Retired lithium battery Q and SOH joint estimation method based on local charging segment

Abstract

The invention discloses a joint estimation method for Q and SOH of a decommissioned lithium battery based on a local charging segment, and the method comprises the steps: obtaining a relaxation stage voltage sequence and a charging stage statistical scalar feature under each cycle, carrying out the standardization, and carrying out the normalization processing of a cycle number; fusing the voltage sequence code and the statistical scalar feature code under each cycle to obtain a joint representation vector; constructing a Q and SOH joint estimation model; and respectively inputting the joint representation vector into the trained Q estimation branch and SOH estimation branch, and finally outputting a Q and SOH joint estimation result. According to the method, time sequence hybrid coding and information fusion are carried out on a unified normalized time axis, embeddable physical prior constraints are introduced into accumulative discharge capacity and SOH estimation branches respectively, training and evaluation are carried out with a weighted joint target, and the large-scale, rapid and reliable evaluation requirements under the retired battery sorting and echelon utilization scene are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium ion batteries, in particular to a retired lithium battery Q and SOH joint estimation method based on local charging segments. BACKGROUND

[0002] Lithium ion batteries are widely used in energy storage, electric vehicles, consumer electronics and other fields. With the increase of service life and cycle number, the performance of the battery gradually declines, showing capacity attenuation, internal resistance rise, etc. After sorting and evaluation, the retired lithium ion battery can be used for secondary use in the scene of gradient energy storage, and its economic value and safety depend on accurate, low-cost and large-scale implementation of health evaluation and grading method. Among them, the accumulated discharge capacity (Accumulated Discharge Capacity, Q for short) and the state of health (State of Health, SOH) are the key indicators to measure the remaining available performance and remaining life of the retired battery.

[0003] In the prior art, battery health evaluation mainly includes the following types: (1) estimation method based on mechanism or equivalent circuit model (ECM), which usually combines open circuit voltage-state of charge (OCV-SOC) characteristics, parameter identification and filter (such as Kalman / particle filter) to realize capacity and SOH inference, but it depends on complete working cycle, stable working condition and high calibration cost, and is sensitive to batch differences and inconsistency of working conditions of retired batteries; (2) statistical learning method based on feature engineering, such as incremental capacity (IC, dQ / dV), differential voltage (DV) feature extraction, and then using traditional regressor for estimation, but in the actual recycling link, complete charge-discharge curve is often unavailable or high cost and long time, and the feature is sensitive to noise and sampling strategy; (3) end-to-end data-driven deep learning method, which has achieved certain results in standardized working conditions, but generally needs long-term, full-condition data, and has insufficient robustness to incomplete data, and lacks physical consistency constraints, which is easy to produce non-physical prediction results (such as SOH out of bounds, abnormal capacity rise over time, etc.).

[0004] For the actual scene of retired batteries, there are also the following common problems: 1) Limited data acquisition: the recycling channel is complex, the working condition is not unified, and it is difficult to obtain complete and standardized full-cycle data; the data that can be stably obtained and controlled in cost is usually short-time and local charging or static segment data.

[0005] 2) Strong sample heterogeneity: there are significant differences in batch, historical use conditions and degradation path of retired batteries, and traditional methods with single working condition or single feature as the core have insufficient generalization.

[0006] 3) Difficulty in index coupling and joint estimation: Cumulative discharge capacity and SOH have a coupled relationship in physical and statistical sense. Independent modeling of each will lead to information fragmentation and uncontrollable error transmission.

[0007] 4) Physical consistency and interpretability requirements: Facing sorting and quality tracing, the method needs to meet the physical prior in terms of output range, boundedness, monotonicity, etc., to improve the credibility and feasibility of the prediction results.

[0008] 5) Computing and deployment overhead: Facing large-scale gradient utilization sorting, the evaluation method should rely on easily collected local segments as much as possible, have fast reasoning ability, and be compatible with existing production line data flow.

[0009] To solve the above problems, in recent years, some researches have tried to combine feature engineering and model optimization. For example, incremental capacity / differential voltage features are used with traditional regressors to improve capacity estimation stability; or end-to-end neural networks are introduced to learn from standardized charge and discharge data to improve SOH prediction accuracy; some works also explore self-supervised pre-training under incomplete conditions to alleviate the generalization problem caused by data scarcity and heterogeneity. However, these explorations are still not comprehensive enough, and have not yet taken into account the influence of retired battery degradation differences on training samples, the effective use of local segment data, and joint estimation under physical prior constraints.

[0010] In summary, the existing lithium-ion battery health management methods have the following key technical problems in practical application: 1. Lack of joint evaluation framework for cumulative discharge capacity and SOH. Existing methods mostly model and evaluate cumulative discharge capacity and SOH separately, lacking a mechanism for coupling and simultaneous evaluation under a unified data and index system, making it difficult to reflect the internal relationship and mutual constraints between the two in the degradation process.

[0011] 2. Evaluation difficulties caused by data incompleteness. Recycling and sorting often lack complete charge and discharge curves, and only short-term local charge / relaxation segments and a small amount of statistical information are available. Under the condition of not being able to obtain complete charge and discharge curves, it is difficult to reliably evaluate the cumulative discharge capacity (Q) and state of health (SOH) of retired lithium-ion batteries based on only local charge / relaxation segment data.

[0012] 3) Difficulty in collaborative use of heterogeneous data. Different batteries have large differences in initial and final cycles, aging progress, and usage conditions, leading to non-comparable time scales and severe feature distribution drift. At the same time, observations of different sources, different sampling strategies, and different granularities (short-term segment and statistical working condition coexist) are difficult to complement each other in the same evaluation framework, which is prone to estimation bias and instability due to uneven information.

[0013] 4. Lack of embeddable physical knowledge constraint mechanism. Existing evaluation methods are generally difficult to introduce prior constraints consistent with the degradation process in the learning framework, resulting in output deviating from physical laws, such as lack of structural constraints on the monotonically non-decreasing capacity with cycles, SOH in [0, 1] and decay trend, etc. At the same time, there is a lack of systematic inclusion of general prior tools that can express monotonicity and decay (such as integral spline type constraints that can be used for monotonic modeling, exponential type priors that can be used for decay modeling, etc.) in the evaluation process, which further affects the interpretability, stability and cross-scene consistency of the results. SUMMARY

[0014] The present application provides a retired lithium battery Q and SOH joint estimation method based on local charging segment, to solve the problems of insufficient data completeness, difficulty in coordinating heterogeneous information, lack of Q-SOH joint evaluation and physical consistency constraints, difficulty in cross-batch time alignment, and insufficient stability and interpretability in the prior art.

[0015] According to a first aspect, a retired lithium battery Q and SOH joint estimation method based on local charging segment is provided in an embodiment, the method comprising: Collecting lithium battery degradation data through charge-discharge cycle test, and aggregating the collected data according to cycle number to obtain relaxation phase voltage sequence and charging phase statistical scalar features under each cycle and standardize them, and normalizing the cycle number to obtain normalized time variable ; Mapping the relaxation phase voltage sequence under each cycle to the latent space and stacking the encoding through multi-scale time mixing and channel mixing, and fusing the corresponding charging phase statistical scalar feature encoding and voltage sequence encoding result to obtain a joint representation vector ; Constructing and training a Q and SOH joint estimation model, the Q and SOH joint estimation model comprising a Q estimation branch based on a monotonicity prior constraint and an SOH estimation branch based on an exponential constraint, the Q estimation branch being used to generate a Q estimation result based on a monotonicity base function layer and a residual correction layer , and the SOH estimation branch being used to generate an SOH estimation result based on exponential decay parameterization ; Inputting the joint representation vector to the trained Q estimation branch and SOH estimation branch respectively, and finally outputting the Q and SOH joint estimation result under the corresponding cycle .

[0016] Furthermore, the voltage sequence of the relaxation stage under each cycle is mapped to the latent space and encoded by multi-scale time mixing and channel mixing stacking. The corresponding statistical scalar feature encoding of the charging stage is then fused with the voltage sequence encoding result to obtain a joint representation vector. Specifically, it includes: Regarding the first The first cell in the first A cycle, the voltage sequence Mapping to the latent space dimension Stacking A TimeMixer basic block is used to realize multi-scale temporal mixing and channel mixing, including: the temporal mixing layer adopts a depthwise separable one-dimensional convolution, and the channel mixing layer is a position-wise two-layer MLP, both of which adopt a Pre-LN residual structure; Statistical scalar characteristics Using a two-layer MLP mapping to The joint representation vector is obtained by adding it element-by-element to the voltage sequence encoding vector. .

[0017] Furthermore, the Q-estimation branch is used based on the monotonic basis function layer and the residual correction layer in The above generates Q-estimation results, specifically including: a. Constructing the baseline trend term: In Construct a set of monotonic basis functions for the integral spline I-spline. The baseline trend term is obtained by combining the constructed basis functions with non-negative coefficients, as follows:

[0018] in, To obtain the baseline trend term, K is the number of monotonic basis functions of the I-spline; Let i be the monotonic basis function of the i-th I-spline; This is the baseline bias term for capacity estimation; The global reference coefficient is shared with all battery cells and is used to control the contribution of each monotonic basis function to the reference trend term. It is constrained to be non-negative to ensure that the reference trend term is monotonically non-decreasing over time. b. Constructing residual correction terms: To express the individual differences of different cells / operating conditions, a joint representation vector is introduced. The relevant residual correction terms are as follows:

[0019] in, For the introduced residual correction term, This is the joint representation vector obtained from multimodal temporal coding; The weight matrix for generating the residual coefficient layer; This is the corresponding bias vector; It is an element-wise non-negative activation function, and its output vector is... The residual basis function coefficients, which are related to the current sample, are constrained to be non-negative to preserve the residual terms. Monotonicity with respect to time; c. Add the baseline trend term and the residual correction term to obtain the final capacity estimate, as shown in the following formula:

[0020] in, Indicates normalization time The cumulative discharge capacity estimation results are as follows; The total coefficients of the i-th monotone basis function are guaranteed to be nonnegative. It monotonically decreases over time.

[0021] Furthermore, the SOH estimation branch is used to parameterize the exponential decay in... The above generates SOH estimation results, specifically including: a. Parameter generation layer, used to generate joint representation vectors Generate by performing linear transformation Group index component parameters, The unconstrained representation of the group exponential component parameters is as follows:

[0022] Where N is the number of exponential components; , , These represent the initial contribution coefficient, decay coefficient, and steady-state plateau term of each index component, respectively. The steady-state plateau term is used to characterize the residual health level that the index component approaches under long-term service conditions. To obtain from the joint representation vector Generate a weight matrix for three sets of parameters. For the corresponding bias vector; Differentiable structural constraints are then applied to the generated parameters to obtain effective parameters for synthesis:

[0023] in, For the Sigmoid function, , It is a very small constant. This represents the upper limit of the platform's amplitude. Constraints As an initial contribution, to avoid magnitude explosion; Strictly negative, ensuring that the exponential component is a decaying term; Limiting to a small platform enhances noise immunity and prevents abnormal tail-end rise; The unconstrained intermediate parameters output by the network are transformed to obtain the effective parameters. ; For index of exponential components, The number of exponential components; b. Exponential hybrid synthesis layer, used to form an uncompressed SOH curve by superimposing multiple exponents:

[0024] in, The uncompressed SOH curve is obtained by superimposing multiple exponential components; c. Bounded layer, used to ensure that the output is within bounded space. Within the health range, apply Sigmoid compression to the uncompressed SOH curve:

[0025] in, After Sigmoid mapping, it is located at Health estimation results within the interval, This is the Sigmoid function.

[0026] Furthermore, the joint estimation model for Q and SOH is trained, specifically including: The model is trained using joint loss, and regularization is introduced to obtain the optimal parameters. The total loss function is as follows:

[0027] in, Weights for the Q-estimated loss term and the SOH-estimated loss term; These are the regularization coefficients, and These represent the mean square error loss of cumulative discharge capacity and SOH, respectively. The L2 norm regularization term for the residual coefficients of the Q-branch. This is the weight decay term applied to all trainable parameters; Q estimates the loss term as follows:

[0028] in, Estimate the loss term for Q. This is the current training mini-batch sample index set. Batch size; This is the tag for the actual cumulative discharge capacity of the nth cell in the jth cycle. The corresponding model prediction value; The estimated loss term for SOH is:

[0029] in, Estimate the loss term for SOH. and These are the actual SOH label and the model prediction value for the nth cell in the jth cycle, respectively. Q estimates the regularization term for the branch residual coefficients:

[0030] in, Estimate the regularization term for the branch residual coefficients of Q; For the residual coefficient vector generated by the joint representation vector in the Q branch, apply L2 regularization to suppress overfitting and smooth the capacity curve; Apply a weight decay term to all trainable parameters:

[0031] in, This indicates the application of a weight decay term; This represents the set of all trainable parameters of the model, to which a L2 regularization is applied to improve the model's generalization ability.

[0032] According to a second aspect, one embodiment provides a joint estimation system for Q and SOH of retired lithium batteries based on partial charging segments, the system comprising: The data acquisition and processing module is used to collect lithium battery degradation data through charge-discharge cycle tests, aggregate the collected data by cycle number, obtain the relaxation stage voltage sequence and charging stage statistical scalar characteristics for each cycle, standardize them, and normalize the cycle number to obtain a normalized time variable. ; The multimodal temporal coding module maps the voltage sequence of the relaxation stage in each cycle to the latent space and performs multi-scale temporal mixing and channel mixing stacking encoding. It then fuses the corresponding statistical scalar feature encoding of the charging stage with the voltage sequence encoding result to obtain a joint representation vector. ; The joint estimation model construction and training module is used to construct and train a joint estimation model of Q and SOH. This model includes a Q estimation branch based on monotonic prior constraints and a SOH estimation branch based on exponential constraints. The Q estimation branch is used to perform estimation based on the monotonic basis function layer and the residual correction layer. The Q-estimation result is generated above, and the SOH estimation branch is used to generate the Q-estimation result based on the exponential decay parameterization. The SOH estimation results are generated above; The joint estimation module is used to perform joint representation vector estimation. The inputs are fed into the Q-estimation branch and the SOH-estimation branch during training, and the corresponding loop is output at the end. The results of the joint estimation of Q and SOH are as follows.

[0033] According to a third aspect, one embodiment provides an electronic device, the device comprising: a processor and a memory; The memory is used to store one or more program instructions; The processor is configured to run one or more program instructions to perform the steps of a joint estimation method for Q and SOH of retired lithium batteries based on local charging segments as described in any of the preceding claims.

[0034] According to a fourth aspect, one embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a joint estimation method for Q and SOH of a retired lithium battery based on a partial charging segment as described in any of the preceding claims.

[0035] This invention provides a joint estimation method for Q and SOH of retired lithium batteries based on partial charging segments. It employs multimodal modeling based on partial charging / relaxation segments and statistical operating conditions, time normalization alignment, a dual regression head with physical prior constraints, and a weighted joint training framework, which has the following beneficial effects: 1) Low data cost and adaptable to sorting scenarios. Relying only on "partial charging / relaxation segments + a small number of statistical features", Q and SOH can be estimated simultaneously without the need for complete charge and discharge curves. Therefore, it significantly reduces the acquisition cost and evaluation time, making it suitable for batch applications in recycling / sorting.

[0036] 2) Improved consistency and robustness across batches. By normalizing the cycle number to a unified time axis and performing multimodal timing coding, the distribution differences and sampling noise caused by different cells / operating conditions are reduced, thus maintaining stable and comparable output across batches and heterogeneous data.

[0037] 3) Physically consistent and interpretable. The Q branch introduces the monotonic prior of the I-spline, and the SOH branch adopts exponential constraints and bounded output, structurally guaranteeing that "Q is monotonically non-decreasing over time and SOH is bounded and non-increasing", thus avoiding non-physical predictions and facilitating threshold setting and traceability.

[0038] 4) Balancing accuracy and stability. Weighted joint training using Q and SOH is employed, with gradients co-constrained within a shared representation space. This reduces the fragmentation of single tasks and error propagation, thus improving the overall accuracy and stability of both metrics. Attached Figure Description

[0039] Figure 1A flowchart illustrating a method for jointly estimating Q and SOH of a retired lithium battery based on a local charging segment, as provided in one embodiment of the present invention; Figure 2 This is an overall architecture diagram of a joint estimation method for Q and SOH of retired lithium batteries based on local charging segments, provided as an embodiment of the present invention. Figure 3 This is a schematic diagram of the logic structure of a joint estimation system for Q and SOH of a retired lithium battery based on a partial charging segment, provided as an embodiment of the present invention. Detailed Implementation

[0040] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0041] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0042] The first embodiment of this invention provides a joint estimation method for Q and SOH of retired lithium batteries based on partial charging segments. The following is a combination of... Figure 1 and Figure 2 Please provide a detailed explanation.

[0043] like Figure 1 As shown, in step S100, lithium battery degradation data is collected through charge-discharge cycle tests, and the collected data is aggregated according to the cycle number. The voltage sequence of the relaxation stage and the statistical scalar features of the charging stage under each cycle are obtained and standardized. The cycle number is normalized to obtain the normalized time variable t_norm∈[0,1].

[0044] The above steps specifically include: S110, Data Acquisition and Preprocessing: A lithium-ion battery degradation dataset containing 24 Panasonic 18650 cells was constructed. All tests were conducted in a 25°C constant temperature chamber. The experimental platform consisted of an MHW-200 constant temperature chamber, a CT-4008Tn-5V12A-S1 test unit, a CT-ZWJ-4'S-T-1U control unit, a switching circuit, and a NewareBTS~8.0.0 host computer. The collected raw data was aggregated by cycle number to form a single-cycle relaxation segment voltage vector, while statistical operating condition features related to the charging stage were introduced. Subsequently, the voltage segment sequence and scalar operating condition features were standardized (mean-variance standardization) to eliminate dimensional differences and numerical scale inconsistencies. Training / validation / testing were performed by battery ID. Only the training set was used to fit the standardized parameters, and the same parameters were applied to the validation and test sets to avoid data leakage. This step utilizes existing mature technologies to ensure the quality of input data, which is the foundation and necessary prerequisite for subsequent analysis.

[0045] S120 addresses the issue of inconsistent cycle ranges and aging progress among different batteries by performing Min–Max normalization on the cycle number to obtain a normalized time variable. This variable serves as an input / index to the model's degradation process and is used in subsequent calculations of prior constraints and the regression head, thereby modeling the degradation trajectory on a uniform time scale.

[0046] like Figure 1 As shown, in step S200, the voltage sequence of the relaxation stage under each cycle is mapped to the latent space and encoded by multi-scale time mixing and channel mixing stacking. The corresponding statistical scalar feature encoding of the charging stage is fused with the voltage sequence encoding result to obtain the joint representation vector h.

[0047] Specifically, to support the synchronous evaluation of retired lithium-ion batteries under incomplete data conditions, the multimodal time-series coding module designed in this invention is designed to address two heterogeneous inputs: "local charging / relaxation segment voltage sequence" and "statistical operating condition scalar". It realizes the modeling of multi-scale time correlation and the effective fusion of cross-modal information in a unified latent space.

[0048] This module employs a lightweight temporal-channel hybrid stacked structure to perform multi-kernel-scale depthwise separable convolutional extraction on short temporal segments, and uses low-order feedforward mapping to perform semantic alignment of scalar conditions. Training is stabilized through residual and normalization strategies, ultimately outputting a single joint representation vector. This characterization maintains controllable parameter / computational overhead while ensuring robustness to noise and sampling discrepancies. It is also decoupled from subsequent capacity monotonic priors and SOH decay priors, thus providing informative and structurally clear high-level semantic input for physical consistency regression based on a normalized time axis. The specific steps are as follows: S210 constructs two types of inputs for each battery and each cycle: 1) Timing Segment Input The relaxation segment voltage sequence is denoted as: (1) in For the segment length, is the dimension of the voltage sequence during the relaxation segment.

[0049] 2) Scalar operating condition input The statistical characteristics (constant voltage charging time, 3.3–3.6V charging time) are denoted as: (2) After undergoing independent standardization processing, it enters the encoding module.

[0050] S220 linearly projects the time-by-time features of the sequence onto a unified latent space dimension. : (3) Therefore, we obtain .

[0051] S230 introduces the TimeMixer basic block to achieve multi-scale time mixing and channel mixing. The TimeMixer basic block consists of two residual layers: time mixing (T-Mix) and channel mixing (C-Mix), both using a Pre-LN residual structure. Let the first... The input for each block is .

[0052] 1) First, prenormalize the input: (4) in, For each channel, use LayerNorm.

[0053] 2) Perform temporal gradient mixing on the normalized input, including Depthwise multi-scale convolution + adaptive weight fusion.

[0054] Will View as The one-dimensional sequence of channels is first transformed to adapt to the convolution implementation: (5) For each preset convolution kernel ( Apply channel-wise depthwise separable convolution (groups= ): (6) in Activate SiLU ( , (For Sigmoid).

[0055] Introducing learnable kernel selection weights After softmax, we get: (7) The features at each scale are weighted and summed, and then... Channel-wise linear blending (equivalent to pointwiseconv): (8) Then apply random dropout to suppress overfitting. Finally, backinject using residuals: (9) 3) The relaxation segment voltage sequence output after time mixing is channel-mixed with the scalar characteristics using a time-by-time feedforward network.

[0056] right Pre-normalization again: (10) At each time step Above, an expansion coefficient of is applied. Position-wise feedforward network (containing two MLP layers + activation): (11) in Activate GELU , The results were combined by time to obtain The results of this block are output as residuals: (12) 4) Stack the above basic blocks. (Code Example) ),get The temporal dimension is converged (time averaging is used in the code) to form sequence semantics: (13) S240 performs scalar load case branching and multimodal fusion. (The scalar load case is then used for...) Mapped to the same latent space via a lightweight fully connected network: (14) in Take ReLU.

[0057] The two feature vectors are fused element-wise in the vector space to obtain the multimodal encoding vector: (15) like Figure 1 As shown, in step S300, a joint estimation model of Q and SOH is constructed and trained. The joint estimation model of Q and SOH includes a Q estimation branch based on monotonic prior constraints and an SOH estimation branch based on exponential constraints. The Q estimation branch is used to generate Q estimation results on t_norm based on monotonic basis function layers and residual correction layers. The SOH estimation branch is used to generate SOH estimation results on t_norm based on exponential decay parameterization.

[0058] The above steps specifically include: S310, Q estimation under monotonic prior constraints: To ensure that the estimation of cumulative discharge capacity conforms to the physical law that the cumulative amount is monotonically non-decreasing with the degradation process on a unified time axis, the joint representation vector obtained from the multimodal time-series encoding in the previous section is used. This invention effectively transforms the output into an interpretable curve, achieving a normalized time... A Q-estimation module with monotonic prior constraints is constructed: the global baseline trend is characterized by monotonic basis functions formed by integral splines (I-splines), and the monotonicity of the curve with respect to time is structurally guaranteed by the combination of nonnegative coefficients; at the same time, a Q-estimation module with monotonic prior constraints is introduced. The data-driven adaptive residual term expresses individual differences and subtle variations. The superposition of these two terms yields a result that satisfies both physical consistency and individualized characterization. function.

[0059] This module is decoupled from the aforementioned encoder: the encoder focuses on extracting degenerate semantics from local charging / relaxation segments and operating condition scalars, while the Q estimation module applies monotonic constraints and residual corrections at a unified time scale, thereby achieving interpretable, stable and scalable capacity evolution estimation at any time.

[0060] ① Normalize time and objectives. Use normalized time variables. This represents the degradation process (obtained by normalizing the cycle number Min–Max). , as recorded below The goal is to estimate the cumulative discharge capacity function on this unified time axis: (16) It also meets the dual requirements of being monotonically non-decreasing over time (physical consistency) and being able to characterize individual differences (data-driven).

[0061] ②I-Spline monotonic base and baseline trend term. In Construct a set of I-spline monotonic basis functions The definition of an I-spline can be derived from that of an M-spline. The integral is given as follows: (17) in, The number of monotonic basis functions of the I-spline; For normalized time variables; It is a dummy variable in the integral; For definition in The first There are M-spline basis functions that satisfy... It also provides localized support; This indicates that the same basis function is integrated with the independent variable as... The value at time; For the reason After integration and within the interval The first normalization obtained One I-spline monotonic basis function; therefore Therefore right It is monotonically non-decreasing; and because the denominator is a constant and normalized, it guarantees... and .

[0062] The baseline is obtained by combining these basis functions with non-negative coefficients: (18) ③ Data-driven residual correction term. To represent the individual differences between different cells / operating conditions, a multimodal encoding vector is introduced. Relevant residual correction terms: (19) in Non-negative activation (used in the code) ),ensure Similarly, we have: (20) therefore It is also monotonic and non-decreasing. The role of the residual term is to adaptively adjust the curve shape and slope for each sample based on the input segment and operating condition characteristics without destroying the monotonicity.

[0063] ④ Final capacity function and boundary properties. The final capacity estimate is obtained by adding the baseline trend and the residuals: (twenty one) because and ,but: (twenty two) thereby exist It strictly satisfies the physical consistency requirement of being monotonically non-decreasing. Furthermore, by... We can obtain: (twenty three) Right now Control the initial value offset, The total cumulative increment is determined, and both together determine the starting and ending levels of the curve; while different Shape adjustment is driven by sample input, achieving individualized but conserved curve fitting.

[0064] S320, SOH estimation under exponential constraints: Given that existing research and engineering experience show that the state of health (SOH) degradation of lithium-ion batteries over their entire lifespan can be robustly approximated by an exponential function, this invention normalizes the time... The above introduces an exponentially constrained SOH estimation module: it takes the joint representation vector obtained from the multimodal temporal encoding in the previous section... The model is mapped to several sets of exponential components of "initial contribution - decay rate - plateau term", and differentiability constraints are imposed on the parameters to ensure decay properties and numerical stability. Subsequently, the synthesis results are bounded and compressed, with the output at... The interval and the overall monotonicity does not increase over time. This design is based on the prior rationality of the exponential family, taking into account interpretability and engineering feasibility, and is decoupled from the aforementioned encoder, thereby forming a health evolution curve that conforms to physical laws on a unified time scale. Specifically: First, on the normalized timeline Above, estimate the health state function: (twenty four) It is required that the vector is bounded and conforms to a degradation pattern (usually exhibiting no increase or slow decrease over time / cycles), and can reflect the differences between different samples. The module input is a multimodal temporal encoding vector. (Obtained by fusing partial charging / relaxation segments with operating condition scalars). The output is in the state of... Interval prediction .

[0065] S321, Parameter Generation Layer Depend on Generate through linear transformation Unconstrained representation of grouped exponential component parameters: (25) By applying differentiable structural constraints to the above parameters, the effective parameters for synthesis are obtained: (26) in It is Sigmoid. , It is a minimal constant (to prevent) ), The upper limit of the platform amplitude (the example can be taken as follows) ). Constraints As an initial contribution, to avoid magnitude explosion; Strictly negative, ensuring that the exponential component is a decaying term; Limiting to a small platform enhances noise immunity and prevents abnormal rise in the tail section.

[0066] S322, Exponential Hybrid Synthesis Layer An uncompressed SOH curve is formed by superimposing multiple exponents: (27) The computational complexity of this form is... It is independent of time step, which facilitates batch inference.

[0067] Endpoint behavior: (28) because Therefore, .

[0068] S323, Bounded Layer To ensure the output is in Within the health range, apply Sigmoid compression to the uncompressed output: (29) If necessary, adjustments can be made in the numerical implementation. Make a very small cut (e.g., trim to) This improves the numerical stability of log loss or probabilistic post-processing (without changing the physical meaning).

[0069] To verify the monotonicity and physical consistency of the method, we can obtain the following from the above equation: (30) in , and .

[0070] According to the chain rule: (31) therefore, exist The overall structure is monotonically non-increasing and bounded, satisfying the degradation prior of battery health. With... Increase, each exponential term , Xiang You The small platform that was decided gradually converged, depicting the slow decay at the end.

[0071] ⑤ Final SOH Prediction Output Let the first The first cell in the first The normalized time for each loop is: (32) Its corresponding multimodal temporal coding vector is: (33) The exponential parameter triplet is obtained from the parameter generation layer: (34) The uncompressed combined output of this sample in this loop is: (35) The final health status prediction is obtained after range compression: (36) To enhance numerical stability, a very small truncation can be used in the implementation: (37) This truncation does not change the physical meaning; it is only used to avoid numerical problems in subsequent statistical / log calculations.

[0072] For batch size The sample, its time vector Parameter tensor Then the vectorized calculation is: (38) During training and evaluation, Directly corresponding real tags Calculate the mean squared error and sum it with the Q-branch loss to complete the end-to-end joint optimization.

[0073] S330, Joint Training Objective: Used to simultaneously optimize cumulative discharge capacity within a unified framework. With health status Two outputs enable the encoder to converge in tandem with the two regression heads.

[0074] Let the training set samples be indexed by cell. With circular index Organization, given: normalized time Multimodal coding vector ; Real Labels , ;predict (See the Q estimation module under monotonic prior constraints). (See the SOH estimation steps under exponential constraints).

[0075] The parameter set is denoted as: (39) in, This refers to the complete set of trainable parameters for the joint model. , , These represent the multi-mode timing encoder and the cumulative discharge capacity, respectively. The parameter subset of the estimation head and the SOH estimation head; For a small batch (Batch size) =B), both tasks use mean squared error (MSE): (40) (41) To suppress overfitting and maintain a smooth / stable curve, two types of regularization terms are added: 1) Q-head residual coefficient regularization. In the Q module, [the following is used]: Generated residual coefficients Apply L2 norm regularization: (42) 2) Apply weight decay (L2) to all trainable parameters: (43) The joint training objective is a weighted sum of the task loss and the regularization term: (44) in, The weights are assigned to the two tasks. represents the regularization coefficient. Weights and coefficients are selected on the validation set to balance the accuracy of cumulative discharge capacity prediction, SOH prediction accuracy, curve smoothness, and generalization.

[0076] Parameter update follows Its gradient is fed back to the encoder parameters through both task branches. To achieve joint learning of shared representations: This enables the encoder to learn segment semantics related to capacity accumulation; This enables the encoder to learn segment semantics related to health decay; the combined effect of these two factors forms a complementary representation of local segments and operating conditions in the shared latent space, reducing single-task bias and error propagation.

[0077] like Figure 3 As shown, in step S400, the joint representation vector is... The inputs are fed into the Q estimation branch and the SOH estimation branch during training, respectively, and the final output is the joint estimation result of Q and SOH under the corresponding loop t_norm.

[0078] The above steps specifically include: Inference and Output: During the deployment phase, after loading the optimal parameters and standardized / time-normalized statistics obtained from training, the local charging / relaxation segments and statistical operating condition characteristics of each cell per cycle are standardized in accordance with the training, and the cycle number is mapped to the normalized time. The joint representation vector is obtained through multimodal temporal coding. Input them into the "Q-estimator of monotonic priors" and the "SOH-estimator of exponential constraints" respectively, and obtain the results. and (as for) If proportional normalization is used, the output is proportionally restored to the physical quantity. The system's structured results and basic evaluation indicators (such as MAE, RMSE, etc.) are standardized according to the cell and overall output. (and corresponding curve / diagonal comparison charts.)

[0079] Based on the methods in the above embodiments, the present invention also provides specific operational examples as follows: (1) Data and Preprocessing Degradation data containing 24 battery cells (18650 type) were selected and collected under constant temperature conditions of 25℃. Timing input was constructed using a "single-cycle relaxation period": , among which examples (7-point relaxation voltage vector); construct scalar inputs using charging stage statistics: (The example includes one dimension each for "constant voltage charging time" and "3.3–3.6V charging time," for a total of two dimensions). The training / validation / test sets are divided by battery ID; two sets of normalizers (one for sequences and one for scalars) are fitted using the training set and used for validation / testing to avoid data leakage. Cycle number. Mapped to a unified time using Min–Max transformation .

[0080] (2) Multimodal timing coding Will Mapping to the latent space dimension (example: ), stack A basic TimeMixer block (e.g.: The temporal hybrid sublayer employs depthwise separable one-dimensional convolution and a multi-kernel ensemble. Adaptive weighting with softmax; the channel mixing sublayer is a position-wise two-layer MLP (expansion coefficient) (Activate GELU), both sub-layers use Pre-LN residual structures and dropout is set (e.g., 0.1). For scalar features Two-layer MLP (ReLU) mapping to The joint representation is obtained by adding it element-wise to the sequence encoding vector. .

[0081] (3) Q estimation exist Construct I-spline monotone bases (example: (cubic splines). Definition (45) in This serves as the global baseline coefficient. Depend on Generated after softplus nonnegation. (From...) With coefficients that are not negative, Follow Monotonically non-decreasing. If the capacity target is constant during training... (Example: 3.5Ah) Normalized, then the output terminal is... Restore physical units.

[0082] (4) SOH estimation generate Group parameters (example: ), and apply differentiability constraints: (46) This results in the uncompressed output. ,final .because , Follow Monotonous and does not increase.

[0083] (5) Joint training objectives By batch size MSE as the loss of two tasks: (47) And add Q-head residual coefficient regularization Weight decay The total loss is: (48) Optimizer Adam (e.g., learning rate) Weight decay Batch size 128, maximum 200 epochs, early stop (patience 10–15), AMP can be enabled. Recommended weights. , , .

[0084] 6) Reasoning and Derivation Loading optimal weights and standardized / time-normalized statistics, outputting the new battery cell cycle by cycle. Derive the MAE and RMSE for each cell and the overall data. With curve / diagonal charts, and generate CSV reports (including...) (and other fields).

[0085] Corresponding to the aforementioned method for jointly estimating Q and SOH of retired lithium batteries based on partial charging segments, this invention also discloses a system for jointly estimating Q and SOH of retired lithium batteries based on partial charging segments, such as... Figure 3 As shown, it specifically includes: The data acquisition and processing module is used to collect lithium battery degradation data through charge-discharge cycle tests, aggregate the collected data by cycle number, obtain the relaxation stage voltage sequence and charging stage statistical scalar characteristics for each cycle, standardize them, and normalize the cycle number to obtain a normalized time variable. ; The multimodal temporal coding module maps the voltage sequence of the relaxation stage in each cycle to the latent space and performs multi-scale temporal mixing and channel mixing stacking encoding. It then fuses the corresponding statistical scalar feature encoding of the charging stage with the voltage sequence encoding result to obtain a joint representation vector. ; The joint estimation model construction and training module is used to construct and train a joint estimation model of Q and SOH. This model includes a Q estimation branch based on monotonic prior constraints and a SOH estimation branch based on exponential constraints. The Q estimation branch is used to perform estimation based on the monotonic basis function layer and the residual correction layer. The Q-estimation result is generated above, and the SOH estimation branch is used to generate the Q-estimation result based on the exponential decay parameterization. The SOH estimation results are generated above; The joint estimation module is used to perform joint representation vector estimation. The inputs are fed into the Q-estimation branch and the SOH-estimation branch during training, and the corresponding loop is output at the end. The results of the joint estimation of Q and SOH are as follows.

[0086] It should be noted that for a detailed description of the joint estimation system of Q and SOH of retired lithium batteries based on partial charging segments provided in the embodiments of the present invention, please refer to the relevant description of the joint estimation method of Q and SOH of retired lithium batteries based on partial charging segments provided in the embodiments of the present invention, which will not be repeated here.

[0087] In addition, embodiments of the present invention also provide an electronic device, the device comprising: a processor and a memory; the memory for storing one or more program instructions; the processor for executing one or more program instructions to perform the steps of a joint estimation method for Q and SOH of a retired lithium battery based on a partial charging segment as described in any of the preceding embodiments.

[0088] It should be noted that for a detailed description of an electronic device provided in the embodiments of the present invention, please refer to the relevant description of a joint estimation method for Q and SOH of retired lithium batteries based on partial charging segments provided in the embodiments of this application, which will not be repeated here.

[0089] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a joint estimation method for Q and SOH of a retired lithium battery based on a local charging segment as described in any of the preceding claims.

[0090] It should be noted that for a detailed description of the computer-readable storage medium provided in the embodiments of the present invention, please refer to the relevant description of the joint estimation method of Q and SOH of retired lithium batteries based on partial charging segments provided in the embodiments of this application, which will not be repeated here.

[0091] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved. The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A method for jointly estimating Q and SOH of retired lithium batteries based on partial charging segments, characterized in that, The method includes: Degradation data of lithium batteries were collected through charge-discharge cycle tests. The collected data were aggregated according to cycle number to obtain the voltage sequence of the relaxation stage and the statistical scalar characteristics of the charging stage under each cycle. These were then standardized, and the cycle number was normalized to obtain the normalized time variable. ; The voltage sequence of the relaxation stage in each cycle is mapped to the latent space and encoded using multi-scale time mixing and channel mixing stacking. The corresponding statistical scalar feature encoding of the charging stage is then fused with the voltage sequence encoding result to obtain a joint representation vector. ; A joint estimation model for Q and SOH is constructed and trained. This model includes a Q estimation branch based on monotonic prior constraints and an SOH estimation branch based on exponential constraints. The Q estimation branch is used to estimate the SOH based on the monotonic basis function layer and the residual correction layer. The Q-estimation result is generated above, and the SOH estimation branch is used to generate the Q-estimation result based on the exponential decay parameterization. The SOH estimation results are generated above; Joint representation vector The inputs are fed into the Q-estimation branch and the SOH-estimation branch during training, and the corresponding loop is output at the end. The results of the joint estimation of Q and SOH are as follows.

2. The method for jointly estimating Q and SOH of retired lithium batteries based on partial charging segments as described in claim 1, characterized in that, The voltage sequence of the relaxation stage in each cycle is mapped to the latent space and encoded using multi-scale time mixing and channel mixing stacking. The corresponding statistical scalar feature encoding of the charging stage is then fused with the voltage sequence encoding result to obtain a joint representation vector. Specifically, it includes: Regarding the first The first cell in the first A cycle, the voltage sequence Mapping to the latent space dimension Stacking A TimeMixer basic block is used to realize multi-scale temporal mixing and channel mixing, including: the temporal mixing layer adopts a depthwise separable one-dimensional convolution, and the channel mixing layer is a position-wise two-layer MLP, both of which adopt a Pre-LN residual structure; Statistical scalar characteristics Using a two-layer MLP mapping to The joint representation vector is obtained by adding it element-by-element to the voltage sequence encoding vector. .

3. The method for jointly estimating Q and SOH of retired lithium batteries based on partial charging segments as described in claim 1, characterized in that, The Q-estimation branch is used based on the monotonic basis function layer and the residual correction layer. The above generates Q-estimation results, specifically including: a. Constructing the baseline trend term: In Construct a set of monotonic basis functions for the integral spline I-spline. The baseline trend term is obtained by combining the constructed basis functions with non-negative coefficients, as follows: in, To obtain the baseline trend term, K is the number of monotonic basis functions of the I-spline; Let i be the monotonic basis function of the i-th I-spline; This is the baseline bias term for capacity estimation; The global reference coefficient is shared with all battery cells and is used to control the contribution of each monotonic basis function to the reference trend term. It is constrained to be non-negative to ensure that the reference trend term is monotonically non-decreasing over time. b. Constructing residual correction terms: To express the individual differences of different cells / operating conditions, a joint representation vector is introduced. The relevant residual correction terms are as follows: in, For the introduced residual correction term, This is the joint representation vector obtained from multimodal temporal coding; The weight matrix for generating the residual coefficient layer; This is the corresponding bias vector; It is an element-wise non-negative activation function, and its output vector is... The residual basis function coefficients, which are related to the current sample, are constrained to be non-negative to preserve the residual terms. Monotonicity with respect to time; c. Add the baseline trend term and the residual correction term to obtain the final capacity estimate, as shown in the following formula: in, Indicates normalization time The cumulative discharge capacity estimation results are as follows; The total coefficients of the i-th monotone basis function are guaranteed to be nonnegative. It monotonically decreases over time.

4. The method for jointly estimating Q and SOH of retired lithium batteries based on partial charging segments as described in claim 1, characterized in that, The SOH estimation branch is used to perform exponential decay parameterization in... The above generates SOH estimation results, specifically including: a. Parameter generation layer, used to generate joint representation vectors Generate by performing linear transformation Group index component parameters, The unconstrained representation of the group exponential component parameters is as follows: Where N is the number of exponential components; , , These represent the initial contribution coefficient, decay coefficient, and steady-state plateau term of each index component, respectively. The steady-state plateau term is used to characterize the residual health level that the index component approaches under long-term service conditions. To obtain from the joint representation vector Generate a weight matrix for three sets of parameters. For the corresponding bias vector; Differentiable structural constraints are then applied to the generated parameters to obtain effective parameters for synthesis: in, For the Sigmoid function, , It is a very small constant. This represents the upper limit of the platform's amplitude. Constraints As an initial contribution, to avoid magnitude explosion; Strictly negative, ensuring that the exponential component is a decaying term; Limiting to a small platform enhances noise immunity and prevents abnormal tail-end rise; The unconstrained intermediate parameters output by the network are transformed to obtain the effective parameters. ; For index of exponential components, The number of exponential components; b. Exponential hybrid synthesis layer, used to form an uncompressed SOH curve by superimposing multiple exponents: in, The uncompressed SOH curve is obtained by superimposing multiple exponential components; c. Bounded layer, used to ensure that the output is within bounded space. Within the health range, apply Sigmoid compression to the uncompressed SOH curve: in, After Sigmoid mapping, it is located at Health estimation results within the interval, This is the Sigmoid function.

5. The method for jointly estimating Q and SOH of retired lithium batteries based on partial charging segments as described in claim 1, characterized in that, Training the joint estimation model of Q and SOH specifically includes: The model is trained using joint loss, and regularization is introduced to obtain the optimal parameters. The total loss function is as follows: in, Weights for the Q-estimated loss term and the SOH-estimated loss term; These are the regularization coefficients, and These represent the mean square error loss of cumulative discharge capacity and SOH, respectively. The L2 norm regularization term for the residual coefficients of the Q-branch. This is the weight decay term applied to all trainable parameters; Q estimates the loss term as follows: in, Estimate the loss term for Q. This is the current training mini-batch sample index set. Batch size; This is the tag for the actual cumulative discharge capacity of the nth cell in the jth cycle. The corresponding model prediction value; The estimated loss term for SOH is: in, Estimate the loss term for SOH. and These are the actual SOH label and the model prediction value for the nth cell in the jth cycle, respectively. Q estimates the regularization term for the branch residual coefficients: in, Estimate the regularization term for the branch residual coefficients of Q; For the residual coefficient vector generated by the joint representation vector in the Q branch, apply L2 regularization to suppress overfitting and smooth the capacity curve; Apply a weight decay term to all trainable parameters: in, This indicates the application of a weight decay term; This represents the set of all trainable parameters of the model, to which a L2 regularization is applied to improve the model's generalization ability.

6. A joint estimation system for Q and SOH of retired lithium batteries based on partial charging segments, characterized in that, The system includes: The data acquisition and processing module is used to collect lithium battery degradation data through charge-discharge cycle tests, aggregate the collected data by cycle number, obtain the relaxation stage voltage sequence and charging stage statistical scalar characteristics for each cycle, standardize them, and normalize the cycle number to obtain a normalized time variable. ; The multimodal temporal coding module maps the voltage sequence of the relaxation stage in each cycle to the latent space and performs multi-scale temporal mixing and channel mixing stacking encoding. It then fuses the corresponding statistical scalar feature encoding of the charging stage with the voltage sequence encoding result to obtain a joint representation vector. ; The joint estimation model construction and training module is used to construct and train a joint estimation model of Q and SOH. This model includes a Q estimation branch based on monotonic prior constraints and a SOH estimation branch based on exponential constraints. The Q estimation branch is used to perform estimation based on the monotonic basis function layer and the residual correction layer. The Q-estimation result is generated above, and the SOH estimation branch is used to generate the Q-estimation result based on the exponential decay parameterization. The SOH estimation results are generated above; The joint estimation module is used to perform joint representation vector estimation. The inputs are fed into the Q-estimation branch and the SOH-estimation branch during training, and the corresponding loop is output at the end. The results of the joint estimation of Q and SOH are as follows.

7. An electronic device, characterized in that, The device includes: a processor and a memory; The memory is used to store one or more program instructions; The processor is configured to run one or more program instructions to perform the steps of a method for jointly estimating Q and SOH of a retired lithium battery based on a partial charging segment as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a method for jointly estimating Q and SOH of a retired lithium battery based on a partial charging segment as described in any one of claims 1 to 6.

Images