Lithium battery residual life real-time prediction method based on dynamic uncertainty modeling

Abstract

The invention belongs to the field of lithium battery residual life prediction, and discloses a lithium battery residual life real-time prediction method based on dynamic uncertainty modeling, and the method comprises the steps: cooperatively collecting the full life cycle data of a lithium battery through multiple sensors, constructing a dynamic health score in combination with a convolutional neural network, and effectively capturing the non-stationary and non-linear characteristics of the data. The uncertainty of health scoring is quantified by further adopting a non-stationary random process, and related parameters are synchronously adjusted through joint optimization of an objective function to enhance the adaptability of the model to a complex degradation mode. And dynamically updating non-stationary random process parameters based on a Bayesian reasoning framework, and combining conjugate prior distribution of historical data and real-time observation values to realize parameter adaptive adjustment and dynamic prediction of the residual life of the lithium battery. According to the method, the defect that a traditional data driving method lacks physical interpretability is overcome, and the problem that precision is insufficient due to the fact that a battery degradation mechanism is complex in a single model driving method is solved.

Description

Technical Field

[0001] The present invention belongs to the field of predicting the remaining life of lithium batteries, and particularly relates to a real-time prediction method for the remaining life of lithium batteries based on dynamic uncertainty modeling. Background Art

[0002] Predicting the remaining life of lithium batteries is a key technology to ensure the safe operation of the power systems of new energy vehicles and grid-level energy storage devices. The core lies in establishing an accurate mathematical model of multi-dimensional degradation paths such as the phase change kinetics of electrode active materials, the oxidation and decomposition mechanism of electrolytes, and the side reactions at the solid-liquid interface. By quantitatively analyzing the internal correlations of these degradation mechanisms, dynamic early warnings for key failure nodes such as sudden capacity drop and internal resistance mutation can be achieved, providing a scientific basis for system health management. In a dynamic working condition environment, the electrochemical polarization and concentration polarization effects inside the battery form a complex non-linear coupling relationship with the external thermal-mechanical-electrical multi-physical fields. This cross-scale interaction not only accelerates micro-scale degradation processes such as SEI (Solid Electrolyte Interphase) film reconstruction and lithium dendrite growth, but also makes the time-varying degradation behavior at the material level exhibit significant non-stationary characteristics. In this complex background, early aging signals are often submerged in strong noise interference, and traditional linear prediction methods based on threshold judgment are difficult to effectively identify the hidden degradation trends.

[0003] The existing remaining useful life prediction method system for lithium batteries has gradually evolved into three major paradigms: physical mechanism models, data-driven models, and hybrid intelligent architectures, in response to the multi-dimensional requirements of mechanism interpretability, operating condition adaptability, and engineering practicality. Among them, data-driven methods extract potential aging laws from a large amount of historical data through machine learning and deep learning algorithms, and have become the mainstream research direction in this field. Its technological evolution mainly unfolds along three dimensions: deep learning architectures, kernel method systems, and hybrid modeling strategies. In the field of deep learning, recurrent neural networks such as LSTM (Long Short-Term Memory) networks and temporal convolutional networks are used to capture the temporal dependence characteristics of capacity decay, and an end-to-end multi-step rolling prediction framework is constructed. The kernel method system relies on the kernel techniques of support vector regression and the probability framework of Gaussian process regression to represent the degradation trajectory through non-linear mapping and quantitatively evaluate the prediction uncertainty. The hybrid modeling strategy constructs a multi-scale prediction framework by integrating relevance vector machines and three-parameter degradation models, or combines ARIMA (Auto Regressive Integrated Moving Average) time series analysis and hidden Markov models to handle non-stationary problems such as capacity regeneration. Compared with physical models that require precise description of electrochemical reactions, data-driven methods effectively avoid the dilemma of complex mechanism modeling. However, their prediction performance highly depends on the completeness of feature engineering, the scale and quality of training data, and there are still inherent limitations in model interpretability and uncertainty quantification.

[0004] In addition, the progress of sensor and computing technologies has provided a large amount of labeled degradation data for data-driven methods, promoting the application and development of deep neural networks. Although such methods can avoid complex electrochemical modeling through self-learning and accurately predict capacity inflection points in specific scenarios, their "black box" nature results in opaque decision-making bases. In practical applications, differences in different battery models and operating conditions make it difficult to transfer model parameters across scenarios, and the full-life cycle labeled data requires months of experimental collection, and the high cost seriously restricts industrial large-scale applications. Summary of the Invention

[0005] The purpose of the present invention is to provide a real-time prediction method for the remaining useful life of lithium batteries based on dynamic uncertainty modeling, which integrates the advantages of data-driven methods in multi-source information processing and the ability of probability models in uncertainty quantification, not only making up for the defect of the lack of physical interpretability of traditional data-driven methods, but also overcoming the problem of insufficient accuracy caused by the complex battery degradation mechanism in single model-driven methods.

[0006] To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0007] A real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling. The real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling includes:

[0008] Obtain the full-life cycle data of the lithium battery collected by multiple sensors, and combine with the discrete Fourier transform to extract the degradation feature vectors corresponding to each charge-discharge cycle in the full-life cycle of the lithium battery;

[0009] Through a convolutional neural network, output the dynamic health score of the lithium battery according to the degradation feature vectors;

[0010] Based on the dynamic health score, construct a degradation process model for the remaining life of the lithium battery using a linear Wiener process model, and calculate the predicted value of the remaining life of the lithium battery based on the obtained degradation process model;

[0011] Calculate the loss function including the remaining life prediction error term and the maximum likelihood estimation constraint term of the linear Wiener process, and update the parameters of the convolutional neural network and the degradation process model according to the loss function until the training ends;

[0012] Obtain the real-time dynamic health score of the target lithium battery through the updated convolutional neural network, and use Bayesian theory to update the drift coefficient of the degradation process model in real time, and output the real-time remaining life prediction value of the target lithium battery using the latest degradation process model.

[0013] The following also provides several optional methods, which are not additional limitations to the above overall solution, but only further supplements or optimizations. Without technical or logical contradictions, each optional method can be combined with the above overall solution separately, or multiple optional methods can be combined with each other.

[0014] Preferably, the combination with the discrete Fourier transform to extract the degradation feature vectors corresponding to each charge-discharge cycle in the full-life cycle of the lithium battery includes:

[0015] For each charge-discharge cycle, perform a discrete Fourier transform on the charging current sequence to calculate the corresponding total spectral energy; perform a discrete Fourier transform on the discharge voltage sequence to calculate the corresponding total spectral energy;

[0016] And calculate the average resistance of the charge-discharge cycle, the time of the constant-voltage charging stage, the time of the constant-current discharge stage, the average slope of the current in the constant-voltage charging stage decaying from A to B, and the average slope of the voltage in the constant-current discharge stage decaying from C to D in the current charge-discharge cycle, where A is greater than B and C is greater than D;

[0017] Normalize the total spectral energy corresponding to the charging current sequence, the total spectral energy corresponding to the discharging voltage sequence, the average resistance of the charge-discharge cycle, the time of the constant voltage charging stage, the time of the constant current discharging stage, the average slope of the current in the constant voltage charging stage decaying from A to B, and the average slope of the voltage in the constant current discharging stage decaying from C to D respectively, and combine all the normalized features as the degradation feature vector.

[0018] Preferably, based on the dynamic health score, a degradation process model for the remaining life of the lithium battery is constructed using a linear Wiener process model, including:

[0019] Use a linear Wiener process to describe the degradation path of the dynamic health score in the whole life cycle of the lithium battery, where the drift coefficient follows a normal distribution and the increment of the dynamic health score follows a normal distribution;

[0020] Define the first passage failure time according to the degradation path of the dynamic health score that conforms to the linear Wiener process;

[0021] Construct the probability density function of the first passage failure time;

[0022] Take the integral process of the probability density function of the first passage failure time as the remaining life prediction function to complete the construction of the degradation process model of the remaining life of the lithium battery.

[0023] Preferably, the remaining life prediction error term includes:

[0024] Calculate the mean square error based on the predicted remaining life value and the true remaining life value corresponding to each charge-discharge cycle in the whole life cycle of the lithium battery, and use the mean square error as the remaining life prediction error term.

[0025] Preferably, the maximum likelihood estimation constraint term of the linear Wiener process includes:

[0026] Calculate the increment of the dynamic health score between two adjacent charge-discharge cycles in the whole life cycle of the lithium battery;

[0027] Construct the logarithmic joint likelihood function of all dynamic health score increments;

[0028] Take the logarithmic joint likelihood function after removing the constant as the maximum likelihood estimation constraint term of the linear Wiener process.

[0029] Preferably, the construction of the logarithmic joint likelihood function of all dynamic health score increments includes:

[0030]

[0031] Among them, represents the logarithmic joint likelihood function, is the total number of charge-discharge cycles in the entire life cycle of the lithium battery, is the increment of the dynamic health score corresponding to the th charge-discharge cycle, is the mean value of the drift coefficient, is the time increment between two adjacent charge-discharge cycles,

[0032] Preferably, the loss function is the weighted sum of the remaining life prediction error term and the maximum likelihood estimation constraint term of the linear Wiener process;

[0033] Use the gradient descent method to minimize the loss function to update the parameters of the convolutional neural network and the degradation process model, and after the training is completed, output the optimal parameter set, the optimal parameter set includes the optimal weight parameters of the convolutional neural network, the optimal drift coefficient of the degradation process model, the optimal diffusion coefficient of the degradation process model, and the optimal failure threshold of the degradation process model, the drift coefficient follows a normal distribution, that is, obtain the mean value of the optimal drift coefficient and the standard deviation of the optimal drift coefficient.

[0034] Preferably, obtaining the real-time dynamic health score of the target lithium battery through the updated convolutional neural network includes:

[0035] Obtain multi-sensor data of the target lithium battery during the current charge-discharge cycle;

[0036] Combined with the discrete Fourier transform, extract the degradation feature vector of the multi-sensor data;

[0037] Through the updated convolutional neural network, output the real-time dynamic health score of the target lithium battery according to the feature vector.

[0038] Preferably, the use of Bayesian theory to update the drift coefficient of the degradation process model in real time includes:

[0039] Use Bayesian theory to update the drift coefficient of the degradation process model in real time. The updated drift coefficient follows the posterior distribution. Calculate the posterior distribution in combination with the optimal parameters of the degradation process model obtained after the training is completed. The calculation formula is as follows:

[0040]

[0041] Among them, is the standard deviation of the drift coefficient after Bayesian theory update, is the mean value of the drift coefficient after Bayesian theory update, is the optimal diffusion coefficient of the degradation process model, is the standard deviation of the optimal drift coefficient of the degradation process model, is the current number of charge-discharge cycles of the target lithium battery, is the time increment between two adjacent charge-discharge cycles, is the mean of the optimal drift coefficient of the degradation process model, For the target lithium battery at the th charge-discharge cycle corresponds to the dynamic health score increment.

[0042] A real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling provided by the present invention realizes the efficient joint optimization of data-driven deep health score extraction of a convolutional neural network and model-driven linear Wiener process degradation modeling by designing a global objective function for multi-dimensional feature fusion and uncertainty collaborative optimization. Different from the traditional life prediction methods that rely on a single sensing signal or a fixed parameter model, the method of the present invention starts from the in-depth mining of full-life cycle data, adaptively fuses battery degradation features by using a convolutional neural network, and combines the Bayesian framework to dynamically update the parameters of a non-stationary random process, accurately depicting the randomness and time-varying characteristics of battery performance degradation. The method of the present invention can realize real-time dynamic prediction of the remaining life, and has the advantages of high reliability, strong interpretability and low deployment cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of a real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling of the present invention;

[0044] Figure 2 is a parameter change diagram of the linear Wiener process of the present invention;

[0045] Figure 3 is a dynamic health score and remaining life degradation trajectory diagram of a convolutional neural network and a degradation process model trained based on the CS2_35 lithium battery dataset of the present invention;

[0046] Figure 4 is a dynamic health score and remaining life degradation trajectory diagram of a convolutional neural network and a degradation process model tested based on the CS2_35 lithium battery dataset of the present invention;

[0047] Figure 5 is an application effect diagram of a real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling of the present invention on the CS2_36 lithium battery dataset. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this invention belongs. The terms used in the description of the present invention herein are for the purpose of describing specific embodiments only and are not intended to limit the present invention.

[0050] As Figure 1 shown, this embodiment proposes a real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling, including the following steps:

[0051] Step 1: Obtain the full-life cycle data of the lithium battery collected by multiple sensors, and combine with the discrete Fourier transform to extract the degradation feature vectors corresponding to each charge-discharge cycle in the full-life cycle of the lithium battery.

[0052] Step 1.1: Use multiple sensors to collect the full-life cycle data of the lithium battery, output the current data in the constant voltage charging stage and the voltage data in the constant current discharging stage in each charge-discharge cycle, and perform discrete Fourier transform on them respectively to calculate the corresponding total spectrum energy and other features.

[0053] In this embodiment, a voltage sensor, a current sensor, and an internal resistance sensor are used to collect the voltage, current, and internal resistance data of the lithium battery throughout its life cycle. The constant voltage mode is adopted during the charging process. As the battery capacity increases, the charging current decays non-linearly until it reaches the current threshold and the charging terminates; the constant current mode is adopted during the discharging process. As the battery capacity decreases, the terminal voltage decreases monotonically until it reaches the voltage threshold and the discharging terminates, thus constituting a complete charge-discharge cycle.

[0054] Taking the current sensor as an example, define the charging current sequence of a charge-discharge cycle process collected as , where represents the sampling time, represents the number of charge-discharge cycles. In this embodiment, the discrete Fourier transform is performed on the charging current sequence of a collected charge-discharge cycle process, and the total spectrum energy of this charging process is calculated.

[0055] After the discrete Fourier transform, the output frequency sequence is expressed as:

[0056]

[0057] where represents the discrete frequency index, represents the sequence length, that is, the total sampling time, is the imaginary part.

[0058] Perform the total spectrum energy calculation on the output frequency sequence :

[0059]

[0060] Among them, represents the total spectral energy of the charging current sequence in the th charge-discharge cycle, represents the L1 norm.

[0061] Then, the total spectral energy of the charging current sequence in the th charge-discharge cycle is obtained. Similarly, the total spectral energy of the discharge voltage sequence in the th charge-discharge cycle can be obtained through the above steps.

[0062] The total spectral energy of the charging current sequence and the total spectral energy of the discharge voltage sequence are used as part of the features. Other features include the average resistance of the charge-discharge cycle, the time of the constant-voltage charging stage, the time of the constant-current discharge stage, the average slope of the current decay from A (e.g., with a value of 0.6 A) to B (e.g., with a value of 0.2 A) in the constant-voltage charging stage of the charge-discharge cycle, and the average slope of the voltage decay from C (e.g., with a value of 3.0 V) to D (e.g., with a value of 2.0 V) in the constant-current discharge stage. Here, A is greater than B, and C is greater than D. The calculation formulas of each feature within the th charge-discharge cycle are as follows:

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] Among them, represents the average resistance of the th charge-discharge cycle, represents the internal resistance value at the th sample of the current sequence (the sequence collected during the entire charging and discharging process), represents the total number of sampling points of the current sequence, represents the time of the constant-voltage charging stage of the charge-discharge cycle, represents the time of the constant-voltage charging stage, represents the time of the constant-current discharge stage of the charge-discharge cycle, represents the time of the constant-current discharge stage, represents the average slope of the current decay from A (e.g., with a value of 0.6 A) to B (e.g., with a value of 0.2 A) in the constant-voltage charging stage of the charge-discharge cycle, The inverse function of the linear interpolation function of the current sequence in the constant voltage charging stage Indicates The average slope of the voltage decay from C to D in the constant current discharge stage during charge and discharge cycles The inverse function of the linear interpolation function of the voltage sequence in the constant current discharge stage

[0069] Step 1.2: Normalize the total spectrum energy and other features to obtain normalized features

[0070] For the total spectrum energy and other statistical features, for any one of the features, taking as an example, through normalization, it is transformed into the same range to reduce the data dimension difference and noise interference, that is:

[0071]

[0072] Among them, Represents the th feature Represents the th charge and discharge cycle data And Correspond to The maximum and minimum values respectively, Is the normalized feature obtained after normalization processing, and the vector composed of all normalized features in the th charge and discharge cycle is represented by the degenerate feature vector That fuses the time-frequency domain characteristics

[0073] Step 2: Through the convolutional neural network, according to the feature vector, output the dynamic health score of the lithium battery. Compared with the traditional time-domain features that can only statically describe the degradation process of the lithium battery through statistics such as mean and variance, the dynamic health score overcomes the limitation of the fixed statistical dimension in characterizing complex degradation patterns by introducing a non-linear dynamic modeling method

[0074] In this embodiment, a dynamic health score is constructed through a convolutional neural network , and the formula is as follows:

[0075]

[0076] Among them, Represents the convolutional neural network Represents the weights of the convolutional neural network

[0077] Step 3: Based on the dynamic health score, use the linear Wiener process model to construct a degradation process model of the remaining life of the lithium battery, and calculate the predicted value of the remaining life of the lithium battery based on the obtained degradation process model

[0078] During the initial iterative training, the dynamic health score output by the convolutional neural network is a non-stationary random process. In this embodiment, to balance model accuracy and computational efficiency, the complex non-stationary degradation process is simplified to a linear Wiener (Linear Wiener Process Model, LWPM) process by parameter constraints. Specifically, the maximum likelihood estimation constraint term of the linear Wiener process is added to the calculation of the loss function. Through iterative training, the transformation from the non-stationary random process to the linear Wiener process is realized, and thus the degradation process model of the remaining life of the lithium battery is constructed. The goal of this embodiment is to tend to the linear Wiener process. Therefore, based on the assumption that the dynamic health score conforms to the linear Wiener process, the degradation process model is constructed, and during the iterative training, the actual dynamic health score is driven to continuously tend to the linear Wiener process. After the iterative training is completed, a degradation process model that conforms to the linear Wiener process is finally obtained.

[0079] Step 3.1: Use the linear Wiener process to describe the degradation path of the dynamic health score during the entire life cycle of the lithium battery, including:

[0080]

[0081] Among them, represents the dynamic health score of the current th charge-discharge cycle, represents the initial dynamic health score, represents the drift coefficient and satisfies the normal distribution , characterizing individual differences, is the mean of the drift coefficient, is the standard deviation that the drift coefficient follows, represents the diffusion coefficient, is the standard Brownian motion. The increment of the dynamic health score is independent and identically distributed, and it follows the normal distribution, that is:

[0082]

[0083] Among them, represents the increment of the dynamic health score, and this increment is a random variable, represents a possible value of the increment of the dynamic health score, represents the time increment between two adjacent charge-discharge cycles, is the probability density function. According to the degradation path of the dynamic health score that conforms to the linear Wiener process, the first passage failure time is defined, expressed as:

[0084]

[0085] Among them, represents the The remaining life value of the lithium battery after the first charge-discharge cycle, and this remaining life value is a random variable, is the supremum, denotes the th possible value of the remaining life of the lithium battery after the th charge-discharge cycle, denotes the degradation path from the th charge-discharge cycle to the th charge-discharge cycle, denotes the failure threshold,

[0086]

[0087] Select 's mathematical expectation as the point estimate value, and the remaining life prediction function is obtained as follows:

[0088]

[0089] where, denotes the predicted remaining life value of the lithium battery after the th charge-discharge cycle.

[0090] Step 3.2. Measure the uncertainty by estimating the standard deviation of using the following formula:

[0091]

[0092] where, denotes the increment of the dynamic health score between two adjacent charge-discharge cycles, denotes the dynamic health score of the current th charge-discharge cycle, is the total number of charge-discharge cycles in the full life cycle of the lithium battery.

[0093] In this step, a linear Wiener process model is used to construct a degradation process model for the dynamic health score of the full life cycle of the lithium battery, preparing for the subsequent construction of the optimization objective function. After obtaining the dynamic health score of the current charge-discharge cycle, substituting it into the degradation process model can obtain the predicted remaining life value of the lithium battery.

[0094] Step 4. Calculate the loss function including the remaining life prediction error term and the maximum likelihood estimation constraint term of the linear Wiener process, and update the parameters of the convolutional neural network and the degradation process model according to the loss function until the training ends.

[0095] In this embodiment, by minimizing the loss function, the parameter weights in the dynamic convolutional neural network and the drift coefficient, diffusion coefficient, and failure threshold in the linear Wiener process are iteratively adjusted to balance the data-driven features and physical statistical characteristics, avoiding overfitting or underfitting caused by single optimization, specifically including:

[0096] Step 4.1: Calculate the mean square error based on the predicted remaining life value and the true remaining life value corresponding to each charge-discharge cycle in the full life cycle of the lithium battery, and use the mean square error as the remaining life prediction error term:

[0097]

[0098] where The mean square error is approximately represents the true remaining life value of the lithium battery. By combining the dynamic health score output by the convolutional neural network with Step 3.1, the corresponding predicted remaining life value can be obtained.

[0099] Step 4.2: Introduce the maximum likelihood estimation constraint term of the linear Wiener process into the loss function, which is the logarithmic joint likelihood function of all dynamic health score increments, expressed as:

[0100]

[0101] where is the logarithmic joint likelihood function.

[0102] Take the logarithmic joint likelihood function after removing the constant as the maximum likelihood estimation constraint term of the linear Wiener process, expressed as:

[0103]

[0104] where represents the linear Wiener process constraint. Weightedly fuse the mean square error constraint and the maximum likelihood estimation constraint of the linear Wiener process into the total loss function, that is:

[0105]

[0106] where is the total loss function, is the weight hyperparameter.

[0107] Step 4.3: Use the gradient descent method to minimize the loss function and solve to obtain the optimal parameter set , represents the optimal parameter weights in the convolutional neural network, represents the mean of the optimal drift coefficient, represents the standard deviation of the optimal drift coefficient, represents the optimal diffusion coefficient, Represents the optimal failure threshold.

[0108] In this embodiment, the gradient descent method is used to minimize the loss function to obtain the optimal parameter set , and finally, the feedback of the dynamic health score and modeling is realized, thereby forming a closed-loop feedback mechanism between the dynamic health score construction and the linear Wiener process model.

[0109] Step 5: Obtain the real-time dynamic health score of the target lithium battery through the updated convolutional neural network, and use Bayesian theory to update the drift coefficient of the degradation process model in real time, and output the remaining life prediction value of the target lithium battery using the latest degradation process model.

[0110] In this embodiment, through the obtained optimal parameter set, the drift parameters corresponding to the target lithium battery data set are updated in real time by Bayesian, and the remaining life of the lithium battery is predicted in real time by calculating the expected value of the defined first-passage failure time, thereby realizing the dynamic integration from offline joint optimization to online probabilistic inference, including the following steps:

[0111] Step 5.1: For the target lithium battery sample, perform a discrete Fourier transform on the data obtained by multiple sensors during the current charge and discharge cycle of the target lithium battery. The discrete Fourier transform method is the same as before.

[0112] Step 5.2: Calculate the corresponding total spectral energy and other features, and perform feature normalization to obtain the degradation feature vector.

[0113] Step 5.3: Use the optimal convolutional neural network parameter weights , and output the real-time dynamic health score according to the degradation feature vector of the target lithium battery.

[0114] Step 5.4: Use Bayesian theory to update the drift parameters, and use the following formula:

[0115]

[0116] Where, represents the observed data of the dynamic health score of the target lithium battery from the 0th to the th charge and discharge cycle, represents the probability of the event occurring under the condition that the event is known, represents the prior of the drift parameter, which fuses the observed prior distribution of the historical dynamic health score and the real-time dynamic health score. The updated drift parameter follows the posterior distribution, and the posterior distribution of the drift parameter can be calculated , and its formula is as follows:

[0117]

[0118] Among them, is the standard deviation of the drift coefficient after Bayesian theory update, is the mean of the drift coefficient after Bayesian theory update, is the diffusion coefficient in the optimal parameter set, is the standard deviation of the drift coefficient in the optimal parameter set, is the current charge-discharge cycle number of the target lithium battery, is the time increment between two adjacent charge-discharge cycles, is the drift coefficient in the optimal parameter set, is from the th charge-discharge cycle corresponding to the dynamic health score increment of the target lithium battery.

[0119] Step 5.5: Substitute the optimal diffusion coefficient and failure threshold obtained at the end of training, and the drift coefficient after the current Bayesian theory update into the degradation process model to obtain the latest degradation process model; then substitute the dynamic health score of the current charge-discharge cycle of the target lithium battery into the latest degradation process model, and output the real-time remaining life prediction value of the current charge-discharge cycle of the target lithium battery. The remaining life prediction function in real-time prediction adopts the following formula:

[0120]

[0121] Among them, represents the real-time remaining life prediction value of the current charge-discharge cycle of the target lithium battery.

[0122] In this embodiment, a dynamic update mechanism for linear Wiener process parameters is established based on the Bayesian inference framework in real-time prediction to realize online real-time prediction and uncertainty quantification of the remaining life of lithium batteries. It should be noted that Steps 1-Step 4 are the training process, which usually requires multiple iterations until the training end condition is reached. And Step 5 is the application process, which is executed according to actual needs for the corresponding number of times. And the training process can be executed alone, the application process can be executed alone, or the training process and the application process can be executed simultaneously.

[0123] The present invention captures the non-stationary and non-linear characteristics of data effectively by collecting the full life cycle data of lithium batteries through multi-sensor collaboration and constructing a dynamic health score in combination with a convolutional neural network. Further, a non-stationary stochastic process is adopted to quantify the uncertainty of the health score, and relevant parameters are synchronously adjusted by jointly optimizing the objective function to enhance the adaptability of the model to complex degradation patterns. Based on the Bayesian inference framework, the parameters of the non-stationary stochastic process are dynamically updated, and the conjugate prior distribution of historical data and real-time observations is combined to achieve adaptive adjustment of parameters and dynamic prediction of the remaining life of lithium batteries. This method combines the advantages of data-driven in multi-source information processing with the ability of probability models in uncertainty quantification, making up for the lack of physical interpretability in traditional data-driven methods and overcoming the problem of insufficient accuracy caused by the complex degradation mechanism of a single model-driven method.

[0124] In a specific embodiment, this embodiment is verified through the publicly available lithium battery dataset of the University of Maryland, as follows:

[0125] It is set that the constant voltage mode (CV, 4.2V) is adopted during the charging process. As the battery capacity increases, the charging current decays non-linearly until it terminates when it reaches 0.2A. The constant current mode (CC, 0.2A) is adopted during the discharging process. As the battery capacity decreases, the terminal voltage decreases monotonically until it terminates when it reaches 2.7V, thus constituting a complete charge-discharge cycle. Seven charge-discharge cycle characteristics are extracted from the full life cycle data, including the total spectral energy of the charging current sequence, the total spectral energy of the discharging voltage sequence, the average resistance of the charge-discharge cycle, the time of the constant voltage charging stage, the time of the constant current discharging stage, the average slope of the charging current decaying from 0.6A to 0.2A during the constant voltage charging stage, and the average slope of the discharging voltage decaying from 3.8V to 2.8V during the constant current discharging stage. The CS2_35 lithium battery dataset is used to model the dynamic health score degradation process, and the optimal convolutional neural network parameters and degradation process model parameters are obtained. On this basis, the online prediction of the remaining life is realized for the CS2_36 lithium battery dataset using Bayesian update. Figure 2 It can be seen from [reference] that the parameters of the linear Wiener process show good convergence characteristics during the iteration process and finally reach a stable state; Figure 3 and Figure 4 It can be seen from [reference] that the convolutional neural network constructed based on the CS2_35 lithium battery dataset can effectively characterize the battery degradation process, and the remaining life prediction results of the degradation process model are highly consistent with the real degradation trajectory; it can be seen from Figure 5 that the application effect of the Bayesian online update strategy on the CS2_36 lithium battery dataset. Through the real-time parameter correction mechanism, the mean square error between the prediction result and the measured value is controlled within the ideal range of 10.81 days. Generally speaking, the fitting accuracy of the online remaining life of the method in this application reaches 99.48%, and the remaining life prediction effect is good.

[0126] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, 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, it should be considered as falling within the scope described in this specification.

[0127] The above-described embodiments merely represent several implementation manners of the present invention, and the description thereof is relatively specific and detailed. However, it should not be construed as a limitation to the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the appended claims.

Claims

1. A real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling, characterized in that, The real-time prediction method for the remaining useful life of a lithium battery based on dynamic uncertainty modeling includes: Obtain the full-life cycle data of the lithium battery collected by multiple sensors, and combine with the discrete Fourier transform to extract the degradation feature vectors corresponding to each charge-discharge cycle in the full-life cycle of the lithium battery; Output the dynamic health score of the lithium battery through a convolutional neural network according to the degradation feature vectors; Based on the dynamic health score, construct a degradation process model for the remaining useful life of the lithium battery using a linear Wiener process model, and calculate the predicted value of the remaining useful life of the lithium battery based on the obtained degradation process model; Calculate the loss function including the remaining useful life prediction error term and the maximum likelihood estimation constraint term of the linear Wiener process, and update the parameters of the convolutional neural network and the degradation process model according to the loss function until the training ends; Obtain the real-time dynamic health score of the target lithium battery through the updated convolutional neural network, and use the Bayesian theory to update the drift coefficient of the degradation process model in real time, and output the real-time predicted value of the remaining useful life of the target lithium battery using the latest degradation process model.

2. The real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling according to claim 1, characterized in that The combination with the discrete Fourier transform to extract the degradation feature vectors corresponding to each charge-discharge cycle in the full-life cycle of the lithium battery includes: For each charge-discharge cycle, perform a discrete Fourier transform on the charging current sequence to calculate the corresponding total spectral energy; perform a discrete Fourier transform on the discharge voltage sequence to calculate the corresponding total spectral energy; And calculate the average resistance of the charge-discharge cycle, the time of the constant voltage charging stage, the time of the constant current discharge stage, the average slope of the current in the constant voltage charging stage decaying from A to B, and the average slope of the voltage in the constant current discharge stage decaying from C to D in the current charge-discharge cycle, where A is greater than B, and C is greater than D; Normalize the total spectral energy corresponding to the charging current sequence, the total spectral energy corresponding to the discharge voltage sequence, the average resistance of the charge-discharge cycle, the time of the constant voltage charging stage, the time of the constant current discharge stage, the average slope of the current in the constant voltage charging stage decaying from A to B, and the average slope of the voltage in the constant current discharge stage decaying from C to D respectively, and combine all the normalized features as the degradation feature vectors.

3. The real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling according to claim 1, characterized in that, The construction of the degradation process model for the remaining useful life of the lithium battery using a linear Wiener process model based on the dynamic health score includes: Use a linear Wiener process to describe the degradation path of the dynamic health score in the full-life cycle of the lithium battery, where the drift coefficient follows a normal distribution, and the increment of the dynamic health score follows a normal distribution; Define the first passage failure time according to the degradation path of the dynamic health score that conforms to the linear Wiener process; Construct the probability density function of the first passage failure time; Take the integral process of the probability density function of the first passage failure time as the remaining useful life prediction function to complete the construction of the degradation process model for the remaining useful life of the lithium battery.

4. The real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling according to claim 1, wherein, The remaining useful life prediction error term includes: Calculate the mean square error based on the predicted value of the remaining useful life and the true value of the remaining useful life corresponding to each charge-discharge cycle in the full-life cycle of the lithium battery, and use the mean square error as the remaining useful life prediction error term.

5. The real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling according to claim 1, characterized in that, The maximum likelihood estimation constraint term of the linear Wiener process includes: Calculate the dynamic health score increment between two adjacent charge-discharge cycles in the whole life cycle of a lithium battery; Construct the logarithmic joint likelihood function of all dynamic health score increments; Take the logarithmic joint likelihood function after removing the constant as the maximum likelihood estimation constraint term of the linear Wiener process.

6. The real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling according to claim 5, wherein The construction of the logarithmic joint likelihood function of all dynamic health score increments includes: ； Among them, represents the logarithmic joint likelihood function, is the total number of charge-discharge cycles in the entire life cycle of the lithium battery, is the increment of the dynamic health score corresponding to the th charge-discharge cycle, is the mean value of the drift coefficient, is the diffusion coefficient.

7. The real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling according to claim 1, characterized in that, The loss function is the weighted sum of the remaining life prediction error term and the maximum likelihood estimation constraint term of the linear Wiener process; Use the gradient descent method to minimize the loss function to update the parameters of the convolutional neural network and the degradation process model, and after the training is completed, output the optimal parameter set, where the optimal parameter set includes the optimal weight parameters of the convolutional neural network, the optimal drift coefficient of the degradation process model, the optimal diffusion coefficient of the degradation process model, and the optimal failure threshold of the degradation process model. The drift coefficient follows a normal distribution, that is, obtain the mean of the optimal drift coefficient and the standard deviation of the optimal drift coefficient.

8. The real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling according to claim 1, characterized in that The obtaining of the real-time dynamic health score of the target lithium battery through the updated convolutional neural network includes: Obtain multi-sensor data in the current charge-discharge cycle of the target lithium battery; Combine the discrete Fourier transform to extract the degradation feature vector of the multi-sensor data; Through the updated convolutional neural network, output the real-time dynamic health score of the target lithium battery according to the feature vector.

9. The real-time prediction method for the remaining life of a lithium battery based on dynamic uncertainty modeling according to claim 1, wherein, The real-time update of the drift coefficient of the degradation process model using Bayesian theory includes: Use Bayesian theory to real-time update the drift coefficient of the degradation process model. The updated drift coefficient follows the posterior distribution. Calculate the posterior distribution by combining the optimal parameters of the degradation process model obtained after the training is completed. The calculation formula is as follows: ； Among them, is the standard deviation of the drift coefficient after Bayesian theory update, is the mean of the drift coefficient after Bayesian theory update, is the optimal diffusion coefficient of the degradation process model, is the standard deviation of the optimal drift coefficient of the degradation process model, is the current charge-discharge cycle number of the target lithium battery, is the time increment between two adjacent charge-discharge cycles, is the mean of the optimal drift coefficient of the degradation process model, is for the target lithium battery at the th charge-discharge cycle corresponding dynamic health score increment.

Images