A cross-condition lithium battery state joint estimation method based on lightweight operating condition coding

Abstract

The application discloses a cross-condition lithium battery state joint estimation method based on lightweight working condition coding and belongs to the technical field of lithium ion batteries. The method comprises the following steps: establishing a fractional order equivalent circuit model, constructing segment statistical characteristics and sequence characteristics; designing an LTCE encoder to extract segment working condition embedding, constructing a working condition-evidence adaptation matrix, mapping working condition category probability to prior adaptation degree under the current working condition, combining a credibility correction factor to form a cross-condition evidence adaptation factor; calculating an original physical credibility, generating a working condition adaptive credibility gating factor, enabling high-credibility evidence to participate in model calibration and filtering update, and inhibiting low-credibility evidence; injecting the credible evidence screened through the gate into an augmented fractional order particle filtering framework to realize joint estimation of SOC, capacity and current deviation. The application realizes the reservation of physical model constraints while improving the stability and engineering applicability of lithium battery SOC and capacity joint estimation based on fragmented cross-condition operation data.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery technology, and more specifically to a method for joint estimation of the state of lithium batteries across operating conditions based on lightweight operating condition coding. Background Technology

[0002] Lithium-ion batteries, as an important energy storage device, are widely used in new energy vehicles and energy storage systems. To ensure the safe, reliable, and efficient operation of lithium batteries, a Battery Management System (BMS) is typically used to monitor and estimate the battery's State of Charge (SOC), available capacity, terminal voltage, and operating limits in real time. SOC characterizes the battery's current remaining charge, while available capacity characterizes the amount of charge the battery can release or receive under its current aging state. Both directly affect vehicle range display, charge / discharge power limits, and safety protection strategies. Therefore, accurate estimation of lithium battery SOC and available capacity is crucial for improving battery system operational safety and extending its lifespan.

[0003] However, actual vehicle operation is not a single, continuous, and standardized experimental condition. Instead, it consists of alternating segments such as shallow charging and discharging, stop-and-go traffic, high-speed cruising, fast charging, low-temperature operation, and dynamic discharge. These complex driving behaviors cause the battery's State of Charge (SOC) operating range to exhibit significant fragmentation. It is difficult to obtain complete discharge cycles, long periods of inactivity, or calibration data covering the entire SOC range during natural vehicle use. Furthermore, temperature changes, battery aging, current sensor bias, and individual battery variations cause model parameters to continuously drift with operating conditions and lifespan. Fixed-parameter models struggle to maintain stable estimation accuracy under long-term operation across various conditions, easily leading to cumulative SOC drift, capacity reference shift, and increased terminal voltage prediction errors.

[0004] To address the fragmented, cross-operating-condition characteristics mentioned above, existing methods often treat current and voltage data as homogeneous time series inputs to the estimator, or only select certain segments for model correction based on fixed thresholds, lacking a collaborative assessment of the physical information attributes of the segments and their applicability to the operating conditions. While different segments may contain information such as open-circuit voltage, zero-bias current, impedance polarization parameters, or capacity measurement, whether and to what extent this information participates in the update depends on the polarization state, excitation sufficiency, metering integrity, and noise level under the current operating condition. Without condition-adaptive correction to the credibility of the evidence, low-credibility segments can easily be misused for OCV calibration, parameter updates, or capacity correction, leading to state estimation jumps and erroneous model updates. Therefore, there is an urgent need for a joint SOC and capacity estimation method that can further evaluate the cross-operating-condition credibility of segment evidence after identifying its type, and systematically inject credible evidence into the physical filtering framework. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a joint estimation method for the state of lithium batteries across operating conditions based on lightweight operating condition coding. This method has advantages such as adaptability to fragmented operating data across operating conditions, clear evidence application, strong interpretability, low computational cost, and the ability to simultaneously estimate SOC and available capacity.

[0006] To achieve the above objectives, this invention provides a joint estimation method for the state of lithium batteries across operating conditions based on lightweight operating condition coding, specifically including the following steps: S1. Collect the charging and discharging current and terminal voltage data of the lithium battery, and establish a fractional-order equivalent circuit model of the lithium battery. S2. A sliding window is used to divide the charging and discharging current and terminal voltage data into candidate segments, construct segment statistical feature vectors, and identify low-current near-steady-state segments, dynamic current excitation segments, and charging metering segments based on physical evidence to determine the types of candidate evidence that each segment may provide. S3. Construct the feature vector of the segment sequence and build a lightweight temporal convolutional condition encoder (LTCE). The LTCE takes the segment sequence features as input, and sequentially performs input projection, causal and depth-separable one-dimensional convolutional residual calculation and attention pooling, and outputs the segment condition embedding vector, condition category probability and confidence correction factor. The LTCE consists of an input projection layer, two layers of causal depth separable one-dimensional convolutional residual blocks, an attention pooling layer, a working condition classification output head, and a confidence correction output head. S4. Establish a working condition-evidence adaptation matrix, map the working condition category probability to the prior fit degree of each type of evidence under the current working condition, and combine it with the credibility correction factor output by LTCE to form a cross-working condition evidence adaptation factor, and construct a working condition-evidence adaptation mechanism. S5. Based on terminal voltage compensation, near-zero current, dynamic voltage residual and charging metering information, form OCV anchor point, current zero bias, impedance polarization parameters and capacity candidate evidence, calculate the original physical credibility of each type of candidate evidence, and fuse the original physical credibility with cross-operating condition evidence adaptation factor to generate operating condition adaptive credibility gating factor. S6. Based on the operating condition adaptive credible gating factor, candidate evidence is screened, and the evidence screening threshold, model parameter update gain, capacity fusion weight and filtering observation noise are adjusted. Under the credible gating constraint, the OCV local correction term, current deviation, model parameters and available capacity are updated. S7. Inject the updated OCV local correction term, current deviation, model parameters, and capacity evidence into the augmented fractional particle filter framework; sequentially through particle state prediction, weight update, normalization, effective particle number resampling, and posterior estimation, output the lithium battery SOC estimate, available capacity estimate, and current deviation estimate, thus completing the cross-operating condition lithium battery state joint estimation method.

[0007] As a preferred technical solution of the present invention, S1 adopts the current direction convention of positive discharge current and negative charging current, and defines the effective value of current, fractional equivalent circuit model and fractional discrete form.

[0008] As a preferred embodiment of the present invention, in step S2, the BMS divides the operating data into sliding windows. For the th candidate segment, the segment's mean current, standard deviation of current, rate of change of current, mean voltage, standard deviation of voltage, rate of change of voltage, segment duration, cumulative charge, and SOC span are calculated. If the segment meets the conditions of low current, low voltage fluctuation, and minimum duration, it is marked as a low-current near-steady-state segment; if the segment has sufficient current change or voltage dynamic response, it is marked as a dynamic current excitation segment; if the segment meets the conditions of charging direction consistency and SOC span, it is marked as a charging metering segment. This step is only used to determine the type of evidence that the segment may generate, and does not directly determine whether the evidence participates in the model update; if a segment meets multiple criteria simultaneously, it is allowed to correspond to multiple types of candidate evidence, but each type of evidence must be subjected to subsequent adaptive confidence gating under operating conditions.

[0009] As a preferred embodiment of the present invention, step S3 further includes the following steps: S3.1 Construct the segment sequence feature matrix. The effective current value, terminal voltage, current change rate, voltage change rate, normalized time, and cumulative charge within each segment are combined to form the sequence feature matrix, which is then normalized. S3.2 Perform lightweight temporal convolutional coding. Input the sequence feature matrix into LTCE, first unify the feature dimension through the input projection layer, and then extract the temporal variation features of the segment through two layers of causal depth-separable one-dimensional convolutional residual blocks; S3.3 Perform attention pooling. Utilize attention weights to highlight sampling points within a segment that contribute significantly to condition identification and evidence reliability assessment, thereby obtaining the condition embedding vector; S3.4 Perform dual-output head calculations. The operating condition classification output head outputs the probability of the operating condition category, and the credibility correction output head outputs the credibility correction factors for OCV anchor point evidence, current deviation evidence, dynamic excitation evidence, and capacity evidence.

[0010] As a preferred embodiment of the present invention, S4 is the core step for achieving cross-working-condition adaptation, and it further includes the following steps: S4.1 Establish a set of operating condition categories and a set of evidence types. Operating condition categories include at least one of shallow charge / discharge, traffic jam start / stop, high-speed cruise, fast charging, and other dynamic operating conditions; evidence types include OCV anchor point evidence, current deviation evidence, dynamic excitation evidence, and capacity evidence. S4.2 Establish the working condition-evidence fit matrix. Each element in the matrix represents the prior fit of a certain type of evidence under a certain working condition, used to describe the impact of different working conditions on the reliability of different evidence; S4.3. Use the working condition category probabilities output by LTCE to weight the working condition-evidence fitting matrix to obtain the working condition prior fitting degree of the current segment on various types of evidence; S4.4. Integrate the prior fit of the operating conditions with the credibility correction factor to form a cross-operating condition evidence fit factor. This factor does not generate specific evidence values, but rather represents the strength of use of a certain type of evidence under the current operating conditions, providing a basis for subsequent evidence gating.

[0011] As a preferred embodiment of the present invention, step S5 further includes the following steps: S5.1. Generate candidate evidence and calculate the original physical credibility. OCV anchor point evidence is generated based on terminal voltage compensation; current deviation evidence is generated based on near-zero current; dynamic excitation evidence is generated based on dynamic voltage residual, parameter regression vector, and information matrix; and capacity evidence is generated based on cumulative charge and start / end SOC anchor points. The original physical credibility of each is then calculated. S5.2 Calculate the adaptive credibility gating factor for different operating conditions. Multiply the original physical credibility of each type of candidate evidence by the cross-operating condition evidence adaptation factor and limit it to a preset range; S5.3 Adjust the evidence screening threshold based on the credibility gating factor. For evidence with high fit and high physical credibility under the current operating conditions, appropriately relax the screening threshold; for evidence with low fit or low physical credibility under the current operating conditions, increase the screening threshold or prohibit it from participating in the update. S5.4 Adjust the update gain, capacity fusion weight, and filtered observation noise according to the confidence gating factor. High-confidence evidence enters the filter with a larger update gain, a larger fusion weight, and a smaller equivalent observation noise, while low-confidence evidence enters the filter with a smaller update gain, a smaller fusion weight, or a larger observation noise, and may not participate in the update if necessary.

[0012] As a preferred embodiment of the present invention, step S6 further includes the following steps: S6.1 For low-current near-steady-state segments that are gated, the BMS performs ohmic voltage drop compensation and polarization compensation on the terminal voltage based on the current SOC and polarization state, and updates the local OCV correction term; when the segment is in the polarization recovery stage after the end of a large current, the participation of OCV anchor point evidence is reduced to avoid erroneous correction of the OCV curve. S6.2 For near-zero current segments that pass through gating, the BMS updates the current deviation based on the segment reference current; if the segment is in a strong dynamic recovery phase, its participation level is reduced. S6.3 For gated dynamic current excitation segments, the BMS updates the ohmic internal resistance, polarization resistance, generalized capacitive coefficient, and fractional order based on the information matrix and terminal voltage residuals; when the segment excitation is insufficient, the voltage residuals are not directly interpreted as parameter drift. S6.4 For gated charging metering segments, the BMS performs slow capacity calibration based on the cumulative charged capacity and SOC span; when the segment is a short-term replenishment, the end of fast charging, or the SOC span is insufficient, the weight of capacity evidence is reduced to avoid capacity miscalibration.

[0013] As a preferred embodiment of the present invention, in step S7, the BMS injects the OCV local correction term, model parameters, capacity evidence, and current deviation evidence into the augmented fractional-order particle filter framework. The particle state includes SOC, fractional-order polarization voltage, available capacity, and current deviation. Each particle performs state prediction based on the fractional-order equivalent circuit model and updates its weights based on the terminal voltage observation residual and the evidence residual passed through a reliable gating system. For cross-operating condition segments, the filter does not directly use fixed observation noise, but instead uses equivalent observation noise determined by an operating condition adaptive reliable gating factor, thereby varying the degree of influence of the same type of evidence on the particle weights under different operating conditions.

[0014] As a preferred embodiment of the present invention, the LTCE is operated by offline training followed by online freezing or slow updating. During the offline training phase, the work condition classification output head and the confidence correction output head are trained using work condition labels and evidence quality labels generated by physical rules. During the online operation phase, when BMS computing resources are limited, network parameters can be frozen, and only forward inference can be performed. When continuous high-confidence segments exist, slow online updates with a limited step size can also be used.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention first determines the physical information that the data can provide by segment recognition, and then determines whether the information is reliable under the current working condition by working condition coding, thereby avoiding the homogenization of all fragmented running data into the filter and reducing the risk of SOC jump, capacity miscorrection and model parameter divergence.

[0016] (2) This invention clearly achieves cross-operating condition adaptation through a chain mechanism of “operating condition category probability - operating condition - evidence adaptation matrix - cross-operating condition evidence adaptation factor - trustworthy gating factor”. This mechanism allows evidence under operating conditions such as shallow charging and discharging, congestion start-stop, high-speed cruising, fast charging and replenishment and dynamic discharge to enter the model calibration and particle filtering according to different trustworthiness, rather than simply relying on a fixed threshold to process all operating conditions.

[0017] (3) The LTCE in this invention does not directly estimate SOC and capacity, but is only used for working condition coding and evidence trust transfer. It retains the physical constraints of fractional-order model and particle filter, and avoids the instability of extrapolation of pure black box model under unseen working conditions.

[0018] (4) This invention integrates local OCV correction, impedance and fractional polarization parameter update, slow capacity calibration and current deviation estimation into the augmented fractional particle filter framework, which can continuously correct OCV drift, parameter drift, capacity decay and sensor zero bias during long-term operation.

[0019] (5) The LTCE used in this invention consists of two layers of causal depth separable one-dimensional convolutional residual blocks and attention pooling. It has a small number of parameters and low computational complexity, making it suitable for deployment in the BMS edge computing environment. When computational resources are limited, it can be run by offline training and online freezing. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method steps proposed in this invention; Figure 2 This is a schematic diagram of the lightweight temporal convolutional encoder (LTCE) structure of the present invention. Figure 3 This is a schematic diagram of the adaptive trusted gating system under operating conditions of the present invention; Figure 4 This is a schematic diagram of the adaptive evidence gating result under LTCE conditions of the present invention, wherein, Figure 4 In the middle (a), the working condition-evidence prior fitting matrix is ​​shown. Figure 4 (b) is a schematic diagram of the OCV plotting results. Figure 4 (c) is a schematic diagram of the zero-bias current result. Figure 4 (d) is a schematic diagram of the dynamic parameter results. Figure 4 (e) is a schematic diagram of the capacity evidence results; Figure 5 This is a diagram showing the SOC estimation results based on LTCE multi-fragment evidence gating in this invention. Figure 5 (a) is a graph comparing the SOC estimation results. Figure 5 (b) SOC estimation error curve; Figure 6 This is a diagram showing the current zero-bias compensation results based on low-current evidence in this invention. Figure 6 (a) is a comparison chart of SOC trajectories. Figure 6 (b) is a comparison chart of SOC error before and after compensation. Figure 6 (c) shows the results of the dynamic zero-bias estimation. Figure 7 This is a graph showing the slow update results based on capacity evidence in this invention. Figure 7 (a) is a graph comparing the capacity estimation results. Figure 7 (b) is the relative error curve. Detailed Implementation

[0021] 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.

[0022] S1. Collect charging and discharging current and terminal voltage data of lithium battery to construct a running dataset; establish a fractional equivalent circuit model and implement the fractional polarization equation using the finite memory Grünwald-Letnikov discrete form; initialize the model state variables and parameters.

[0023] The expression for constructing the running dataset is: in, To run the dataset, For the first Current sampling value at each sampling time, For the first The terminal voltage sample value at each sampling time. For the first Each sampling time, This represents the total number of sampling points. Adopting the current direction convention of positive discharge current and negative charging current, and using the current sensor deviation as an independent estimator, the effective current value is obtained: The effective value of the current is: in, For the first The effective value of the current at each sampling time. For the first Current sampling value at each sampling time, For the first Current deviation at each sampling time This is the sampling time sequence number.

[0024] based on A fractional-order equivalent circuit model is established for the structure. The SOC state equation, fractional-order polarization equation, and terminal voltage observation equation are as follows: in, and The first The and the first SOC at each sampling time, The sampling time interval, For Coulomb efficiency, For the first Available capacity at each sampling time, for Fractional differential operators of order, It is a fractional order, and , It is a fractional polarization voltage. For generalized capacitive coefficient, Polarization resistor, Terminal voltage, Based on the OCV-SOC relationship Open circuit voltage at the point, This is a local correction term for OCV. Let be the internal resistance of the ohm. To facilitate discretization in the BMS, the fractional polarization equation is discretized using the finite-memory GL (Grünwald-Letnikov) form, as shown in the following formula: in, For finite memory length, For fractional-order memory item numbers, For the first Fractional memory coefficients, for Fractional memory coefficient at time, For a historic moment Fractional polarization voltage.

[0025] The BMS initializes the SOC, fractional polarization voltage, available capacity, current deviation, and model parameters based on the above model. .

[0026] S2. A sliding window is used to divide the charging and discharging current and terminal voltage data into candidate segments, construct segment statistical features and sequence features, and identify low-current near-steady-state segments, dynamic current excitation segments and charging metering segments based on physical evidence to determine the types of candidate evidence that each segment may provide. S2.1, Candidate fragment segmentation and construction of statistical feature vectors. BMS uses a sliding window to process the running dataset. Perform candidate segmentation, the first The candidate segments are represented as follows: in, Number the segments. For the first A set of sampling sequence numbers for candidate segments. and These are the start and end sampling numbers of the segment, respectively. The segment length is given. For each candidate segment, BMS constructs a segment statistical feature vector. : in, The average current of the segment. The standard deviation of the segment current. The maximum rate of change of current in the segment. This represents the average voltage across the segment. The standard deviation of the segment terminal voltage. The maximum voltage change rate of the segment. For the duration of the segment, Accumulate battery power for the segment. The SOC span of the segment.

[0027] S2.2. Based on the segment statistical feature vector and physical basis, identify and label the low current near steady state segment, dynamic current excitation segment and charging metering segment; The low-current near-steady-state segment satisfies: The dynamic current excitation segment satisfies at least one of the following conditions: The charging metering segment satisfies: in, For low flow threshold, The threshold value for the standard deviation of the low-current segment current. The near-steady-state voltage standard deviation threshold. The minimum near-steady-state segment length, The threshold for the standard deviation of dynamic current. The threshold for the rate of change of dynamic current. The threshold for the dynamic voltage change rate. This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. The threshold for charging direction consistency. This is the minimum SOC span threshold.

[0028] If a fragment satisfies multiple criteria, it is allowed to correspond to multiple candidate evidence types, but each candidate evidence type must pass the subsequent adaptive trusted gating.

[0029] S3. Lightweight Temporal Convolutional Condition Encoder (LTCE) Construction and Condition Embedding Extraction: Construct a segment sequence feature vector and build a lightweight temporal convolutional condition encoder (LTCE). The LTCE takes the segment sequence features as input and sequentially performs input projection, causal depth separable one-dimensional convolution residual calculation and attention pooling, and outputs segment condition embedding vector, condition category probability and confidence correction factor. This step does not directly estimate SOC or capacity, but instead maps the fragment sequence to the probability of the working condition category and the evidence credibility correction factor, specifically including the following sub-steps: S3.1. Normalize the original sequence feature vector by mean-variance to construct the segment sequence feature vector; The expression for the original sequence feature vector is: in, The sampling sequence number within the segment. For the first The first segment The original sequence feature vector of each sampling point The effective value of the current at the sampling point. The terminal voltage at the sampling point. The rate of change of current, The rate of change of voltage. Normalized time within the segment, From the starting point of the segment to the... The cumulative electricity consumption at each sampling point The normalized sequence feature vector, The feature mean vector of the training set. The standard deviation vector of the features in the training set. To prevent positive numbers with a denominator of zero, For the first The sequence feature matrix of each segment, This represents the total number of sampling points within the segment. Indicates vector transpose; S3.2 Based on the feature vector of the segment sequence, perform input projection and causal depth separable one-dimensional convolution residual calculation to obtain the hidden feature matrix, and use two layers of causal depth separable one-dimensional convolution residual blocks to extract the segment temporal features; The expression for the hidden feature matrix is ​​as follows: A two-layer causal depthwise separable one-dimensional convolutional residual block is used to extract segment temporal features, as shown in the following expression: S3.3 Calculate the attention weights for the latent feature matrix and perform a weighted summation, then perform attention pooling to obtain the segment condition embedding vector; The calculation expression is as follows: S3.4. Based on the fragment working condition embedding vector, perform dual-output head calculation. The working condition classification output head outputs the working condition category probability, and the credibility correction output head outputs the credibility correction factor for each type of evidence. The expression for the probability of the working condition category is: The expression for the credibility correction factor is: in, The input is the projected latent feature matrix. and These are the weight matrix and bias vector of the input projection layer, respectively. It is a non-linear activation function. This represents the convolutional residual block layer number. For the first Intermediate features after depthwise convolution For the expansion rate Causal depthwise one-dimensional convolution, For the first Layer convolution dilation rate, For batch normalization, For point-by-point one-dimensional convolution, For the first Layer-by-layer pointwise convolution output, For layer normalization, For the first Layer residual block output, for The Middle The latent feature vector at time 1. This is the output of the second-level residual block. For the first Attention weight at each moment, , and These are the trainable parameters for the attention pooling layer. It is the hyperbolic tangent function. For the first The working condition embedding vector of each segment, For the first The probability vector of the working condition category for each segment. For normalized exponential functions, and The weight matrix and bias vector of the output head are used for classifying operating conditions. For the credibility correction factor vector, For logical functions, and To improve reliability, the weight matrix and bias vector of the output header are adjusted. This indicates vector concatenation.

[0030] The credibility correction factors include: credibility correction factors for OCV anchor point evidence, current deviation evidence, dynamic excitation evidence, and capacity evidence.

[0031] S4. Establish a working condition-evidence adaptation matrix, map the working condition category probability to the prior fit degree of each type of evidence under the current working condition, and combine it with the credibility correction factor output by LTCE to form a cross-working condition evidence adaptation factor, and construct a working condition-evidence adaptation mechanism. This step, which transforms the "identified working conditions" into the "intensity of use of various types of evidence," is the core of this invention's cross-working-condition estimation and specifically includes the following sub-steps: S4.1 Establish a set of working condition categories and a set of evidence types; Let the set of working condition categories be The set of evidence types is ,in, , , and These represent OCV anchor point evidence, current deviation evidence, dynamic excitation evidence, and capacity evidence, respectively. S4.2 Establishing a working condition-evidence adaptation matrix The expression is as follows: in, For the first Under the first type of working condition Prior fit of evidence, Index for working condition categories, For evidence category index, This represents the total number of operating condition categories. This matrix can be preset by expert rules, obtained from historical data statistics, or trained offline.

[0032] S4.3. Use the working condition category probabilities output by LTCE to weight the fitting matrix to obtain the working condition prior fitness of the i-th type of evidence under the r-th segment: in, For the first The first segment Prior fit of similar evidence to working conditions For the first The segment belongs to the first The probability of different working conditions. This step allows for a soft fit of evidence to different types of evidence based on the probabilities of multiple working conditions, rather than hard classifying the fragments into a single working condition.

[0033] S4.4. Integrate the prior fit of the operating conditions with the LTCE confidence correction factor to form a cross-operating condition evidence fit factor: in, The first output of LTCE Evidence credibility correction factor For the first The strength of condition correction for similar evidence. For the first The first segment Cross-condition evidence fit factor for similar types of evidence. and The first The lower and upper limits of the fit factor for class evidence. This represents the amplitude limiting function.

[0034] This step does not directly generate evidence observations, but transforms the impact of operating conditions into the degree of fit of various types of evidence, thereby enabling cross-operating condition adaptation to be implemented as differences in subsequent evidence gating and update intensity.

[0035] S5. Construction of Operating Condition Adaptive Trusted Gating. The BMS forms candidate evidence observations based on terminal voltage compensation, near-zero current, dynamic voltage residual, and charging metering information. These observations are then fused with their original physical trustworthiness and cross-operating condition evidence adaptation factors to obtain the final gating factor. Based on this factor, the evidence screening threshold, updated gain, capacity fusion weights, and filtered observation noise are adjusted. This process includes the following sub-steps: S5.1 Generate candidate evidence observations and calculate the physical credibility of the segments. For OCV anchor point evidence, the physical credibility is determined by segment current fluctuation, voltage fluctuation, and OCV residual; for current deviation evidence, the physical credibility is determined by near-zero current segment stability; for dynamic excitation evidence, the physical credibility is determined by the minimum eigenvalue of the information matrix and the variance of the terminal voltage residual; for capacity evidence, the physical credibility is determined by the SOC span of the charging segment and the credibility of the start and end OCV anchor points.

[0036] S5.2 Calculate the adaptive reliability gating factor for the operating condition: in, For the first The first segment The original physical credibility of such evidence, For the first The first segment Cross-condition evidence fit factor for similar types of evidence. For the first The first segment Condition-adaptive credibility gating factor for similar evidence; When the fragment itself meets the physical conditions but the current working conditions are not suitable for providing a certain type of evidence. This reduces the likelihood of the candidate evidence being updated; when the fragment is physically reliable and has a high degree of fit to the operating conditions... With increased strength, candidate evidence is allowed to participate in updates with greater intensity.

[0037] S5.3 Adjust the evidence screening threshold, update the gain, capacity fusion weight, and filter observation noise according to the adaptive confidence gating factor based on the operating conditions; The formulas for calculating the screening threshold, update gain, capacity fusion weight, and filtered observation noise are as follows: in, For the first The first segment Adaptive screening threshold for class evidence, For the first The baseline threshold for this type of evidence, This is the threshold adjustment coefficient. and These are the lower threshold and the upper threshold, respectively. For the first The first segment Update gain of evidence class Update the gain based on the baseline. For the first The first segment The equivalent observation noise variance when similar evidence enters the filter. The variance of the baseline observation noise. To prevent positive constants with a denominator of zero, the operating condition category does not directly determine the SOC or capacity. Instead, it alters the way different pieces of evidence enter the physical model and filters through evidence adaptation, gating thresholds, update gain, and observation noise.

[0038] Through this step, high-confidence evidence enters the model with a larger update gain, a larger capacity fusion weight, and smaller observation noise, while low-confidence evidence participates with a smaller update gain, a smaller fusion weight, or larger observation noise, and does not participate in the update when necessary.

[0039] S6. Perform multi-fragment evidence updates based on trusted gating. BMS only uses candidate evidence that passes the screening threshold for subsequent model calibration and filtering updates. Candidate evidence that fails the gating does not participate in the update, or the corresponding state only maintains strong regularized slow change.

[0040] S6.1 For low-current near-steady-state segments, the BMS generates candidate evidence for OCV anchor points. First, the average SOC of the segment and the OCV anchor points after ohmic voltage drop and polarization voltage compensation are calculated: in, For the first The mean SOC of a near-steady-state segment with low flow. This is the OCV anchor point after ohmic voltage drop and polarization voltage compensation. The original physical credibility of the OCV anchor point evidence is: in, For OCV anchor point residuals, To assess the original physical credibility of OCV anchor evidence, This is the voltage stability scaling factor. This is the current stability scaling factor. This is a local correction term for OCV.

[0041] When a segment is in the polarization recovery phase after a large current has ended, even if the current is temporarily small, LTCE will reduce its OCV evidence fit factor, thereby avoiding miscorrection of the OCV curve.

[0042] The OCV local correction term is represented using normalized local basis functions: in, ' is the total number of local basis functions. Number the local basis functions. For the first OCV local correction coefficients For the first 'Normalized local basis functions, For the first Local basis function centers For the first The width of each local basis function; For Temporary index, The current state of charge; The local correction coefficients for OCV are derived recursively as follows: in, and These are the OCV local correction coefficient vectors before and after the update, respectively. For OCV correction coefficients, physical constraint projection operator. Update the gain for OCV evidence. As the reliable gating factor for adaptive OCV evidence under operating conditions, For Huber robust functions, The threshold for the Huber function. These are local basis function vectors.

[0043] S6.2 For the near-zero current segment, the BMS generates candidate evidence for current deviation. Its original physical confidence and current deviation update formulas are as follows: in, To establish the original physical credibility of the evidence regarding current deviation, This is the current stability scaling factor. This is the voltage stability scaling factor. and These are the current deviations before and after the update, respectively. Update the gain to account for current deviation. The reliable gating factor for adaptive current deviation under operating conditions. This is the reference current corresponding to the near-zero current segment.

[0044] S6.3 For the dynamic current excitation segment, the BMS generates candidate evidence for impedance and fractional polarization parameters. The model parameter vector and regression vector are defined as follows: in, For the first The model parameter vector at each sampling time point, For ohmic internal resistance, Polarization resistor, For generalized capacitive coefficient, For fractional order, This is the regression vector of the predicted terminal voltage on the model parameters. This is the predicted terminal voltage calculated from the parameter vector.

[0045] For the Construct an information matrix from dynamic segments: in, For the first Information matrix of dynamic segments The smallest eigenvalue of the information matrix. and These are the lower and upper thresholds for the excitation quality, respectively. For the variance of the voltage residuals in the dynamic segment, The voltage residual scaling factor. To dynamically incentivize the original physical credibility of the evidence, the model parameters are updated using robust recursive updates with a forgetting factor and physical constraints. in, For the first The recursive gain at each sampling time, and These are the parameter covariance matrices before and after the update, respectively. Forgetting factor, and These are the model parameter vectors before and after the update, respectively. To dynamically incentivize evidence to update gain, For the reliable gating factor of adaptive dynamic excitation under working conditions, The physical feasible region projection operator is used to constrain the model parameters. , , , And limit the magnitude and rate of change of the model parameters in a single update.

[0046] S6.4 For charging metering segments, the BMS generates capacity candidate evidence. The cumulative charging volume and capacity evidence for the segment are as follows: in, Accumulate charge for the charging segment. For charging coulomb efficiency, This indicates that only the current in the charging direction is counted. For capacity evidence obtained from charging metering segments, and These are the SOC anchor points at the beginning and end of the charging segment, respectively. The original physical credibility of capacity evidence and the capacity slow calibration formula are as follows: in For the original physical credibility of the capacity evidence, and These are the trusted gating factors for the OCV anchor points at the start and end, respectively. For reference to SOC span, and These are the available capacities before and after the capacity update, respectively. For capacity fusion gain, As a reliable gating factor for adaptive capacity evidence under operating conditions, This is a capacity physical constraint projection operator used to limit the capacity located at... Within the range. If the segment belongs to short-term power replenishment, the end of fast charging constant voltage, or insufficient SOC span, then Reduced capacity means that evidence is not included in the update or is only included in the update with a low weight.

[0047] S7. Inject the updated OCV local correction term, current deviation, model parameters, and capacity evidence into the augmented fractional particle filter framework; and output the lithium battery SOC estimate, available capacity estimate, and current deviation estimate by sequentially performing particle state prediction, weight update, normalization, effective particle number resampling, and posterior estimation.

[0048] The first step of the BMS augmented fractional particle filter framework The states of the particles are: The first The formula for predicting the state of a single particle is: in, For the first The sampling time of the first sampling moment The state vector of each particle Number the particles. For the first SOC of individual particles For the first Fractional polarization voltage of each particle For the first The available capacity of each particle. For the first Current deviation of individual particles It is a fractional polarization state transition function. , , and These are process noises related to SOC, polarization voltage, capacitance, and current deviation, respectively. For the first The predicted terminal voltage of each particle.

[0049] The formulas for particle weight update, normalization, effective particle number resampling, and posterior estimation are as follows: in, For the first The unnormalized weights of each particle, For the first The normalized weights of each particle at the previous time step For the first The normalized weights of each particle at the current moment The variance of the voltage observation noise at the lower end of the current segment. For the first Each fragment is passed through a trusted gated set of evidence. For the first The first segment Class of evidence for the first Evidence residuals formed by individual particles For the first The first segment Equivalent observation noise variance of similar evidence The total number of particles, The effective number of particles, This is the posterior estimate of SOC. This is the posterior estimate of the available capacity. This is the posterior estimate of the current deviation; when Resampling is performed at that time. This is the resampling threshold.

[0050] S8. Slow Capacity Fusion, Model Feedback, and Final Output. When reliable capacity evidence exists for the current segment, the BMS slowly fuses the charging metering capacity evidence with the particle-filtered capacity posterior. When reliable capacity evidence is absent, the capacity only maintains a strongly positive, slow change. The capacity estimation result is fed back to the SOC ampere-hour integral reference, the current deviation estimation result is fed back to the current RMS calculation, and the model parameter update result is fed back to the terminal voltage prediction model. Finally, the BMS outputs the SOC estimate, the available capacity estimate, and the current deviation estimate, and uses the updated OCV local correction term, model parameters, capacity estimate, and current deviation as priors for the calculation of the next segment.

[0051] S9, LTCE Training and Online Deployment. During offline training, the BMS or host computer trains the LTCE using historical segments labeled with work condition tags and physical rule evidence quality tags. The loss function includes work condition classification loss, evidence quality supervision loss, confidence correction factor smoothing loss, and network parameter regularization loss. During online operation, when BMS computational resources are limited, the LTCE network parameters can be frozen, and only forward inference can be performed. When consecutive high-confidence segments appear, a limited step size can be used for slow online updates. Therefore, the LTCE is not used as a SOC or capacity black-box estimator, but rather as a cross-work condition evidence confidence transfer module, working together with fractional-order models and particle filter frameworks to complete constrained joint estimation.

[0052] The LTCE method obtains network parameters by offline training followed by online freezing or slow updates. The loss function for offline training is: in, Let be the total loss function for the training segments. Losses are classified according to operating conditions. Loss due to evidence quality supervision To smooth the loss using a credibility correction factor, For network parameter regularization loss, , and These are the weights of the evidence quality supervision loss, the credibility correction factor smoothing loss, and the network parameter regularization loss, respectively. This represents the total number of operating condition categories. For the operating condition category number, For the first The first segment Type of working condition label, For the first The first segment Probability of similar working conditions For the first The first segment Class of evidence quality labels The first output of LTCE Evidence credibility correction factor This is the credibility correction factor vector for the previous segment. For all trainable parameters of LTCE, This represents the L2 norm.

[0053] The evidence quality labels are generated by physical rules, specifically including: in, For OCV anchor point evidence quality labeling, To dynamically incentivize evidence quality labels, For the quality label of capacity evidence, This represents the upper limit of the OCV anchor point residual. This represents the minimum confidence threshold for OCV anchor points. When updating slowly online, only use Fragments exceeding a preset threshold update the LTCE network parameters, and satisfy the following conditions: in, To update the serial number online, and The first LTCE network parameters before and after the last online update To update the step size of network parameters online, The gradient of the loss function with respect to the network parameters. Update the magnitude constraint projection operator for network parameters.

[0054] To verify the feasibility and effectiveness of the proposed cross-operating condition joint estimation method for lithium batteries based on lightweight operating condition coding and multi-segment evidence fusion, the method was validated using the constructed cross-operating condition operating data. The LTCE model was trained offline and then frozen online or updated slowly. Its model training and main structural settings are as follows: Table 1: LTCE Model Training and Main Structure Settings To comprehensively evaluate the estimation performance of the method of this invention, root mean square error (RMSE), mean absolute error (AAE), and relative improvement rate were selected as the main evaluation indicators. When SOC is expressed as a percentage, the SOC error is uniformly represented by "%"; capacity estimation is evaluated using the RMSE of the relative capacity error. The improvement rate is calculated as follows: Improvement rate = (RMSE benchmark) RMSE (in this invention) / RMSE benchmark × 100% The verification results of this invention are as follows: Figures 4 to 7 As shown in Table 2. Figure 4 Used to verify LTCE condition coding and condition-adaptive evidence gating mechanism. Figures 5 to 7 The effectiveness of the proposed joint estimation method was verified from three aspects: SOC estimation, current zero bias compensation, and slow capacity update. The main quantitative results are summarized in Table 2.

[0055] (1) LTCE adaptive evidence gating results; like Figure 4 As shown, Figure 4 (a) The prior fit relationships between different operating conditions and various candidate evidence types are presented. The prior fit of the low-current near-steady-state operating condition to the OCV anchor point evidence and the current zero bias evidence are 1.00 and 0.85, respectively; the prior fit of the dynamic driving condition to the dynamic parameter evidence is 1.00; the prior fit of the charging metering condition to the capacity evidence is 1.00, while the prior fit of various evidence types under uncertain operating conditions remains at a low level overall. The above results show that the constructed operating condition-evidence fit matrix can assign differentiated prior weights to the OCV anchor point, current zero bias, dynamic parameters, and capacity evidence based on the physical information contained in different operating segments.

[0056] Figure 4 (b) to Figure 4 (e) Further, the actual gating results of various types of evidence after LTCE condition coding and credibility correction are presented. Among them, OCV anchor point and current zero bias evidence only participate in the update when the physical conditions of the corresponding segment are met and the gating is effective; dynamic parameter evidence maintains a high gating strength in the operating range with sufficient dynamic excitation; and capacity evidence participates in capacity calibration at a lower frequency and slower time scale. It can be seen that the present invention can transform the condition information identified by LTCE into the actual usability of different evidence, enhance the high credibility evidence that matches the current condition, and suppress the inapplicable or low credibility evidence, thereby verifying the cross-condition evidence adjustment mechanism of "condition coding - evidence adaptation - credibility gating".

[0057] (2) SOC estimation results; like Figure 5 As shown, Figure 5 (a) SOC estimation results are presented using the reference SOC, UPF, residual fusion, and LTCE evidence fusion methods. Figure 5 (b) The corresponding SOC estimation error is given. As can be seen from the figure, all methods can follow the overall trend of SOC change, but the LTCE evidence fusion result has a higher degree of agreement with the reference SOC and maintains a smaller estimation error in most operating ranges.

[0058] Based on the quantitative results in Table 2, the SOC RMSE of the DEKF, DPF, UPF, residual fusion, and LTCE evidence fusion methods are 2.1044%, 1.1194%, 0.8876%, 0.7209%, and 0.6887%, respectively. Among them, the LTCE evidence fusion method achieves the lowest SOC RMSE, reducing it by 22.41% compared to UPF and by approximately 4.47% compared to the method using only residual fusion; its mean absolute error is 0.4929%. Further statistical analysis based on the operation process divided into the pre-operation, mid-operation, and post-operation stages shows that the SOC RMSE of LTCE evidence fusion is reduced by 14.15%, 40.43%, and 19.41% compared to UPF, respectively, indicating that the accuracy improvement of the proposed method is not limited to a single local interval but can maintain relatively stable estimation performance in different operation stages.

[0059] (3) Results of zero-bias current compensation; Figure 6 The compensation results are presented under the condition that the current sensor has zero bias. Among them, Figure 6 (a) The SOC estimation trajectories were compared with those without zero bias, with injected zero bias, and with compensation. Figure 6 (b) The SOC error before and after zero bias compensation is given; Figure 6 (c) The dynamic estimation process and valid evidence samples for zero current bias under strict gating conditions are given.

[0060] Depend on Figure 6 It is evident that without zero-bias compensation, current measurement deviations accumulate gradually through ampere-hour integration, causing continuous SOC drift. After zero-bias calibration using the low-current near-steady-state evidence of this invention, the estimated zero bias can gradually track the injected zero bias, significantly suppressing the accumulated SOC error. Quantitative results show that the SOC RMSE after zero-bias injection is 1.2443%, which decreases to 0.0287% after zero-bias compensation, representing a 97.69% reduction in error. This result demonstrates that the low-current near-steady-state segment can provide effective calibration information for the current sensor's zero bias and improve the stability of the long-term SOC integration reference through zero-bias closed-loop correction.

[0061] (4) Slow capacity update results; Figure 7 The estimation results after incorporating capacity evidence into slow timescale updates are presented. Among them, Figure 7 (a) The results of reference capacity, UPF prior, HPPC econometric evidence update, and gated multi-segment fusion update were compared. Figure 7 (b) The corresponding relative capacity error is given. The reference capacity is only used for result evaluation and is not involved in capacity evidence generation or online updates.

[0062] Depend on Figure 7 It is evident that when using HPPC metrological evidence alone for updating capacity, the capacity estimation is easily affected by local metrological fluctuations and the quality of fragment information, resulting in significant fluctuations in the estimation curve. After adaptive reliability gating and slow timescale fusion, the capacity update results are smoother and can more stably remain near the reference capacity. Quantitative results show that the RMSE of the prior capacity using UPF is 0.4523%, while the RMSE of the capacity updated using HPPC metrological evidence is 0.5211%, an increase of 15.21% compared to UPF. After gating and multi-fragment fusion for updating, the capacity RMSE decreases to 0.3050%, a decrease of 32.56% compared to UPF. These results indicate that unfiltered single metrological evidence does not necessarily improve the accuracy of capacity estimation. This invention, through reliable gating and slow fusion, limits the update intensity of low-quality or locally fluctuating evidence, thereby improving the stability and accuracy of capacity estimation.

[0063] (5) Summary of key verification indicators; Table 2 summarizes the main quantitative results of SOC estimation, current zero-bias compensation, and capacity estimation. Overall, the SOC RMSE of the LTCE evidence fusion method decreased from 0.8876% of the UPF to 0.6887%; after using low-current evidence for current zero-bias compensation, the SOC RMSE decreased from 1.2443% to 0.0287%; and after using gated multi-segment capacity fusion update, the capacity RMSE decreased from 0.4523% to 0.3050%.

[0064] Table 2: Summary of Main Verification Indicators of the Invention In summary, the verification results demonstrate that this invention can utilize LTCE to encode fragmented operating segments based on their operating conditions. Through operating condition-evidence adaptation and a trusted gating mechanism, it differentiates and adjusts various physical evidences, such as OCV anchor points, current bias, dynamic parameters, and capacity, allowing trusted evidence matching the current operating state to participate in model correction and state estimation. Further verification results for SOC, current bias, and capacity show that this mechanism not only reduces SOC estimation errors but also effectively suppresses long-term integral drift caused by sensor bias and improves the stability of slow-timescale capacity estimation. This verifies the feasibility and effectiveness of the cross-operating condition joint estimation method of this invention.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A joint estimation method for the state of lithium batteries across operating conditions based on lightweight operating condition coding, characterized in that, Includes the following steps: S1. Collect the charging and discharging current and terminal voltage data of the lithium battery, and establish a fractional-order equivalent circuit model of the lithium battery. S2. A sliding window is used to divide the charging and discharging current and terminal voltage data into candidate segments, construct segment statistical feature vectors, and identify low-current near-steady-state segments, dynamic current excitation segments, and charging metering segments based on physical evidence to determine the candidate evidence type. S3. Construct fragment sequence feature vectors and build a lightweight temporal convolutional condition encoder (LTCE). The LTCE takes fragment sequence features as input, and sequentially performs input projection, causal depth separable one-dimensional convolutional residual calculation and attention pooling. It outputs fragment condition embedding vectors, condition category probabilities and confidence correction factors. S4. Establish a working condition-evidence adaptation matrix, map the working condition category probability to the prior fit degree of each type of evidence under the current working condition, and combine it with the credibility correction factor output by LTCE to form a cross-working condition evidence adaptation factor, and construct a working condition-evidence adaptation mechanism. S5. Based on terminal voltage compensation, near-zero current, dynamic voltage residual and charging metering information, form OCV anchor point, current zero bias, impedance polarization parameters and capacity candidate evidence, calculate the original physical credibility of each type of candidate evidence, and fuse the original physical credibility with cross-operating condition evidence adaptation factor to generate operating condition adaptive credibility gating factor. S6. Based on the operating condition adaptive credible gating factor, candidate evidence is screened, and the evidence screening threshold, model parameter update gain, capacity fusion weight and filtering observation noise are adjusted. Under the credible gating constraint, the OCV local correction term, current deviation, model parameters and available capacity are updated. S7. Inject the updated OCV local correction term, current deviation, model parameters, and capacity evidence into the augmented fractional particle filter framework; sequentially through particle state prediction, weight update, normalization, effective particle number resampling, and posterior estimation, output the lithium battery SOC estimate, available capacity estimate, and current deviation estimate, thus completing the cross-operating condition lithium battery state joint estimation method.

2. The method for joint estimation of lithium battery state across operating conditions based on lightweight operating condition coding according to claim 1, characterized in that, In S1, the expression for constructing the running dataset is: in, To run the dataset, For the first Current sampling value at each sampling time, For the first The terminal voltage sample value at each sampling time. For the first Each sampling time, The total number of sampling points. For sampling time index; The SOC state equation, fractional polarization equation, and terminal voltage observation equation of the fractional equivalent circuit model are as follows: in, and The first The and the first SOC at each sampling time, For Coulomb efficiency, The sampling time interval, For the first The effective value of the current at each sampling time. For the first Available capacity at each sampling time, for Fractional differential operators of order, For fractional order, For the first Fractional polarization voltage at each sampling time For generalized capacitive coefficient, Polarization resistor, Terminal voltage, Based on the OCV-SOC relationship Open circuit voltage at the point, This is a local correction term for OCV. It is the internal resistance of the Ohm.

3. The method for joint estimation of lithium battery state across operating conditions based on lightweight operating condition coding according to claim 2, characterized in that, The fractional polarization equation is implemented in a finite-memory GL discrete form, and the specific formula is as follows: in, For finite memory length, For fractional-order memory item numbers, For the first Fractional memory coefficients, for Fractional memory coefficient at time, For a historic moment Fractional polarization voltage.

4. The method for joint estimation of lithium battery state across operating conditions based on lightweight operating condition coding according to claim 3, characterized in that, S2 specifically includes the following steps: S2.

1. Use a sliding window to divide the running dataset into candidate segments and construct segment statistical feature vectors; No. The candidate segments are represented as follows: in, Number the segments. For the first A set of sampling sequence numbers for candidate segments. For sampling time index, and These are the start and end sampling numbers of the segment, respectively. The length of the segment; For each candidate segment, construct a segment statistical feature vector, as shown below: in, For the first The statistical feature vector of each segment, The average current of the segment. The standard deviation of the segment current. The maximum rate of change of current in the segment. The average voltage at the segment terminals. The standard deviation of the segment terminal voltage. The maximum voltage change rate of the segment. For the duration of the segment, Accumulate battery power for the segment. For the SOC span of the segment, Indicates vector transpose; S2.

2. Based on the segment statistical feature vector and physical basis, perform segment identification to identify low current near steady state segments, dynamic current excitation segments, and charging metering segments; Based on physical principles, the fragment recognition method is as follows: The low-current near-steady-state segment satisfies: The dynamic current excitation segment satisfies at least one of the following conditions: The charging metering segment satisfies: in, For low flow threshold, The threshold value for the standard deviation of the low-current segment current. The near-steady-state voltage standard deviation threshold. The minimum near-steady-state segment length, The threshold for the standard deviation of dynamic current. The threshold for the rate of change of dynamic current. The threshold for the dynamic voltage change rate. For indicator functions, The threshold for charging direction consistency. This is the minimum SOC span threshold.

5. The method for joint estimation of lithium battery state across operating conditions based on lightweight operating condition coding according to claim 4, characterized in that, S3, the LTCE is composed of an input projection layer, two layers of causal depth separable one-dimensional convolutional residual blocks, an attention pooling layer, a working condition classification output head and a confidence correction output head; Specifically, the following steps are included: S3.

1. Normalize the original sequence feature vector by mean-variance to construct the segment sequence feature vector; Original sequence feature vector The expression is: in, The sampling sequence number within the segment. For the first The first segment The original sequence feature vector of each sampling point The effective value of the current at the sampling point. The terminal voltage at the sampling point. The rate of change of current, The rate of change of voltage. Normalized time within the segment, From the starting point of the segment to the number The cumulative electricity consumption at each sampling point The normalized sequence feature vector, The feature mean vector of the training set. The standard deviation vector of the features in the training set. To prevent positive numbers with a denominator of zero, For the first The sequence feature matrix of each segment, This represents the total number of sampling points within the segment. Indicates vector transpose; S3.2 Based on the feature vector of the segment sequence, perform input projection and causal depth separable one-dimensional convolution residual calculation to obtain the hidden feature matrix, and use two layers of causal depth separable one-dimensional convolution residual blocks to extract the segment temporal features; The expression for the hidden feature matrix is ​​as follows: in, The input is the projected latent feature matrix. and These are the weight matrix and bias vector of the input projection layer, respectively. It is a non-linear activation function; A two-layer causal depthwise separable one-dimensional convolutional residual block is used to extract segment temporal features, as shown in the following expression: in, This represents the convolutional residual block layer number. For the first Intermediate features after depthwise convolution For the expansion rate Causal depthwise one-dimensional convolution, For the first Layer convolution dilation rate, For batch normalization, For point-by-point one-dimensional convolution, For the first Layer-by-layer pointwise convolution output, For layer normalization, For the first Layer residual block output; S3.3 Calculate the attention weights for the latent feature matrix and perform a weighted summation, then perform attention pooling to obtain the segment condition embedding vector; The calculation expression is as follows: in, for The Middle The latent feature vector at time 1. This is the output of the second-level residual block. For the first Attention weight at each moment, , and These are the trainable parameters for the attention pooling layer. It is the hyperbolic tangent function. For the first The working condition embedding vector of each segment; S3.

4. Based on the fragment working condition embedding vector, perform dual-output head calculation. The working condition classification output head outputs the working condition category probability, and the credibility correction output head outputs the credibility correction factor for each type of evidence. The expression for the probability of the working condition category is: in, For the first The probability vector of the working condition category for each segment. For normalized exponential functions, and The weight matrix and bias vector of the output head are used for classifying working conditions; The expression for the credibility correction factor is: in, For the credibility correction factor vector, For logical functions, and To improve reliability, the weight matrix and bias vector of the output header are adjusted. This indicates vector concatenation; The credibility correction factors include: credibility correction factors for OCV anchor point evidence, current deviation evidence, dynamic excitation evidence, and capacity evidence.

6. The method for joint estimation of lithium battery state across operating conditions based on lightweight operating condition coding according to claim 5, characterized in that, S4 specifically includes the following steps: S4.1 Establish a set of working condition categories and a set of evidence types; Let the set of working condition categories be The set of evidence types is ,in, , , and These represent OCV anchor point evidence, current deviation evidence, dynamic excitation evidence, and capacity evidence, respectively. S4.2 Establishing a working condition-evidence adaptation matrix The expression is as follows: in, For the first Under the first type of working condition Prior fit of evidence, For working condition category index, For evidence category index, This represents the total number of operating condition categories. S4.

3. Weight the fitting matrix using the probabilities of different working conditions to obtain the prior fitness of the working conditions, as shown in the following expression: in, For the first The first segment Prior fit of similar evidence to working conditions For the first The segment belongs to the first The probability of a particular type of working condition; S4.4 Integrate the prior fit of the operating conditions with the credibility correction factor to form a cross-operating condition evidence fit factor. The expression is as follows: in, For the first The strength of condition correction for similar evidence. The first output of LTCE Evidence credibility correction factor For the first The first segment Cross-condition evidence fit factor for similar types of evidence and The first The lower and upper limits of the fit factor for class evidence. This represents the amplitude limiting function.

7. The method for joint estimation of lithium battery state across operating conditions based on lightweight operating condition coding according to claim 6, characterized in that, S5 specifically includes the following steps: S5.1 Generate candidate evidence observations and calculate the physical credibility of the fragments; For OCV anchor evidence, physical reliability is determined by segment current fluctuations, voltage fluctuations, and OCV residuals; For evidence of current deviation, physical reliability is determined by the stability of the near-zero current segment; For dynamic excitation evidence, physical credibility is determined by the minimum eigenvalue of the information matrix and the variance of the terminal voltage residual; For capacity evidence, physical credibility is determined by the credibility of the SOC span of the charging segment and the start and end OCV anchor points; S5.2 Calculate the adaptive reliability gating factor for the operating condition, as shown in the following expression: in, For the first The first segment The original physical credibility of such evidence, For the first The first segment Cross-condition evidence fit factor for similar types of evidence For the first The first segment Condition-adaptive credibility gating factor for similar evidence; S5.3 Adjust the evidence screening threshold, update the gain, capacity fusion weight, and filter observation noise according to the adaptive confidence gating factor based on the operating conditions; The formulas for calculating the screening threshold, update gain, capacity fusion weight, and filtered observation noise are as follows: in, For the first The first segment Adaptive screening threshold for class evidence, For the first The baseline threshold for this type of evidence This is the threshold adjustment coefficient. and These are the lower threshold and the upper threshold, respectively. For the first The first segment Update gain of evidence class Update the gain based on the baseline. For the first The first segment The equivalent observation noise variance when similar evidence enters the filter. The variance of the baseline observation noise. To prevent positive numbers with a denominator of zero.

8. The method for joint estimation of lithium battery state across operating conditions based on lightweight operating condition coding according to claim 7, characterized in that, S6 specifically includes the following steps: S6.1 For low-current near-steady-state segments, first calculate the mean SOC of the segment, then calculate the OCV anchor point after compensation for ohmic voltage drop and polarization voltage, and then generate OCV anchor point candidate evidence by calculating the OCV anchor point residual and the original physical credibility of the OCV anchor point evidence. The OCV local correction term is calculated by normalized local basis function and updated recursively. The expression for calculating the local correction term of OCV using normalized local basis functions and recursively updating it is as follows: in, and These are the OCV local correction coefficient vectors before and after the update, respectively. For OCV correction coefficients, physical constraint projection operator, Update the gain for OCV evidence. As the working condition adaptive OCV evidence credibility gating factor, For Huber robust functions, The threshold for the Huber function. These are local basis function vectors; S6.2 For near-zero current segments, candidate evidence for current deviation is generated. The original physical confidence and current deviation update formulas are as follows: in, To establish the original physical credibility of the evidence regarding current deviation, The standard deviation of the segment current. The standard deviation of the segment terminal voltage. This is the current stability scaling factor. This is the voltage stability scaling factor. and These are the current deviations before and after the update, respectively. Update the gain to account for current deviation. The reliable gating factor for adaptive current deviation under operating conditions. This is the reference current corresponding to the near-zero current segment; S6.3 For the dynamic current excitation segment, define the model parameter vector, regression vector, information matrix and the original physical credibility of the dynamic excitation, generate candidate evidence for impedance and fractional polarization parameters, and update the model parameters using robust recursion with forgetting factor and physical constraints. The model parameters are updated using a robust recursive method with a forgetting factor and physical constraints, as shown in the following formula: in, For the first The recursive gain at each sampling time, and These are the parameter covariance matrices before and after the update, respectively. Forgetting factor, and These are the model parameter vectors before and after the update, respectively. This is the regression vector of the predicted terminal voltage on the model parameters. To dynamically incentivize evidence to update gain, For the reliable gating factor of adaptive dynamic excitation under working conditions, For the model parameters, the projection operator of the physical feasible region. The predicted terminal voltage is calculated from the parameter vector. The physical feasible region is used to constrain the actual terminal voltage calculated from the parameter vector. , , , And limit the magnitude and rate of change of model parameters in a single update; S6.4 For the charging metering segment, define the calculation formulas for the cumulative charging amount, capacity evidence, original physical credibility of capacity evidence, and capacity slow speed, generate capacity candidate evidence, and update the capacity. The expression for capacity update is: in, and These are the available capacities before and after the capacity update, respectively. For capacity fusion gain, As a reliable gating factor for adaptive capacity evidence under operating conditions, For capacity-physical-constrained projection operators, This is capacity evidence obtained from charging metering segments.

9. The method for joint estimation of lithium battery state across operating conditions based on lightweight operating condition coding according to claim 1, characterized in that, In step S3, the LTCE obtains network parameters by either offline training followed by online freezing or slow online updates. The loss function for offline training is: in, Let be the total loss function for the training segments. Losses are classified according to operating conditions. Loss due to evidence quality supervision To smooth the loss using a credibility correction factor, For network parameter regularization loss, , and These are the weights of the evidence quality supervision loss, the credibility correction factor smoothing loss, and the network parameter regularization loss, respectively. This represents the total number of operating condition categories. For the operating condition category number, For the first The first segment Type of working condition label, For the first The first segment Probability of similar working conditions For the first The first segment Class of evidence quality labels The first output of LTCE Evidence credibility correction factor This is the credibility correction factor vector for the previous segment. For all trainable parameters of LTCE, This represents the L2 norm.

10. The method for joint estimation of lithium battery state across operating conditions based on lightweight operating condition coding according to claim 9, characterized in that, The evidence quality labels are generated by physical rules, specifically including: in, For OCV anchor point evidence quality labeling, To dynamically incentivize evidence quality labels, For the quality label of capacity evidence, This represents the upper limit of the OCV anchor point residual. This represents the minimum confidence threshold for OCV anchor points. When updating slowly online, only use Fragments exceeding a preset threshold update the LTCE network parameters, and satisfy the following conditions: in, To update the serial number online, and The first LTCE network parameters before and after the last online update. To update the step size of network parameters online, The gradient of the loss function with respect to the network parameters. Update the magnitude constraint projection operator for network parameters.

Images