Battery multi-dimensional aging diagnosis method based on random short-time charging and multi-task learning

Abstract

The application discloses a battery multi-dimensional aging diagnosis method based on random short-time charging and multi-task learning, and the method comprises the following steps: constructing a half-cell model based on positive and negative electrode open-circuit potential curves and full-cell open-circuit voltage curves, identifying electrode aging parameters by using a cuckoo search algorithm, combining with test capacity to calculate lithium inventory loss LLI and negative active material loss LAMNE, and constructing an aging diagnosis label; determining an aging feature extraction voltage interval based on the main peak position of the capacity increment curve, selecting a continuous short-time charging sequence with a variable starting position in the voltage interval, and constructing an input matrix; and adopting a two-stage training mode to obtain the estimation results of capacity, lithium inventory loss LLI and negative active material loss LAMNE at the same time. The application reduces the dependence on complete charging and discharging data, and realizes the joint diagnosis of the electrode layer and the single layer aging state of the lithium ion battery.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery health status assessment technology, and in particular to a multidimensional aging diagnosis method for batteries based on random short-time charging and multi-task learning. Background Technology

[0002] With the development of clean energy utilization and transportation electrification, lithium-ion batteries are widely used in electric vehicles, energy storage systems, and electronic products. However, with increasing usage time, lithium-ion batteries inevitably age, mainly manifested as capacity decay, increased internal resistance, reduced charge / discharge efficiency, and decreased safety performance. Their internal aging modes primarily include lithium inventory loss and electrode active material loss. Currently, methods for diagnosing the health status and aging of lithium-ion batteries are mainly divided into data-driven methods and model-based methods. Data-driven methods typically use data such as voltage and current during charge / discharge processes to estimate battery capacity or health status, but most methods rely on complete charge / discharge data. Model-based methods can describe the internal behavior of the battery, but mechanistic models often suffer from numerous unknown parameters, complex modeling, and high computational costs. Furthermore, parameters such as resistance and capacitance in equivalent circuit models are often fitted values, making it difficult to directly reflect the true electrochemical aging mechanism of the battery.

[0003] In practical applications, the charging initiation state, charging duration, and operating conditions of batteries are uncertain, making it difficult to continuously obtain complete charge and discharge data. Existing studies using partial or random charging data mostly focus on capacity or health state estimation, with less emphasis on further quantifying electrode layer aging modes such as lithium inventory loss and electrode active material loss. Furthermore, existing studies typically perform battery capacity estimation and aging mode analysis separately, failing to fully utilize the correlation between capacity degradation and different aging modes. Some aging mode diagnostic methods still rely on complete charge and discharge data, making it difficult to meet the need for joint diagnosis of battery electrode layer and cell layer aging states under complex operating conditions.

[0004] Furthermore, existing aging diagnosis methods based on capacity increment curves typically use peak height, peak position, or peak area as artificial health features. Their feature extraction process relies on curve smoothing and numerical differentiation, and requires charging data to fully cover the main peak region of the capacity increment, making them susceptible to measurement noise and missing data. The aforementioned low-dimensional statistical features also struggle to fully preserve the local temporal coupling relationships between charging voltage, charging current, and charging capacity. Therefore, how to reduce reliance on complete charging data and artificial feature extraction while preserving physical interpretability is a problem that needs to be addressed in lithium-ion battery aging diagnosis under real-world operating conditions. In addition, some existing aging diagnosis methods are mainly validated based on ideal constant current operating condition data, and their applicability under complex discharge conditions such as real-world vehicle driving cycles still needs improvement. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-dimensional aging diagnosis method for batteries based on random short-time charging and multi-task learning, which reduces the dependence on complete charge and discharge data and realizes the joint diagnosis of the aging state of lithium-ion battery electrode layers and single cell layers.

[0006] To achieve the above objectives, the present invention provides the following solution: A multidimensional battery aging diagnosis method based on random short-time charging and multi-task learning includes the following steps: S1. Based on the charging operation data of lithium-ion battery samples, the open-circuit voltage curve of the full cell and the open-circuit potential curves of the positive and negative electrodes, a half-cell model is constructed to identify electrode aging parameters. The lithium inventory loss (LLI) and the negative electrode active material loss (LAMNE) are calculated in combination with the test capacity. A multi-dimensional aging diagnostic label set integrating the single-cell layer capacity and the electrode layer aging mode is constructed. S2. Construct a capacity increment curve based on charging operation data, determine the aging feature extraction voltage range according to its main peak position, and select continuous short-time charging sequence segments with non-fixed starting positions within the voltage range to construct an input matrix containing charging voltage, charging current and charging capacity. S3. Based on the input matrix and the multidimensional aging diagnosis label set, a two-stage training method is used to train the pre-set multi-task learning model. In the first training stage, the backbone of the shared convolutional feature is extracted based on the capacity label for pre-training. In the second training stage, multi-task joint training is performed based on the capacity, LLI and LAMNE labels. S4. Obtain continuous short-time charging sequence segments of the battery to be diagnosed within the aging feature extraction voltage range, input them into the trained multi-task learning model, and simultaneously output the capacity estimation results, LLI estimation results, and LAMNE estimation results, which are used as the aging diagnosis results of the single cell layer and the electrode layer, respectively, to realize the joint diagnosis of multi-dimensional aging of lithium-ion battery electrode-single cell.

[0007] Preferably, in S1, in the half-cell model, the open-circuit voltage of the full cell is expressed by the following formula:

[0008] in, This represents the open-circuit voltage of the full battery under the state of charge (SoC). Indicates the degree of lithiation of the positive electrode. The open circuit potential below, Indicates the degree of lithiation of the negative electrode The open-circuit potential below; and The value range is [0,1]. A value of 1 indicates that the corresponding electrode is fully lithiated, and a value of 0 indicates that the corresponding electrode is delithiated.

[0009] Preferably, in S1, constructing a half-cell model to identify electrode aging parameters specifically includes: An optimization algorithm was used to optimize the parameters of the half-cell model so that the open-circuit voltage curve reconstructed based on the half-cell model matched the open-circuit voltage curve of the full cell, thus obtaining the electrode aging parameters of the lithium-ion battery samples under different aging states. Electrode aging parameters include the degree of lithiation of the positive electrode at the end of discharge. Parameters of lithiation degree of the positive electrode at the end of charging Parameters of lithiation degree of the negative electrode at the end of discharge The degree of lithiation of the negative electrode at the end of charging. ;in, EoD This indicates the end of the discharge process. EoC This indicates the end of charging; the optimization algorithm is the Cuckoo Search algorithm; the Cuckoo Search algorithm is used, with the fitting error between the open-circuit voltage curve reconstructed based on the half-cell model and the full-cell open-circuit voltage curve as the objective function. and Optimize the identification; based on the open-circuit potential difference relationship between the positive and negative electrodes satisfied by the minimum and maximum cutoff voltages of the lithium-ion battery samples, determine respectively and .

[0010] Preferably, in S1, the lithium inventory loss (LLI) and the negative electrode active material loss (LAMNE) are calculated based on the test capacity, specifically including: For any of the lithium-ion battery samples t The aging state, according to the first t Test capacity under aging conditions Parameters of lithiation degree of the positive electrode at the end of discharge. The degree of lithiation parameters of the positive electrode at the end of charging. Determine the first t Available capacity of the positive electrode under aging conditions ; Based on test capacity Parameters of lithiation degree of the negative electrode at the end of charging Parameters of lithiation degree of the negative electrode at the end of discharge Determine the first t Available capacity of the negative electrode under aging conditions ; According to the t The available capacity of the positive electrode and the available capacity of the negative electrode under aging and initial conditions, as well as the degree of lithiation of the positive and negative electrodes under the end of charging conditions, are used to determine the lithium inventory loss (LLI). According to the tAvailable capacity of the negative electrode under aging conditions and the available capacity of the negative electrode in the initial state Determine the loss of negative electrode active material LAM NE ; Where t represents the current aging state of the lithium-ion battery sample, and 0 represents the initial state of the lithium-ion battery sample.

[0011] Preferred, available capacity of the positive electrode Available capacity of negative electrode Lithium inventory loss (LLI) and negative electrode active material loss (LAM) NE The following formulas are used to determine the values ​​respectively:

[0012]

[0013]

[0014]

[0015] in, Indicates the lithium-ion battery sample in the first t Test capacity under aging conditions, and They represent the first t Available capacity of the positive electrode and available capacity of the negative electrode under aging conditions. and These represent the available capacity of the positive electrode and the available capacity of the negative electrode in the initial state, respectively.

[0016] Preferably, in S2, the aging characteristic extraction voltage range is determined based on the main peak position of the capacity increment curve, specifically including: The charging capacity of the lithium-ion battery sample during the charging process is calculated based on the charging current and the corresponding sampling time. The capacity increment curve of the lithium-ion battery sample under different aging states is constructed based on the rate of change of the charging capacity relative to the charging voltage. The main peak position of the capacity increment curve under different aging states is determined, and the aging feature extraction voltage range is determined based on the main peak position.

[0017] Preferably, in S3, the multi-task learning model includes an input layer, four convolutional modules, and multiple task output branches; the four convolutional modules constitute a shared convolutional feature extraction backbone, and each convolutional module includes a two-dimensional convolutional layer, a batch normalization layer, and a linear rectified function ReLU layer. The first three convolutional modules also include a max pooling layer. In the first training phase, the shared convolutional feature extraction backbone connection capacity prediction branch is pre-trained; in the second training phase, the pre-trained shared convolutional feature extraction backbone connection capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM) are used to extract the negative electrode active material loss. NE The system has multiple corresponding task output branches, each of which outputs its own estimation results through the corresponding fully connected layer and regression layer.

[0018] Preferably, in S3, a two-stage training method is used to train the pre-defined multi-task learning model, specifically including: In the first training phase, the capacity in the multidimensional aging diagnostic label set is used as the supervision label. The input matrix is ​​input into the shared convolutional feature extraction backbone, and the shared convolutional feature extraction backbone is pre-trained based on the mean square error between the capacity prediction result and the corresponding test capacity. In the second training phase, the output branches of the capacity task, the lithium inventory loss (LLI) task, and the negative electrode active material loss (LAMNE) task are connected to the pre-trained shared convolutional feature extraction backbone. The task losses of capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE) are calculated respectively, and the weighted sum of the above three task losses is used as the joint loss function to jointly train the shared convolutional feature extraction backbone and multiple task output branches.

[0019] Preferably, the first two convolutional modules of the four convolutional modules are used to extract battery aging information that can be shared among capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE), while the latter two convolutional modules are used to extract battery aging information corresponding to different aging modes.

[0020] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning as described above.

[0021] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: 1. This invention constructs a half-cell model using the positive electrode open-circuit potential curve, the negative electrode open-circuit potential curve, and the full-cell open-circuit voltage curve. It also uses the Cuckoo Search algorithm to identify electrode aging parameters under different aging states, thereby quantifying lithium inventory loss (LLI) and negative electrode active material loss (LAMNE). This enables quantitative diagnosis of the aging mode of lithium-ion battery electrode layers, which is beneficial for characterizing the aging state of lithium-ion batteries from multiple dimensions of electrode layers and single cell layers.

[0022] 2. This invention determines the voltage range for extracting aging features with physical interpretation based on capacity increment analysis, and directly uses the original sequence of charging voltage, charging current and charging amount for any continuous preset duration within the voltage range for automatic feature extraction. This avoids relying on complete charging curves and manually extracting peak height, peak position and peak area of ​​capacity increment curves, and improves the adaptability to random charging start position and incomplete data.

[0023] 3. This invention employs a multi-task learning model to extract shareable battery aging information among capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE), and outputs corresponding estimation results through multiple task output branches, thereby achieving joint estimation of single-cell capacity degradation and electrode layer aging modes. Compared with establishing single-task learning models for different aging indicators, this invention reduces the model size and computational burden, and improves the accuracy, computational efficiency, and robustness of multi-dimensional aging diagnosis of lithium-ion battery electrodes and single cells under complex operating conditions.

[0024] 4. This invention adopts a two-stage training method. First, the shared convolutional feature extraction backbone is pre-trained using capacity labels, so that the convolutional module learns stable basic aging features. Then, joint training is performed based on three related tasks: capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE). By sharing aging information, the utilization efficiency of limited aging labels is improved, and the repetitive computation and storage overhead caused by building multiple independent single-task models is avoided. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning provided by the present invention. Figure 2 This is a schematic diagram of electrode aging parameter identification based on positive and negative open-circuit potential matching according to the present invention. Figure 3 This is a schematic diagram illustrating the present invention of determining the voltage range for aging feature extraction based on the capacity increment curve and selecting a short-time charging sequence; Figure 4 This is a schematic diagram of the multi-task learning model structure based on two-stage training according to the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] Example 1 This application provides a battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning. Figure 1 A flowchart of a battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning is shown. (See attached diagram) Figure 1 The method mainly includes the following steps: Step 1: Obtain charging operation data and full-cell open-circuit voltage curves of lithium-ion battery samples under different aging conditions, and obtain the positive electrode open-circuit potential curves and negative electrode open-circuit potential curves corresponding to the lithium-ion battery samples.

[0030] The charging operation data includes charging voltage, charging current, and the sampling time corresponding to the charging voltage and charging current. After aging cycles, the lithium-ion battery samples are subjected to characteristic tests under different aging conditions to obtain the corresponding test capacity and full-cell open-circuit voltage curve. To obtain the positive electrode open-circuit potential curve and the negative electrode open-circuit potential curve, before the start of the aging cycle test, a reference battery cell with the same electrode material system as the lithium-ion battery sample is selected. The reference battery cell is disassembled in an argon atmosphere, and the positive electrode material and negative electrode material are extracted separately. The positive electrode button cell and the negative electrode button cell are then assembled. The positive electrode button cell and the negative electrode button cell are tested at preset temperatures to obtain the positive electrode open-circuit potential curve and the negative electrode open-circuit potential curve.

[0031] Step 2: Construct a half-cell model based on the positive and negative open-circuit potential curves, so that the open-circuit voltage of the full cell is expressed as the potential difference between the positive and negative open-circuit potentials.

[0032] Step 3: The cuckoo search algorithm is used to optimize the parameters of the half-cell model so that the open-circuit voltage curve reconstructed based on the half-cell model matches the open-circuit voltage curve of the full cell, thereby obtaining the electrode aging parameters of the lithium-ion battery sample under different aging conditions. Figure 2 A schematic diagram of electrode aging parameter identification based on the matching of positive and negative open-circuit potentials is shown.

[0033] Step 4: Obtain the test capacity of lithium-ion battery samples under different aging conditions, and calculate the lithium inventory loss (LLI) and negative electrode active material loss (LAM) based on the test capacity and electrode aging parameters. NE The test capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM) under different aging conditions were compared. NE This constitutes a set of diagnostic labels for aging.

[0034] Here, the test capacity is used to characterize the aging state of the single-cell layer of the lithium-ion battery sample, lithium inventory loss (LLI), and negative electrode active material loss (LAM). NE The electrode layer aging state of lithium-ion battery samples is used to characterize the aging state of the samples, and the three together serve as the output labels of the multi-task learning model.

[0035] Step 5: Calculate the charging capacity based on the charging current and the corresponding sampling time, construct the capacity increment curve based on the charging voltage and the charging capacity, and determine the aging characteristic extraction voltage range of the lithium-ion battery sample based on the main peak position of the capacity increment curve. Figure 3 A schematic diagram is shown illustrating the determination of the voltage range for aging feature extraction based on the capacity increment curve and the selection of a short-time charging sequence.

[0036] Step 6: Select any continuous charging sequence segment of a preset duration from the charging data corresponding to the voltage range of the aging feature extraction, and arrange the charging voltage, charging current and charging capacity in the charging sequence segment according to the sampling time sequence to form a training input matrix.

[0037] The aging feature extraction voltage range is 3.6V to 3.8V, and the preset duration is 100s. Any continuous charging sequence segment of the preset duration refers to a continuous 100s charging data segment within the aging feature extraction voltage range, where the starting position is not fixed.

[0038] Step 7: The training input matrix and its corresponding aging diagnostic labels constitute the training samples, and a two-stage training method is used to train the multi-task learning model based on the convolutional neural network. In the first training stage, the shared convolutional feature extraction backbone is pre-trained based on the capacity label. In the second training stage, capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM) are connected after the pre-trained shared convolutional feature extraction backbone. NE The corresponding task output branches are then used for multi-task joint training. Figure 4 A schematic diagram of a multi-task learning model structure based on two-stage training is shown.

[0039] The multi-task learning model includes a shared feature extraction backbone and multiple task output branches. The shared feature extraction backbone is used to extract shareable battery aging features from the input matrix, and the multiple task output branches are used to output the estimation results of capacity, LLI and LAMNE, respectively.

[0040] Step 8: Obtain any continuous 100s charging sequence segment of the lithium-ion battery to be diagnosed within the aging feature extraction voltage range, and arrange the charging voltage, charging current and charging capacity in the charging sequence segment according to the sampling time sequence to form an online diagnostic input matrix.

[0041] Step 9: Input the online diagnostic input data into the trained multi-task learning model to obtain the capacity estimation results, lithium inventory loss (LLI) estimation results, and negative electrode active material loss (LAM) estimation results for the lithium-ion battery to be diagnosed. NE Estimation results.

[0042] Step 10: Use the capacity estimation result as the monomer layer aging diagnosis result, and combine the lithium inventory loss (LLI) estimation result and the negative electrode active material loss (LAM) result. NE The estimation results are used as the electrode layer aging diagnosis results to obtain the electrode-cell multidimensional aging diagnosis results of the lithium-ion battery to be diagnosed.

[0043] Therefore, by extracting charging data for any consecutive 100 seconds within the voltage range based on aging characteristics, it is possible to estimate the capacity degradation of lithium-ion batteries and to assess lithium inventory loss (LLI) and negative electrode active material loss (LAM). NE The characterized electrode aging modes are quantified to achieve multidimensional aging diagnosis of lithium-ion battery electrodes and cells.

[0044] This embodiment determines the aging feature extraction voltage range based on capacity increment analysis, and uses any continuous 100s of charging data within this voltage range as the input of the multi-task learning model. By sharing the feature extraction backbone, it extracts shareable battery aging information among capacity degradation, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE), and outputs the estimation results of capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE) through multiple task output branches, thereby reducing the dependence on complete charge and discharge data and realizing the joint diagnosis of the aging state of lithium-ion battery electrode layers and single cell layers.

[0045] In one embodiment, in step 1, multiple lithium-ion battery samples are subjected to aging cycle tests and characteristic tests are performed under different aging conditions to obtain charging operation data, test capacity, and full-cell open-circuit voltage curves under different aging conditions.

[0046] The aging cycle test includes a charging process and a discharging process based on actual vehicle operating conditions. The charging process adopts a 0.3C constant current-constant voltage charging method, and the discharging process adopts the recorded actual vehicle driving cycle conditions, and the cycle test is carried out within the range of 10% to 90% state of charge. During the charging process, the charging voltage, charging current and corresponding sampling time of the lithium-ion battery sample are collected as charging operation data under different aging conditions.

[0047] To obtain the positive and negative open-circuit potential curves, before the aging cycle test began, the battery cells corresponding to the lithium-ion battery samples were disassembled in an inert gas environment, and the positive and negative electrode materials were extracted separately. The positive and negative button cells were then assembled. The positive and negative button cells were tested to obtain the positive and negative open-circuit potential curves, respectively.

[0048] In one specific embodiment, 20 LiNiCoAlO2 / silicon-based graphite lithium-ion battery cells were used as lithium-ion battery samples. The rated capacity of the lithium-ion battery samples was 3.4Ah, the battery weight was 48g, the charging cut-off voltage was 4.2V, the discharging cut-off voltage was 2.65V, and the maximum discharge current was 13A. Capacity testing was conducted using the manufacturer's recommended rate; the full-cell open-circuit voltage test was performed using a 0.1C charge-discharge method to obtain the full-cell open-circuit voltage curve; the test temperature for the positive and negative electrode open-circuit potential curves was 25℃. In this embodiment, the negative electrode active material loss (LAMNE) was used to characterize the loss of active material on the negative electrode side. For the silicon-based graphite negative electrode system used, the volume change of the silicon-based material during cycling easily causes the active material particles to break and fail; therefore, this embodiment mainly focuses on the quantitative diagnosis of negative electrode active material loss.

[0049] Before the aging test began, initial characteristic tests were performed on each lithium-ion battery sample. This was followed by 10 real-vehicle driving cycle tests, and another characteristic test. Subsequently, a characteristic test was performed every 30 real-vehicle driving cycle tests until the end of the test. The data obtained from each characteristic test corresponded to one aging state of the lithium-ion battery sample.

[0050] In one embodiment, in steps 2 and 3, a half-cell model is constructed based on the positive electrode open-circuit potential curve, the negative electrode open-circuit potential curve, and the full-cell open-circuit voltage curve, and the cuckoo search algorithm is used to identify the electrode aging parameters of the lithium-ion battery sample under different aging states. Figure 2 A schematic diagram of electrode aging parameter identification based on the matching of positive and negative open-circuit potentials is shown.

[0051] Specifically, nonlinear fitting can be performed on the positive electrode open-circuit potential curve and the negative electrode open-circuit potential curve to obtain the correspondence between the positive electrode open-circuit potential and the positive electrode lithiation degree, as well as the correspondence between the negative electrode open-circuit potential and the negative electrode lithiation degree.

[0052] To reconstruct the open-circuit voltage curves of lithium-ion battery samples under different aging states, it is assumed that the electrochemical behavior of the positive and negative electrodes remains consistent during the aging process; that is, the correspondence between the positive electrode open-circuit potential and the degree of lithiation, as well as the correspondence between the negative electrode open-circuit potential and the degree of lithiation, remains unchanged. Based on this assumption, the open-circuit voltage of the full cell at any state of charge (SoC) is determined by the potential difference between the positive and negative electrode open-circuit potentials at the corresponding degree of lithiation, expressed by the following formula:

[0053] in, Indicates that the full battery is in a state of charge. Open circuit voltage below, Indicates the degree of lithiation of the positive electrode. The open circuit potential below, Indicates the degree of lithiation of the negative electrode The open-circuit potential below; and The range of values ​​is A value of 1 indicates that the corresponding electrode is fully lithiated, and a value of 0 indicates that the corresponding electrode is delithiated.

[0054] Under the current aging state, the relationship between the state-of-charge SoC and the degree of lithiumization of the positive and negative electrodes is expressed by the following formula:

[0055] in, and These parameters represent the degree of lithiation of the positive electrode at the end of discharge and the end of charging, respectively. and These parameters represent the degree of lithiation of the negative electrode at the end of discharge and the end of charge, respectively; EoD represents the end of discharge and EoC represents the end of charge.

[0056] At the end of discharge, the minimum cutoff voltage of the lithium-ion battery sample satisfies:

[0057] At the end of charging, the maximum cutoff voltage of the lithium-ion battery sample satisfies:

[0058] Based on the above relationship, the open-circuit voltage of the full battery can be expressed as the state of charge (SoC) and the degree of lithiation of the positive electrode at the end of discharge. And the degree of lithiation parameters of the negative electrode at the end of charging. The function.

[0059] Using the cuckoo search algorithm to and Optimization and identification are performed. The fitting error between the full-cell open-circuit voltage curve reconstructed based on the half-cell model and the full-cell open-circuit voltage curve obtained from actual testing is used as the objective function. The parameter identification result that minimizes the fitting error is obtained through the Cuckoo Search algorithm.

[0060] Furthermore, the degree of lithiation parameters of the negative electrode in the discharged state are determined based on the open-circuit potential difference relationship between the positive and negative electrodes satisfied by the minimum cutoff voltage. The degree of lithiation parameters of the positive electrode in the end-of-charge state are determined based on the open-circuit potential difference relationship between the positive and negative electrodes satisfied by the maximum cutoff voltage. Thus, the aging parameters of the four electrodes under the current aging state are obtained.

[0061] The electrode aging parameter identification process described above is performed on different aging states of lithium-ion battery samples to obtain the electrode aging parameters of lithium-ion battery samples under different aging states, and the open-circuit voltage curve of the full cell under the corresponding aging state is reconstructed based on the identified electrode aging parameters.

[0062] In one embodiment, in step 4, based on the electrode aging parameter identification results, and according to the test capacity of the lithium-ion battery sample under different aging states, the available capacity of the positive electrode, the available capacity of the negative electrode, the lithium inventory loss (LLI), and the negative electrode active material loss (LAM) are calculated. NE .

[0063] For the t-th aging state of a lithium-ion battery sample, the available capacity of the positive electrode Calculate using the following formula:

[0064] Available capacity of negative electrode Calculate using the following formula:

[0065] in, This represents the test capacity of a lithium-ion battery sample under the t-th aging condition. and Let represent the available capacity of the positive electrode and the available capacity of the negative electrode under the t-th aging state, respectively; and These parameters represent the degree of lithiation of the positive electrode in the end-of-discharge and end-of-charge states, respectively, under the t-th aging condition. and These represent the degree of lithiation parameters of the negative electrode in the t-th aging state at the end of discharge and the end of charge, respectively.

[0066] Based on the available capacity of the positive electrode and the available capacity of the negative electrode in the t-th aging state and the initial state, as well as the degree of lithiation of the positive and negative electrodes in the end-of-charge state, the lithium inventory loss LLI is calculated using the following formula:

[0067] Loss of negative electrode active material LAM NE Calculate using the following formula:

[0068] Wherein, the subscript t represents the current aging state of the lithium-ion battery sample, and the subscript 0 represents the initial state of the lithium-ion battery sample; and These represent the available capacity of the positive electrode and the available capacity of the negative electrode in the initial state, respectively.

[0069] The test capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE) under different aging conditions are used as aging diagnostic tags for the corresponding charging sequence segments to form an aging diagnostic tag set.

[0070] In one embodiment, in steps 5 and 6, the charging data of lithium-ion battery samples under different charge-discharge cycles are selected for capacity increment analysis to determine the aging feature extraction voltage range, and short-time charging sequence segments are selected within the aging feature extraction voltage range. Figure 3 The capacity increment curves of lithium-ion battery samples under different aging conditions are shown.

[0071] Specifically, the charging capacity of the lithium-ion battery sample during the charging process is determined based on the charging current and its corresponding sampling time, and the capacity increment curve is obtained based on the change of charging capacity with charging voltage.

[0072] The capacity increment curves of lithium-ion battery samples under different charge-discharge cycles were analyzed. As the lithium-ion battery samples aged, the capacity increment curves under different charge-discharge cycles shifted to varying degrees, and the peak shape and position of the capacity increment curves changed; however, the main peak position of the capacity increment curves under different aging states remained generally between 3.6V and 3.8V. Therefore, 3.6V–3.8V was determined as the aging characteristic extraction voltage range.

[0073] As the number of charge-discharge cycles increases, the peak value of the capacity increment curve gradually decreases and the peak position shifts, indicating that the capacity increment curve changes significantly with battery aging. The above changes in the capacity increment curve are related to battery aging changes such as lithium inventory loss and active material loss.

[0074] Within the voltage range of 3.6V to 3.8V, the capacity increment curve exhibits a distinct main peak. This main peak reflects electrochemical reactions such as lithium-ion insertion and deintercalation in the electrode active material, providing information related to the aging of lithium-ion battery samples. Therefore, charging data within the 3.6V to 3.8V voltage range were used for aging feature extraction.

[0075] Furthermore, considering the uncertainty of actual charging behavior, any continuous 100-second charging sequence segment is selected from the charging data corresponding to the voltage range for aging feature extraction. Any continuous 100-second charging sequence segment refers to 100 seconds of continuous charging data within the voltage range of 3.6V to 3.8V without a fixed starting position.

[0076] The charging voltage, charging current, and corresponding charging capacity in the charging sequence segments are used as input features and arranged according to the sampling time sequence to form a training input matrix. To verify the rationality of the three-channel input of charging voltage, charging current, and charging capacity, different combinations of inputs V, I, Q, VI, VQ, IQ, and VIQ are compared under the same model structure, training parameters, data partitioning, and input segment length. The results show that the VIQ three-channel input can comprehensively characterize the local electrochemical response, charging condition constraints, and capacity evolution information, and has good comprehensive performance and stability in the multi-task diagnosis of capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE). Therefore, this embodiment uses the VIQ three-channel as the input of the multi-task learning model.

[0077] Furthermore, to determine the preset duration of the charging sequence segment, VIQ charging sequence segments of 30s, 50s, 100s, 150s, and 200s were compared. As the input segment length increased from 30s to 100s, the diagnostic errors for capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE) significantly decreased. Further increasing to 150s or 200s resulted in only a small improvement in diagnostic accuracy, while training time, online inference time, and data requirements increased significantly. Therefore, in this embodiment, 100s was determined as the preset duration that balances diagnostic accuracy, data requirements, and computational efficiency.

[0078] In one embodiment, step 7 involves using the training input matrix and its corresponding aging diagnostic labels to form training samples, and training a multi-task learning model based on a convolutional neural network using a two-stage training method. In the first training stage, the shared convolutional feature extraction backbone is pre-trained based on the capacity label. In the second training stage, the task output branches corresponding to capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE) are connected after the pre-trained shared convolutional feature extraction backbone, and multi-task joint training is performed. Figure 4 A schematic diagram of a multi-task learning model structure based on two-stage training is shown.

[0079] The multi-task learning model takes into account charging voltage, charging current, and charging capacity from charging sequence segments. Its outputs include capacity estimation, lithium inventory loss (LLI) estimation, and negative electrode active material loss (LAMNE) estimation. The training input data originates from the charging process of lithium-ion battery samples. The labels for capacity, LLI, and LAMNE are derived from characteristic tests and electrode aging parameter identification results under different aging conditions.

[0080] Since the charging voltage, charging current, and charging capacity have different numerical ranges, the Zero Center method is used to normalize the training input matrix in the input layer. The normalized training input matrix is ​​then input into the convolution module for battery aging feature extraction. At the same time, the output labels corresponding to capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE) are normalized.

[0081] The multi-task learning model consists of an input layer, four convolutional modules, a fully connected layer, and a regression layer. Each convolutional module includes a two-dimensional convolutional layer, a batch normalization layer, and a Rectified Linear Unit (ReLU) layer. The first three convolutional modules also include a max pooling layer.

[0082] The system consists of four layers: a two-dimensional convolutional layer to extract battery aging information from the input data; a batch normalization layer to accelerate model training; a Rectified Linear Unit (ReLU) layer to increase nonlinearity between adjacent convolutional modules; and a max pooling layer to retain key battery aging information and prevent overfitting. The four convolutional modules sequentially extract battery aging information from the input data, and the extracted information is then passed through a fully connected layer and a regression layer to output corresponding estimation results.

[0083] The multi-task learning model includes a shared feature extraction backbone as well as capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM). NE The corresponding task output branch. The first two convolutional modules are used to extract capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM). NE The battery aging information can be shared between the modules, and the latter two convolutional modules are used to further extract battery aging information corresponding to different tasks. Each task's output branch outputs capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM) through a fully connected layer and a regression layer, respectively. NE The estimation results.

[0084] In one specific embodiment, the kernel size of the four two-dimensional convolutional layers is set to 5×3, and the number of filters is set to 8, 16, 32, and 64 respectively. The three task output branches are each configured with a fully connected layer of 16 nodes, and each outputs a scalar result through its corresponding regression layer. The network weights are initialized using the Glorot method, and the network biases are initialized to zero.

[0085] In the first training phase, capacity labels are used as supervision information, and the backbone of the shared convolutional feature extraction is pre-trained based on the mean square error between the capacity prediction results and the corresponding test capacity.

[0086]

[0087] in, This represents the loss function for the first training phase. N Indicates the number of training samples. C i This represents the test capacity corresponding to the i-th training sample. Indicates the first i The capacity prediction results corresponding to each training sample.

[0088] In the second training phase, after extracting the backbone from the shared convolutional features after pre-training, the connection capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM) are calculated. NE The corresponding task output branches calculate the mean squared error for each of the three tasks, and use the weighted sum of the losses from the three tasks as the joint loss function for multi-task joint training.

[0089]

[0090] in, L C , L LLI and L LAM These represent capacity loss task, lithium inventory loss (LLI) task, and negative electrode active material loss (LAM) task, respectively. NE The mean squared error loss corresponding to the task. , and These represent the loss weights for the corresponding tasks.

[0091] Since all three output labels were normalized before training, the three tasks used equal loss weights. Both the first and second training phases used the Adam optimizer and backpropagation to update the model parameters. The initial learning rate was set to 0.01, and it was reduced by 50% every 10 training epochs. The maximum training epoch was set to 100, and the minimum batch size was set to 1000. After completing the two-stage training, a model capable of simultaneously estimating capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM) based on short-time charging sequences was obtained. NE A multi-task learning model.

[0092] In one embodiment, in steps 8 to 10, short-time charging data of the lithium-ion battery to be diagnosed is obtained, and a trained multi-task learning model is used to perform electrode-cell multi-dimensional aging diagnosis.

[0093] Specifically, after the charging data of the lithium-ion battery to be diagnosed reaches the aging feature extraction voltage range, any continuous 100s charging sequence segment is selected from the charging data corresponding to that voltage range.

[0094] The aging feature extraction voltage range is 3.6V to 3.8V; any continuous 100s charging sequence segment refers to 100s of continuous charging data with an unfixed starting position within this voltage range, in order to adapt to the uncertainty of actual charging behavior.

[0095] Based on the charging voltage, charging current, and charging capacity in the charging sequence segment, an online diagnostic input matrix is ​​constructed using the same data processing method as in the model training phase. The online diagnostic input data is then input into the trained multi-task learning model to obtain the capacity estimation results, lithium inventory loss (LLI) estimation results, and negative electrode active material loss (LAMNE) estimation results for the lithium-ion battery to be diagnosed.

[0096] The capacity estimation results were used as the single-cell layer aging diagnosis results for the lithium-ion battery to be diagnosed, and the lithium inventory loss (LLI) estimation results and the negative electrode active material loss (LAM) results were used as the aging diagnosis results. NE The estimation results are used as the electrode layer aging diagnosis results for the lithium-ion battery to be diagnosed, thus obtaining the electrode-cell multidimensional aging diagnosis results for the lithium-ion battery to be diagnosed.

[0097] To verify the technical effectiveness of this application, aging data from lithium-ion battery samples were used for testing. Under average operating conditions, when a 100-second continuous VIQ charging sequence within the aging feature extraction voltage range was used as model input, the capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM) were measured. NEThe root mean square errors (RMSE) of the estimation results were 0.37%, 0.55%, and 2.10%, respectively, with corresponding mean absolute errors (MAE) of 0.21%, 0.36%, and 1.75%. This indicates that using a 100-second short VIQ charging sequence can achieve good joint diagnostic accuracy for capacity and electrode aging modes with lower data requirements, thus validating the rationality of using 100 seconds as the preset duration. Furthermore, compared with the single-task learning model using the same training set, test set, model parameters, and input parameters, the multi-task learning model reduced the average RMSE by 17.2% under optimal conditions and by 27.2% under worst conditions, indicating that the multi-task learning model can improve the accuracy of multidimensional aging diagnosis by sharing relevant information between different aging diagnosis tasks. The multi-task learning model has a size of 11.2 KB, which is about one-third of the total size of the three single-task learning models. The multi-task learning model has a running time of 0.035 s, while the total running time of the three single-task learning models is 0.105 s. This indicates that the multi-task learning model has good performance in terms of estimation accuracy, model size, and computational efficiency.

[0098] Employing the same inventive concept as the battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning, this embodiment also provides a corresponding battery multidimensional aging diagnosis system based on random short-time charging and multi-task learning, including: The sample data acquisition module is used to acquire charging operation data and full-cell open-circuit voltage curves of lithium-ion battery samples under different aging conditions, and to acquire the positive electrode open-circuit potential curves and negative electrode open-circuit potential curves corresponding to the lithium-ion battery samples; the charging operation data includes charging voltage, charging current, and sampling time corresponding to the charging voltage and charging current. The half-cell model building module is used to build a half-cell model based on the positive open-circuit potential curve and the negative open-circuit potential curve, so that the open-circuit voltage of the full cell is expressed as the potential difference between the positive open-circuit potential and the negative open-circuit potential. The electrode aging parameter identification module is used to optimize the parameters of the half-cell model using the Cuckoo Search algorithm, so that the open-circuit voltage curve reconstructed based on the half-cell model matches the open-circuit voltage curve of the full cell, and obtains the electrode aging parameters of the lithium-ion battery sample under different aging states. The aging diagnostic label generation module is used to obtain the test capacity of lithium-ion battery samples under different aging conditions, and calculate the lithium inventory loss (LLI) and negative electrode active material loss (LAM) based on the test capacity and electrode aging parameters. NE The test capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM) under different aging conditions were also measured. NE As diagnostic labels for aging, they constitute a set of diagnostic labels for aging. The aging feature extraction voltage range determination module is used to calculate the charging capacity based on the charging current and the corresponding sampling time, construct the capacity increment curve based on the charging voltage and the charging capacity, and determine the aging feature extraction voltage range of the lithium-ion battery sample according to the main peak position of the capacity increment curve under different aging states. The training sample construction module is used to select any continuous charging sequence segment of preset duration from the charging data corresponding to the voltage range of aging feature extraction, arrange the charging voltage, charging current and charging capacity in the charging sequence segment according to the sampling time sequence to form an input matrix, and use the input matrix and its corresponding aging diagnosis label to form a training sample. The multi-task learning model training module is used to train a multi-task learning model based on a convolutional neural network using a two-stage training approach. The first training stage pre-trains the shared convolutional feature extraction backbone based on capacity labels. The second training stage connects the capacity task output branch, the lithium inventory loss (LLI) task output branch, and the negative electrode active material loss (LAM) branch to the pre-trained shared convolutional feature extraction backbone. NE The task output branches are used for joint training based on the task losses corresponding to multiple task output branches; The online diagnostic module acquires any continuous, preset-duration charging sequence segment of the lithium-ion battery to be diagnosed within the aging feature extraction voltage range. Based on the charging voltage, charging current, and corresponding charging capacity in the charging data, it constructs an online diagnostic input matrix. This online diagnostic input data is then fed into a trained multi-task learning model to obtain capacity estimation results, lithium inventory loss (LLI) estimation results, and negative electrode active material loss (LAM) estimation results for the lithium-ion battery to be diagnosed. NE Estimation results; The diagnostic result generation module is used to use the capacity estimation results as the monomer layer aging diagnostic results, and to combine the lithium inventory loss (LLI) estimation results and the negative electrode active material loss (LAM) results. NE The estimation results are used as the electrode layer aging diagnosis results to obtain the electrode-cell multidimensional aging diagnosis results of the lithium-ion battery to be diagnosed.

[0099] The battery multidimensional aging diagnosis system based on random short-time charging and multi-task learning in this embodiment has the same inventive concept as the battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning described above. Therefore, the specific implementation of this device can be found in the embodiment section of the lithium-ion battery electrode-cell multidimensional aging diagnosis method based on random short-time charging sequence and multi-task learning described above, and its technical effects correspond to the technical effects of the above method, so they will not be repeated here.

[0100] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning as described above.

[0101] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0102] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-dimensional aging diagnosis method for batteries based on random short-time charging and multi-task learning, characterized in that, Includes the following steps: S1. Based on the charging operation data of lithium-ion battery samples, the open-circuit voltage curve of the full cell and the open-circuit potential curves of the positive and negative electrodes, a half-cell model is constructed to identify electrode aging parameters. The lithium inventory loss (LLI) and the negative electrode active material loss (LAMNE) are calculated in combination with the test capacity. A multi-dimensional aging diagnostic label set integrating the single-cell layer capacity and the electrode layer aging mode is constructed. S2. Construct a capacity increment curve based on charging operation data, determine the aging feature extraction voltage range according to its main peak position, and select continuous short-time charging sequence segments with non-fixed starting positions within the voltage range to construct an input matrix containing charging voltage, charging current and charging capacity. S3. Based on the input matrix and the multidimensional aging diagnosis label set, a two-stage training method is adopted to train the preset multi-task learning model. In the first training stage, the backbone of the shared convolutional feature is pre-trained based on the capacity label. In the second training stage, multi-task joint training is carried out based on the capacity, LLI and LAMNE labels. S4. Obtain continuous short-time charging sequence segments of the battery to be diagnosed within the aging feature extraction voltage range, input them into the trained multi-task learning model, and simultaneously output the capacity estimation results, LLI estimation results, and LAMNE estimation results, which are used as the aging diagnosis results of the single cell layer and the electrode layer, respectively, to realize the joint diagnosis of multi-dimensional aging of lithium-ion battery electrode-single cell.

2. The battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning according to claim 1, characterized in that, In S1, the open-circuit voltage of the full cell in the half-cell model is expressed by the following formula: in, This represents the open-circuit voltage of the full battery under the state of charge (SoC). Indicates the degree of lithiation of the positive electrode. The open circuit potential below, Indicates the degree of lithiation of the negative electrode The open-circuit potential below; and The value range is [0,1]. A value of 1 indicates that the corresponding electrode is fully lithiated, and a value of 0 indicates that the corresponding electrode is delithiated.

3. The battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning according to claim 1, characterized in that, In step S1, constructing a half-cell model to identify electrode aging parameters specifically includes: The parameters of the half-cell model are optimized using an optimization algorithm so that the open-circuit voltage curve reconstructed based on the half-cell model matches the open-circuit voltage curve of the full cell, thereby obtaining the electrode aging parameters of the lithium-ion battery sample under different aging states. Electrode aging parameters include the degree of lithiation of the positive electrode at the end of discharge. Parameters of lithiation degree of the positive electrode at the end of charging Parameters of lithiation degree of the negative electrode at the end of discharge The degree of lithiation of the negative electrode at the end of charging. ;in, EoD This indicates the end of the discharge process. EoC This indicates the end of charging; the optimization algorithm is the Cuckoo Search Algorithm; the Cuckoo Search Algorithm uses the fitting error between the open-circuit voltage curve reconstructed based on the half-cell model and the full-cell open-circuit voltage curve as the objective function. and Optimize the identification; based on the open-circuit potential difference relationship between the positive and negative electrodes satisfied by the minimum and maximum cutoff voltages of the lithium-ion battery samples, determine respectively and .

4. The battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning according to claim 1, characterized in that, In step S1, the calculation of lithium inventory loss (LLI) and negative electrode active material loss (LAMNE) based on the test capacity specifically includes: For any of the lithium-ion battery samples t The aging state, according to the first t Test capacity under aging conditions Parameters of lithiation degree of the positive electrode at the end of discharge. The degree of lithiation parameters of the positive electrode at the end of charging. Determine the first t Available capacity of the positive electrode under aging conditions ; According to the test capacity Parameters of lithiation degree of the negative electrode at the end of charging Parameters of lithiation degree of the negative electrode at the end of discharge Determine the first t Available capacity of the negative electrode under aging conditions ; According to the first t The available capacity of the positive electrode and the available capacity of the negative electrode under aging and initial conditions, as well as the degree of lithiation of the positive and negative electrodes under the end of charging conditions, are used to determine the lithium inventory loss (LLI). According to the first t Available capacity of the negative electrode under aging conditions and the available capacity of the negative electrode in the initial state Determine the loss of negative electrode active material LAM NE ; Where t represents the current aging state of the lithium-ion battery sample, and 0 represents the initial state of the lithium-ion battery sample.

5. The battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning according to claim 4, characterized in that, The available capacity of the positive electrode Available capacity of negative electrode Lithium inventory loss (LLI) and negative electrode active material loss (LAM) NE The following formulas are used to determine the values ​​respectively: in, This indicates that the lithium-ion battery sample is in the first... t Test capacity under aging conditions, and They represent the first t Available capacity of the positive electrode and available capacity of the negative electrode under aging conditions. and These represent the available capacity of the positive electrode and the available capacity of the negative electrode in the initial state, respectively.

6. The battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning according to claim 1, characterized in that, In step S2, the aging feature extraction voltage range is determined based on the main peak position of the capacity increment curve, specifically including: The charging capacity of the lithium-ion battery sample during the charging process is calculated based on the charging current and the corresponding sampling time; the capacity increment curve of the lithium-ion battery sample under different aging states is constructed based on the rate of change of the charging capacity relative to the charging voltage; the main peak position of the capacity increment curve under different aging states is determined, and the aging feature extraction voltage range is determined based on the main peak position.

7. The battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning according to claim 1, characterized in that, In S3, the multi-task learning model includes an input layer, four convolutional modules, and multiple task output branches; the four convolutional modules constitute a shared convolutional feature extraction backbone, and each convolutional module includes a two-dimensional convolutional layer, a batch normalization layer, and a linear rectified function ReLU layer. The first three convolutional modules also include a max pooling layer. In the first training phase, the shared convolutional feature extraction backbone connection capacity prediction branch is pre-trained; in the second training phase, the pre-trained shared convolutional feature extraction backbone connection capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAM) are used to extract the negative electrode active material loss (LAM). NE The system has multiple corresponding task output branches, each of which outputs its own estimation results through the corresponding fully connected layer and regression layer.

8. The battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning according to claim 7, characterized in that, In step S3, a two-stage training method is used to train the preset multi-task learning model, specifically including: In the first training phase, the capacity in the multidimensional aging diagnostic label set is used as the supervision label. The input matrix is ​​input into the shared convolutional feature extraction backbone, and the shared convolutional feature extraction backbone is pre-trained based on the mean square error between the capacity prediction result and the corresponding test capacity. In the second training phase, the capacity task output branch, the lithium inventory loss (LLI) task output branch, and the negative electrode active material loss (LAMNE) task output branch are connected to the pre-trained shared convolutional feature extraction backbone. The task losses for capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE) are calculated respectively, and the weighted sum of the above three task losses is used as the joint loss function to jointly train the shared convolutional feature extraction backbone and multiple task output branches.

9. The battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning according to claim 8, characterized in that, The first two convolutional modules are used to extract battery aging information that can be shared among capacity, lithium inventory loss (LLI), and negative electrode active material loss (LAMNE), while the latter two convolutional modules are used to extract battery aging information corresponding to different aging modes.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a battery multidimensional aging diagnosis method based on random short-time charging and multi-task learning as described in any one of claims 1-9.

Images