Method and system for predicting remaining service life of lithium ion battery
By combining external and internal parameters through a parallel hybrid deep neural network, and utilizing temporal convolutional neural networks and gated recurrent units, the problem of insufficient accuracy in lithium-ion battery life prediction is solved, achieving life prediction with higher accuracy and wider applicability.
Patent Information
- Application Number
- CN202511070138.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-31
- Publication Date
- 2025-09-19
AI Technical Summary
Existing methods for predicting the remaining useful life of lithium-ion batteries suffer from insufficient accuracy, especially data-driven methods that ignore the complex electrochemical reactions within lithium-ion batteries, resulting in inaccurate predictions.
A parallel hybrid deep neural network is used to combine external parameters such as battery current and voltage with internal parameters such as core temperature and expansion force. A multi-dimensional feature space is constructed through a dual-channel feature extraction method, and feature fusion prediction is performed using temporal convolutional neural networks and gated recurrent units.
The prediction accuracy of the remaining service life of lithium-ion batteries has been significantly improved, the error has been reduced, and the applicability and generalization ability in complex scenarios have been improved.
Smart Images
Figure CN120669152A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of lithium-ion battery life prediction, and in particular to a method and system for predicting the remaining service life of a lithium-ion battery. Background Art
[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.
[0003] Lithium-ion batteries are widely used in energy storage, aerospace, electric vehicles, and other fields, boasting high energy density, recyclability, and low self-discharge. However, when battery performance degrades to a certain level, the equipment running on them will experience performance degradation. If not promptly replaced, this can lead to downtime or even safety incidents. For example, extreme operating conditions such as high and low temperatures, and high charge and discharge rates can accelerate battery wear, significantly shortening battery life and generating increased heat, ultimately leading to safety incidents. Therefore, accurately predicting a battery's remaining useful life (RUL) is crucial to ensuring system reliability and safety.
[0004] In the existing technology, there are two main methods for predicting the remaining useful life of batteries: equivalent circuit model-based prediction methods and data-driven prediction methods. Among them, the prediction accuracy of the equivalent circuit model-based prediction method depends on the model parameters and is easily affected by the battery status and environmental noise, resulting in poor model adaptability. The data-driven prediction method achieves life prediction by analyzing the actual operating data of the battery, avoiding complex mechanism modeling. However, most existing data-driven prediction methods only consider external parameters such as voltage, current, and temperature, while ignoring the complex electrochemical reactions within lithium-ion batteries. Therefore, it is difficult to fully characterize the complex multi-physics field coupling state within the battery, which affects the prediction accuracy. Summary of the Invention
[0005] To address the above issues, the present invention proposes a method and system for predicting the remaining service life of lithium-ion batteries. By comprehensively considering external operating parameters such as battery current and voltage and internal state parameters such as internal core temperature and expansion force, a parallel hybrid deep neural network is used to fully exploit the correlation characteristics of internal and external data through a dual-channel feature extraction method, thereby achieving accurate prediction of the remaining service life of lithium-ion batteries.
[0006] In some embodiments, the following technical solutions are adopted: A method for predicting the remaining service life of a lithium-ion battery, comprising: Obtain the charging current and voltage data of the lithium-ion battery under test within any time interval, as well as the core temperature and expansion force data at the corresponding moment; The acquired data is input into the trained parallel hybrid deep neural network model to obtain the prediction results of battery capacity and remaining battery life; Among them, the parallel hybrid deep neural network includes a first encoder, a second encoder and a decoder. The first encoder extracts high-dimensional features related to life based on the input charging current and voltage data. The second encoder extracts high-dimensional features related to life based on the input core temperature and expansion force data. The high-dimensional features extracted by the two encoders are fused and input into the decoder. The decoder includes two independent gated recurrent units, which process feature sequences from forward and backward respectively. Among them, the forward gated recurrent unit calculates the forward hidden state from the starting point to the end point of the feature sequence to capture historical information; the backward gated recurrent unit calculates the backward hidden state from the end point to the starting point of the feature sequence to capture future information.
[0007] As a further solution, the first encoder or the second encoder adopts a temporal convolutional neural network to expand the receptive field of the convolution kernel by expanding the causal convolution structure; for the input data x exist t The operational formula of the dilated causal convolution at the moment is: ; in, F(x t ) is the output of the temporal convolutional neural network, F d is the receptive field, X is the input sequence, x t is the current input value, k is the convolution kernel size, f i represents the convolution kernel i values, x (t-d*i) express td*i The input value at the moment.
[0008] As a further solution, the size of the receptive field is specifically: ; in, F is the size of the receptive field, k is the size of the convolution kernel, d is the expansion factor.
[0009] As a further solution, the hidden state of the decoder at each time step is composed of the forward hidden state and the backward hidden state Jointly determine the final hidden state for: .
[0010] As a further solution, the improved dung beetle optimization algorithm is used to adjust the parameters of the parallel hybrid deep neural network model, and to optimize the number of hidden layer nodes, network training times and learning rate of the network.
[0011] As a further solution, in the improved dung beetle optimization algorithm, the global exploration and local development of the algorithm are balanced by dynamic inertia weight adjustment; based on the weighted mixed normal distribution and uniform distribution U The random number, early enhanced global search, and later accelerated convergence are described as follows; ; in, Time step t The inertia weight of and is the maximum and minimum value of the dynamic inertia weight, controlling the balance between global exploration and local development; is the maximum number of iterations, R is the adjustment factor of the mixture of Gaussian random exploration and uniform random; The angular momentum of the rolling step is dynamically adjusted through an adaptive angle adjustment strategy. As the exploration angle is narrowed over the iteration, the search gradually shifts from a wide range scan to a fine search. Specifically: ; in, The current iteration number t The angle of the lower output; The upper limit of the angle is set to prevent the system from being unstable due to excessive angle; The lower limit of the set angle is to prevent the adjustment from being stagnant due to the angle being too small; is the shape parameter of the Lévy distribution, which controls the heavy-tail characteristic of the step size; Then the position of the j-th behavior is updated as follows: ; in, is the position before updating, is the updated position, X best is the global optimal position.
[0012] As a further solution, the improved dung beetle optimization algorithm introduces a heavy-tailed distribution step size to achieve cross-region search, and adopts an improved boundary processing mechanism to reduce the loss of effective search areas and avoid falling into local optimality. Specifically: ; in, and They are the upper and lower bounds of the set position variable respectively.
[0013] In other embodiments, the following technical solutions are adopted: A lithium-ion battery remaining service life prediction system, comprising: a data acquisition module configured to acquire charging current and voltage data of the lithium-ion battery under test within any time interval, as well as core temperature and expansion force data at the corresponding moment; A prediction module is configured to input the acquired data into a trained parallel hybrid deep neural network model to obtain prediction results of battery capacity and remaining battery life; Among them, the parallel hybrid deep neural network includes a first encoder, a second encoder and a decoder. The first encoder extracts high-dimensional features related to life based on the input charging current and voltage data. The second encoder extracts high-dimensional features related to life based on the input core temperature and expansion force data. The high-dimensional features extracted by the two encoders are fused and input into the decoder. The decoder includes two independent gated recurrent units, which process feature sequences from forward and backward respectively. Among them, the forward gated recurrent unit calculates the forward hidden state from the starting point to the end point of the feature sequence to capture historical information; the backward gated recurrent unit calculates the backward hidden state from the end point to the starting point of the feature sequence to capture future information.
[0014] In other embodiments, the following technical solutions are adopted: A terminal device includes a processor and a memory, wherein the processor is used to implement instructions; the memory is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executing the above-mentioned method for predicting the remaining service life of a lithium-ion battery.
[0015] In other embodiments, the following technical solutions are adopted: A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device to implement the above-mentioned method for predicting the remaining service life of a lithium-ion battery.
[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention comprehensively considers external operating parameters such as battery current and voltage and internal state parameters such as internal core temperature and expansion force to construct a multi-dimensional feature space, which can more comprehensively characterize the electrical-thermal-mechanical coupling characteristics of the battery aging process, reducing the battery life prediction error by more than 40% compared with traditional methods.
[0017] (2) The present invention designs a parallel hybrid network of dual encoding heads (external parameter encoder + internal state encoder) and a joint decoder. Through sub-modal feature extraction and cross-modal feature fusion, the characterization capability of multi-dimensional information such as current / voltage (external) and temperature / expansion force (internal) is significantly improved. Experiments show that the feature extraction integrity is improved by more than 60% compared with the single encoder structure.
[0018] (3) The present invention can perform adaptive feature extraction based on data from any operating interval, and can achieve high-dimensional aging feature mining without the need for specific charging and discharging conditions, significantly improving the applicability of the method in complex practical scenarios (such as intermittent charging and discharging, dynamic loads).
[0019] Other features and advantages of additional aspects of the present invention will be given in part in the following description and in part will become obvious from the following description or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 This is a flow chart of a method for predicting the remaining service life of a lithium-ion battery according to an embodiment of the present invention; Figure 2 Schematic diagram of the parallel hybrid deep neural network model structure in an embodiment of the present invention; Figure 3 Schematic diagram of the terminal device structure for deploying the method for predicting the remaining service life of a lithium-ion battery in an embodiment of the present invention. DETAILED DESCRIPTION
[0021] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.
[0022] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.
[0023] Example 1 In one or more embodiments, a method for predicting the remaining service life of a lithium-ion battery is disclosed, combining Figure 1 , specifically including the following process: S101: Acquire charging current and voltage data of the lithium-ion battery to be tested within any time interval, as well as core temperature and expansion force data at the corresponding moment; Because the correlation between the data throughout the charging process and the battery life is not fixed, if the data used to extract features has a poor correlation with the lifespan, the prediction results will also be poor. Therefore, traditional methods require correlation analysis based on the data to find the charging range most relevant to the battery life to extract features. This brings certain limitations to the acquisition of charging data and the application scenarios of the prediction method.
[0024] In this embodiment, charging current and voltage data can be obtained at any time during the charging process of the lithium-ion battery under test. Because this embodiment uses multidimensional data to improve the robustness of the input terminal, and the trained parallel hybrid deep neural network model has the ability to automatically extract features and has good generalization performance, data within a subinterval obtained at any time can be directly used as input.
[0025] The core temperature can be obtained by embedding a temperature sensor inside the battery or implanting a fiber optic sensor; the expansion force data can be obtained by attaching a strain gauge to the battery surface, implanting a fiber optic sensor inside the battery, or using a thin film pressure measurement system.
[0026] S102: Input the acquired data into the trained parallel hybrid deep neural network model to obtain prediction results of battery capacity and remaining battery life.
[0027] In this embodiment, a parallel hybrid deep neural network (PHDNN) is constructed, and the correlation characteristics of internal and external data are fully exploited through a dual-channel feature extraction network. The parallel hybrid deep neural network (PHDNN) includes a first encoder, a second encoder, and a decoder. The first encoder extracts high-dimensional features related to the external signal and lifespan based on the input charging current and voltage data. The second encoder extracts high-dimensional features related to the internal signal and lifespan based on the input core temperature and expansion force data. The high-dimensional features extracted by the two encoders are fused and input into the decoder for prediction, resulting in a prediction result of the remaining battery life.
[0028] As a specific example, the inputs of the two encoders in this embodiment are: ; Where Input1 and Input2 are the input features of the first encoder and the second encoder respectively. I is the battery current, U is the battery voltage, T is the battery core temperature, F is the battery expansion pressure; the subscript represents A data series within a time period.
[0029] Combine Figure 2The first encoder and the second encoder have the same structure, both using the temporal convolutional neural network TCN, which expands the receptive field of the convolution kernel through a unique dilated causal convolution structure and obtains high-dimensional information related to the remaining battery life based on the input internal and external data.
[0030] Taking the structure of the first encoder as an example, through the hierarchical structure of convolutional layers, pooling layers, and fully connected layers, it automatically extracts high-dimensional features of the input data, gradually compresses information, reduces redundancy, and improves generalization ability.
[0031] Temporal Convolutional Neural Networks (TCNs) capture long-term dependencies in a hierarchical manner through the collaborative work of dilated causal convolutions and residual connections. The output of a TCN, in the temporal dimension, is only related to the data at the current time t and the data before time t, and does not contain any future data information. This means that there is a cause before there is an effect, ensuring causality. The input and output of the TCN network are guaranteed to be consistent in length, and the sequence is padded on the left to ensure consistency. The formula is as follows: ; in, For the l Layer in time t The hidden state, For the l The weights of the convolution kernels of the layers, K is the convolution kernel size, d (l) is the expansion factor of the lth layer, x The input characteristic parameters include the charging current ( I )、Charging voltage( V ), core temperature ( T ) and expansion pressure ( F )wait.
[0032] Dilated causal convolution refers to setting the dilation factor d to increase the receptive field of the convolution kernel. During the convolution operation, the interval determined by the dilation factor forms an interval sampling. Under the condition of the same convolution kernel, dilated causal convolution can see more "front" historical information than causal convolution. The size of the receptive field is shown in the following formula: ; In the formula F is the size of the receptive field, k is the size of the convolution kernel, d is the expansion factor.
[0033] For input data x exist t The operational formula of the dilated causal convolution at the moment is: ; in, F(x t ) is the result of TCN output, F d is the receptive field, X is the input sequence, x t is the current input value, k is the convolution kernel size, f i represents the convolution kernel i values, x (t-d*i) express td*i The input value at the moment.
[0034] Assume that the vector form of the convolution kernel is , the vector form of the receptive field is , the offset size is set to b , then the output of a neuron is y It can be expressed as: ; in, As the activation function, the ReLU function is generally selected.
[0035] The output of a dilated causal convolution depends only on the timesteps preceding it, not on future ones. Therefore, the output of a TCN at the current timestep only considers the inputs from the previous timesteps covered by the convolution kernel, avoiding the possibility that the output of the current timestep will depend on future inputs. TCNs use dilated convolutions to expand the receptive field of the convolution kernel. By inserting "gaps" between the elements of the convolution kernel, the receptive field of the convolution is expanded while maintaining the kernel size, enabling the model to capture long-range dependencies.
[0036] The activation function is the core component of the nonlinear transformation of temporal convolutional neural networks (TCNs). It enables TCNs to learn and represent complex nonlinear characteristics, preventing models composed solely of linear transformations from failing to solve more common and widespread nonlinear problems. It also alleviates the problem of vanishing or exploding gradients and accelerates convergence.
[0037] In this embodiment, the activation function used by the temporal convolutional neural network TCN is the ReLU function, and its calculation formula is: ; The output range of the Relu function is , only in x<0 The gradient is 0 when θ is less than 0, and 1 at other times, which effectively alleviates the problem of gradient disappearance.
[0038] The principle of the Dropout layer can be called random inactivation, which makes some neurons fail with a certain probability during the forward propagation process of model training; the Dropout layer can alleviate the overfitting phenomenon in the network architecture because of its ability to have the role of integrated learning and regularization terms, thereby improving the generalization ability of the model.
[0039] In temporal convolutional neural networks (TCNs), the vanishing gradient phenomenon is very likely to occur due to the increase in the number of network layers. The residual connection structure is a means of combining shallow features with deep features to avoid the problem of vanishing gradients. Each TCN residual block consists of two layers of dilated causal convolution, weight normalization operation, ReLU function and Dropout layer. The mathematical expression can be expressed as: ; Where, x i Indicates the i The input of units, O i Indicates the i The output of the unit, h(x i ) yes x i The identity mapping of is a shallow feature, F(x i ) is the deep feature of the convolution operation, and Activation is the activation function.
[0040] In this embodiment, the decoder achieves high-precision and high-reliability prediction of the remaining life of the lithium battery through a bidirectional gated recurrent unit Bi-GRU neural network.
[0041] Bi-GRU introduces two independent GRU network layers to process the sequence from forward and backward directions respectively. The forward network layer calculates the forward hidden state from the start to the end of the sequence to capture historical information; the backward network layer calculates the backward hidden state from the end to the start of the sequence to capture future information.
[0042] Finally, the output of each time step is combined with the forward and backward hidden states by concatenation, summation or weighted averaging, thereby utilizing both past and future contextual information. In Bi-GRU, the hidden state of each time step is composed of the forward hidden state and the backward hidden state Jointly decide. Under the splicing method, the final hidden state is: .
[0043] In this embodiment, the training process of the parallel hybrid deep neural network PHDNN is as follows: Under set control conditions, multiple groups of lithium-ion batteries were charged and discharged separately. The experiment was terminated after a certain number of charge and discharge cycles. The termination criterion for the battery life was that the discharge capacity could reach up to 70% of the rated capacity. The battery charge and discharge test parameters (including charge and discharge voltage, current, battery core temperature and battery expansion pressure) and the corresponding battery capacity and remaining service life were obtained. Basic preprocessing operations were performed and used as the data set for parallel hybrid deep neural network (PHDNN) training.
[0044] Assume that a complete charging process requires T Seconds, the corresponding time interval can be recorded as [ t 0, t end ], let the time window be , the entire charging interval can be divided into N Equally timed subintervals: ; For the above N For each sub-interval, there is a corresponding data sequence of voltage, current, etc.
[0045] Based on 70% of the lithium battery experimental test data set, the neural network was trained according to each group of battery life prediction data; the remaining 30% of the experimental test data was used as the test set to test the trained neural network; the actual battery capacity and remaining service life were compared with the experimental measurement results data, and the predicted capacity error and prediction error of the prediction method were recorded and calculated to evaluate the prediction performance of the parallel hybrid deep neural network PHDNN.
[0046] This embodiment uses battery charge and discharge test parameters within multiple time periods to train the parallel hybrid deep neural network PHDNN, so that the model has the ability to automatically extract features. Therefore, in actual use, the battery data within the subinterval obtained at any time can be directly used as input.
[0047] In this embodiment, during the model training process, the parameter space is searched by improving the dung beetle optimization algorithm, the parameters of the parallel hybrid deep neural network PHDNN model are adaptively updated, and the number of hidden layer nodes, the number of network training times and the learning rate of the network are optimized.
[0048] Specifically, the Dung Beetle Optimizer (DBO) simulates the ball-rolling, foraging, and reproduction behaviors of dung beetles. This algorithm simultaneously considers global exploration and local exploitation, resulting in fast convergence and high accuracy. It can effectively solve complex optimization problems. Its basic optimization equation is shown below: ; in, Contains direction adjustment and step length control strategies; 、 They are the location information before and after the update respectively.
[0049] On this basis, this embodiment adds improved mechanisms including a dynamic inertia weight mechanism, an adaptive angle adjustment strategy, a Lévy flight hybrid strategy, and an improved boundary processing mechanism to enhance the global search capability.
[0050] Specifically, the dynamic inertia weight mechanism aims to balance global exploration and local development, based on the weighted mixed normal distribution and uniform distribution U The random number is adjusted dynamically through inertia weights to balance the algorithm's global exploration (high weight) and local development (low weight). This enhances global search in the early stages and accelerates convergence in the later stages, as described below. ; in, Time step t The inertia weight of and are the maximum and minimum values of the dynamic inertia weight, respectively, which control the balance between global exploration and local development; is the maximum number of iterations, R is a regulation factor for the mixture of Gaussian random exploration and uniform random.
[0051] The angular momentum of the rolling step is dynamically adjusted through an adaptive angle adjustment strategy. As the exploration angle is narrowed over the iteration, the search gradually shifts from a wide range scan to a fine search. Specifically: ; in, Current iteration number t The angle of the lower output; The upper limit of the angle is set to prevent the system from being unstable due to excessive angle; The lower limit of the set angle is to prevent the adjustment from being stagnant due to the angle being too small; is the shape parameter of the Lévy distribution, which controls the heavy-tail characteristic of the step size; Then the position of the j-th behavior is updated as follows: ; is the position before updating, is the updated position, X best is the global optimal position.
[0052] The Lévy flight hybrid strategy is used to enhance the ability to escape local extreme values. Its steps include generating standard normal distribution random vectors, calculating scaling factors, and constructing Lévy step sizes. It can generate long jump steps with low probability, breaking the algorithm's premature convergence. It is particularly suitable for high-dimensional multimodal problems. The details are as follows: ; In the formula, typical parameters =1.5, represents element-wise product, is the update step size, is the global optimal position, is the current location, and are normally distributed and uniformly distributed random numbers, is the gamma function, is the scaling factor, Generated for the Lévy step size.
[0053] By introducing a heavy-tailed distribution step size to achieve cross-region search, an improved boundary processing mechanism is adopted to reduce the loss of the effective search area and avoid falling into the local optimum. Specifically: ; in, and The upper and lower bounds of the position variable are set.
[0054] Improve the random reset strategy in the original boundary replacement algorithm to reduce the loss of effective search area.
[0055] As a specific implementation method, after obtaining the prediction results of the parallel hybrid deep neural network PHDNN through the above process, the root mean square error (RMSE) and absolute error (AE) are used to evaluate the battery capacity prediction and RUL prediction results respectively; the specific formulas are as follows: ; ; Where, N is the number of prediction points, Q r,i is the actual capacity value of the test sample, Q p,i is the capacity prediction value of the sample point, RUL r is the actual number of remaining available cycles, RUL p To predict the remaining available cycles.
[0056] This embodiment can comprehensively verify the accuracy of the parallel hybrid deep neural network PHDNN in predicting the remaining life of lithium-ion batteries by calculating the predicted capacity error and prediction error.
[0057] The dual-decoder architecture of this embodiment can extract features separately from sensor signals at different scales, improving prediction accuracy. During the generation process, the decoder gradually decompresses the information compressed by the encoder and converts it into meaningful output. The proposed data-driven framework encodes the entire input sequence into an internal representation, allowing the model to take into account the global contextual information in the input sequence and capture complex input-to-output mappings. Furthermore, its design structure is very simple and involves few parameters, significantly reducing the algorithmic difficulty of designing experiments and significantly improving the model's fit.
[0058] Example 2 In one or more embodiments, a system for predicting the remaining useful life of a lithium-ion battery is disclosed, specifically comprising: a data acquisition module configured to acquire charging current and voltage data of the lithium-ion battery under test within any time interval, as well as core temperature and expansion force data at the corresponding moment; A prediction module is configured to input the acquired data into a trained parallel hybrid deep neural network model to obtain prediction results of battery capacity and remaining battery life; Among them, the parallel hybrid deep neural network includes a first encoder, a second encoder and a decoder. The first encoder extracts high-dimensional features related to life based on the input charging current and voltage data. The second encoder extracts high-dimensional features related to life based on the input core temperature and expansion force data. The high-dimensional features extracted by the two encoders are fused and input into the decoder. The decoder includes two independent gated recurrent units, which process feature sequences from forward and backward respectively. Among them, the forward gated recurrent unit calculates the forward hidden state from the starting point to the end point of the feature sequence to capture historical information; the backward gated recurrent unit calculates the backward hidden state from the end point to the starting point of the feature sequence to capture future information.
[0059] It should be noted that the specific implementation process of each of the above modules is exactly the same as that in Example 1 and will not be described in detail.
[0060] Example 3 In one or more embodiments, a terminal device is disclosed, which includes a processor and a memory, wherein the processor is used to implement instructions; the memory is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executing the lithium-ion battery remaining service life prediction method described in Example 1.
[0061] As a specific example, the method in the first embodiment is deployed in an embedded computer system, the structure of which is as follows: Figure 3 shown.
[0062] Embedded computer systems are computing terminals designed specifically for specific tasks. They primarily consist of hardware, middleware, and software layers. They can be deployed locally for edge computing and offer advantages such as strong real-time response capabilities, low power consumption, and high reliability. They serve as the physical carrier for this example.
[0063] The hardware layer of an embedded system is its physical foundation, responsible for performing calculations, storing data, connecting to external devices, and implementing specific functions. Among these layers are the embedded microprocessor, its "brain," which performs its core computing functions; storage systems that balance speed, capacity, and reliability, including ROM and SDRAM; peripheral interfaces for connecting sensors, actuators, and communication modules, including DACs, ADCs, and SPIOs; general interfaces for external communication, including serial ports, Ethernet, and USB; and human-computer interaction interfaces that respond to user operations, including touch screens, mice, and keyboards.
[0064] The driver layer of an embedded system is a key module that connects hardware devices to the operating system / middleware. It is responsible for directly controlling hardware peripherals and providing standardized operational interfaces. Its core task is to transform physical hardware behavior into logical services that can be called by software.
[0065] The software layer of an embedded system is the top-level architecture directly oriented toward business logic and user interaction, implementing high-level tasks such as functional integration, data processing, and decision-making control. It includes a real-time operating system (RTOS), various libraries, protocol stacks, and middleware such as communication protocols and file systems. The embedded software layer is the ultimate vehicle for delivering business value, and its design requires a deep integration of domain knowledge, hardware features, and algorithmic capabilities, such as the algorithm design in this example.
[0066] This embodiment operates on a lithium-ion battery system, comprising hundreds or even thousands of lithium-ion battery cells, as well as auxiliary subsystems and supporting components. The battery management system's core tasks include balancing management, safety protection, and communication interaction. The power conversion system is responsible for converting AC and DC power, regulating voltage and frequency, and controlling bidirectional energy flow. Auxiliary equipment includes air conditioning and fire protection systems. Sensors used for battery status monitoring include current sensors, voltage sensors, and temperature sensors.
[0067] Example 4 In one or more embodiments, a computer-readable storage medium is disclosed, in which a plurality of instructions are stored. The instructions are suitable for being loaded by a processor of a terminal device and executing the method for predicting the remaining useful life of a lithium-ion battery described in Example 1.
[0068] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.
Claims
1. A method for predicting the remaining service life of a lithium-ion battery, characterized in that: include: Obtain the charging current and voltage data of the lithium-ion battery under test within any time interval, as well as the core temperature and expansion force data at the corresponding moment; The acquired data is input into the trained parallel hybrid deep neural network model to obtain the prediction results of battery capacity and remaining battery life; Among them, the parallel hybrid deep neural network includes a first encoder, a second encoder and a decoder. The first encoder extracts high-dimensional features related to life based on the input charging current and voltage data. The second encoder extracts high-dimensional features related to life based on the input core temperature and expansion force data. The high-dimensional features extracted by the two encoders are fused and input into the decoder. The decoder includes two independent gated recurrent units, which process feature sequences from forward and backward respectively. Among them, the forward gated recurrent unit calculates the forward hidden state from the starting point to the end point of the feature sequence to capture historical information; the backward gated recurrent unit calculates the backward hidden state from the end point to the starting point of the feature sequence to capture future information.
2. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, wherein: The first encoder or the second encoder uses a temporal convolutional neural network to expand the receptive field of the convolution kernel by expanding the causal convolution structure; for the input data x exist t The operational formula of the dilated causal convolution at the moment is: ； in, F(x t ) is the output of the temporal convolutional neural network, F d is the receptive field, X is the input sequence, x t is the current input value, k is the convolution kernel size, f i represents the convolution kernel i values, x (t-d*i) express td*i The input value at the moment.
3. The method for predicting the remaining service life of a lithium-ion battery according to claim 2, wherein: The size of the receptive field is specifically: ； in, F is the size of the receptive field, k is the size of the convolution kernel, d is the expansion factor.
4. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, wherein: The hidden state of the decoder at each time step is composed of the forward hidden state and the backward hidden state Jointly determine the final hidden state for: .
5. The method for predicting the remaining service life of a lithium-ion battery according to claim 1, wherein: The improved dung beetle optimization algorithm is used to adjust the parameters of the parallel hybrid deep neural network model, and to optimize the number of hidden layer nodes, network training times and learning rate of the network.
6. A method for predicting the remaining service life of a lithium-ion battery according to claim 5, characterized in that: In the improved dung beetle optimization algorithm, the global exploration and local development of the algorithm are balanced by dynamic inertia weight adjustment; based on the weighted mixed normal distribution and uniform distribution U The random number, early enhanced global search, and later accelerated convergence are described as follows; ； in, Time step t The inertia weight of and is the maximum and minimum value of the dynamic inertia weight, controlling the balance between global exploration and local development; is the maximum number of iterations, R is the adjustment factor of the mixture of Gaussian random exploration and uniform random; The angular momentum of the rolling step is dynamically adjusted through an adaptive angle adjustment strategy. As the exploration angle is narrowed over the iteration, the search gradually shifts from a wide range scan to a fine search. Specifically: ； in, The current iteration number t The angle of the lower output; The upper limit of the angle is set to prevent the system from being unstable due to excessive angle; The lower limit of the set angle is to prevent the adjustment from being stagnant due to the angle being too small; is the shape parameter of the Lévy distribution, which controls the heavy-tail characteristic of the step size; Then the position of the j-th behavior is updated as follows: ； in, is the position before updating, is the updated position, X best is the global optimal position.
7. A method for predicting the remaining service life of a lithium-ion battery according to claim 6, characterized in that: In the improved dung beetle optimization algorithm, cross-region search is achieved by introducing a heavy-tailed distribution step size, and an improved boundary processing mechanism is used to reduce the loss of effective search areas and avoid falling into local optimality. Specifically: ； in, and They are the upper and lower bounds of the set position variable respectively.
8. A lithium-ion battery remaining service life prediction system, characterized in that: include: a data acquisition module configured to acquire charging current and voltage data of the lithium-ion battery under test within any time interval, as well as core temperature and expansion force data at the corresponding moment; A prediction module is configured to input the acquired data into a trained parallel hybrid deep neural network model to obtain prediction results of battery capacity and remaining battery life; Among them, the parallel hybrid deep neural network includes a first encoder, a second encoder and a decoder. The first encoder extracts high-dimensional features related to life based on the input charging current and voltage data. The second encoder extracts high-dimensional features related to life based on the input core temperature and expansion force data. The high-dimensional features extracted by the two encoders are fused and input into the decoder. The decoder includes two independent gated recurrent units, which process feature sequences from forward and backward respectively. Among them, the forward gated recurrent unit calculates the forward hidden state from the starting point to the end point of the feature sequence to capture historical information; the backward gated recurrent unit calculates the backward hidden state from the end point to the starting point of the feature sequence to capture future information.
9. A terminal device comprising a processor and a memory, wherein the processor is used to implement instructions; the memory is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method for predicting the remaining service life of a lithium-ion battery according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method for predicting the remaining service life of a lithium-ion battery according to any one of claims 1 to 7.