weighted summation of the remaining lifetime prediction error term, the nonlinear Wiener process constraint term, and the degradation capacity prediction error term. The gradient descent method is used to minimize the total loss function to update the model parameters of the first neural network model, the degradation process model, and the second neural network model.

5. The lithium battery remaining life prediction method based on a stochastic degradation state-space model as described in claim 4, characterized in that: The remaining life prediction error term is the mean square error of the remaining life calculated based on the predicted and actual remaining life values ​​after each charge-discharge cycle in the entire life cycle of each lithium battery. The constraint terms for the nonlinear Wiener process are obtained as follows: The difference in the composite health status between two adjacent charge-discharge cycles throughout the entire life cycle of each lithium battery is calculated to form the corresponding composite health status increment. Construct the negative log-likelihood function of all composite health state increments based on the probability density function of the composite health state increments; The negative log-likelihood function after removing the constant is used as a constraint term for the nonlinear Wiener process; The degradation capacity prediction error term is the mean square error of the degradation capacity calculated based on the predicted degradation capacity value and the actual degradation capacity value for each charge-discharge cycle throughout the entire life cycle of each lithium battery.

6. The lithium battery remaining lifetime prediction method based on a stochastic degradation state-space model as described in claim 1, characterized in that: The first neural network model is a convolutional neural network, and the second neural network model is a channel aggregation network. 7.The lithium battery remaining life prediction method based on a random degradation state space model of claim 1, wherein: The variance of the model parameters of the nonlinear Wiener process model is calculated based on the trained degradation process model, as follows: S61. The mean square error of the predicted remaining lifetime and the actual remaining lifetime output by the trained degradation process model is used as the objective function. S62. Calculate the Hessian matrix of the objective function. Use the diagonal elements of the inverse of the Hessian matrix as the variance of the model parameters of the nonlinear Wiener process model. That is, take the top left and bottom right elements of the inverse of the Hessian matrix as the variance of the degradation rate and the variance of the set of kernel function parameters, respectively. 8.The lithium battery remaining life prediction method based on a random degradation state space model of claim 1, wherein: The particle filter is used to update the state space model a preset number of times, as detailed below: S71. Calculate the observation likelihood of each particle in the current charge-discharge cycle; S72. Update the weights of the corresponding particles based on the observation likelihood of the current charge-discharge cycle and Bayes' theorem. S73. Normalize the weights of each particle; S74. Perform a weighted update on the composite health status of all particles to obtain a weighted composite health status. S75. Based on the weighted composite health status accumulated from the 0th charge-discharge cycle to the current charge-discharge cycle, the remaining lifetime prediction value after the end of the current charge-discharge cycle is calculated using the full probability formula. S76. Set the current charge / discharge cycle to the next charge / discharge cycle, and return to step S71 until the preset number of times has been updated. 9.The lithium battery remaining life prediction method based on a random degradation state space model of claim 1, wherein: The lithium battery remaining lifetime prediction method based on the stochastic degradation state-space model further includes the following steps: S8. Using the constraint that the target lithium battery remains unchanged throughout its entire life cycle, and combining the current number of charge-discharge cycles and the corresponding updated composite health state, calculate the final predicted value of the target lithium battery's remaining life. Update the final predicted value of the target lithium battery's remaining life to the predicted result of the target lithium battery's remaining life. 10.The lithium battery remaining life prediction method based on a random degradation state space model of claim 9, wherein: The final predicted remaining lifespan of the target lithium battery is calculated by utilizing the constraint that the target lithium battery remains unchanged throughout its entire lifespan, combined with the current number of charge-discharge cycles and the corresponding updated composite health state-of-the-art predicted remaining lifespan value, as detailed below: S81. Add the historical remaining life prediction value of the current charge-discharge cycle to the corresponding number of charge-discharge cycles to obtain the full life cycle of the target lithium battery at the end of the corresponding charge-discharge cycle. S82. Calculate the average value of the target lithium battery over its entire life cycle at the end of the historical charge-discharge cycles; S83. Subtract the current charge / discharge cycle number from the average value of the target lithium battery's total lifespan at the end of the historical charge / discharge cycles to obtain the predicted remaining lifespan of the target lithium battery at the end of the current charge / discharge cycle.

Images