A battery soh adaptive estimation method and system based on conditional reconstruction

Abstract

The application discloses a battery SOH self-adaptive estimation method and system based on conditional reconstruction, and belongs to the technical field of battery health state evaluation. The method comprises the following steps: obtaining target battery data to construct a candidate voltage window, obtaining a real health feature corresponding to each candidate health window, constructing a characteristic vector based on the real health feature and feeding the SOH evaluation model to obtain an OIS; setting a source information interval for a fragmented charging segment in advance, and performing an overlap relationship judgment on the current fragmented charging segment, the OIS and the source information interval to generate a direct estimation routing instruction or a conditional reconstruction routing instruction; pre-establishing and training a cross-interval conditional reconstruction network model, and obtaining a reconstructed health feature corresponding to the conditional reconstruction routing instruction OIS; aligning the real health feature or the reconstructed health feature object format, and feeding the SOH fusion prediction model to output an SOH prediction mean value of the target battery.

Description

Technical Field

[0001] This invention belongs to the field of battery health status assessment technology, and particularly relates to an adaptive estimation method and system for battery state of health (SOH) based on condition reconstruction. Background Technology

[0002] Lithium-ion batteries are widely used in new energy vehicles and energy storage systems. With increasing cycle count and operating time, aging phenomena such as loss of active lithium, electrode material degradation, increased internal resistance, and enhanced polarization gradually appear, leading to a decrease in usable capacity and power output. The State of Health (SOH) is a crucial indicator characterizing the degree of battery capacity decay and safety margin. Accurately estimating SOH is essential for charge / discharge control, safety warnings, energy management, lifespan prediction, and secondary utilization.

[0003] In laboratory environments, State of Harm (SOH) estimation can typically be performed based on complete charge-discharge cycle data. However, in real-world vehicles and energy storage systems, the Battery Management System (BMS) often cannot obtain complete charging curves. User behaviors such as charging and leaving frequently, shallow charging and discharging, charging interruptions, and mixed start and end SOCs result in the BMS only collecting fragmented charging data that is localized, has a narrow voltage range, and is discontinuous. This fragmented data makes traditional SOH estimation methods that rely on complete charging curves, large-scale charging curves, or fixed voltage range characteristics difficult to apply stably. Existing technologies include methods for SOH estimation based on partial charging curves, and schemes that reconstruct partial charging segments from low-correlation voltage ranges to high-correlation voltage ranges, then extract health features and combine them with regression models for SOH estimation. These methods improve the utilization rate of partial charging data, but still have the following problems: 1. Target range selection is biased. Existing methods typically select the target reconstruction range based on correlation coefficients or empirical rules, easily assuming that high-voltage ranges contain more aging information. However, the distribution of aging information along the voltage axis is not fixed or monotonic under different material systems, battery structures, and aging conditions; 2. There is a lack of online conditional routing mechanisms. Some existing reconstruction methods often directly perform reconstruction after obtaining local segments without first determining whether the real-time acquired segments have covered the key aging information intervals. If the real-time segments have already covered the effective optimal information intervals, continuing to call the deep reconstruction model will cause unnecessary computational overhead and may also introduce additional reconstruction errors; 3. There is a lack of handling for differences in the quality of reconstructed features. Existing methods usually input the reconstructed health features as a whole into the back-end regression model without adaptively reducing the weights of low-correlation components, high-noise components, or components with large reconstruction errors in different feature dimensions. Direct fusion can easily lead to an amplification of SOH estimation errors; 4. Insufficient generalization ability across material systems and aging paths. Under different material systems and aging conditions, the optimal information interval may shift between low-voltage, medium-voltage, or high-voltage intervals. Fixed target interval reconstruction strategies are difficult to adapt to such nonlinear changes.

[0004] Therefore, it is necessary to provide a battery SOH adaptive estimation method and system that can first determine the optimal information interval under a specific battery system and operating condition through error scanning, then select direct estimation or conditional reconstruction based on the coverage relationship between online fragmented segments and the optimal information interval, and further perform conditional reconstruction-based estimation of the real or reconstructed health characteristics. Summary of the Invention

[0005] In view of this, the present invention proposes a condition-reconstruction-based adaptive estimation method and system for battery SOH, which can determine the optimal information interval (OIS), output direct estimation routing instructions or conditional reconstruction routing instructions through an online adaptive data scheduling and routing allocation environment, obtain reconstruction health characteristics of the optimal information interval (OIS) based on the conditional reconstruction routing instructions, and predict SOH using a SOH fusion prediction model combined with the results of the direct estimation routing instructions or conditional reconstruction routing instructions.

[0006] On the one hand, the present invention provides an adaptive estimation method for battery SOH based on conditional reconstruction, comprising the following steps: S1: Obtain target battery data to construct candidate voltage windows, obtain the real health characteristics corresponding to each candidate health window, construct a characteristic vector based on the real health characteristics and send it into the SOH evaluation model, and take the candidate voltage window whose SOH estimation error meets the preset optimization criteria as the optimal information interval (OIS). S2: During the actual operation of the target battery, the BMS acquires the current fragmented charging segment, determines the effective voltage range of the charging segment, pre-sets the source information interval LII for the fragmented charging segment, and judges the coverage relationship between the current fragmented charging segment and the optimal information interval OIS and the source information interval. When the current fragmented charging segment covers the optimal information interval OIS, a direct estimation routing instruction is generated; when the current fragmented charging segment does not cover the optimal information interval OIS but covers the source information interval LII, a conditional reconstruction routing instruction is generated; otherwise, no routing instruction is generated. S3: Pre-establish and train the cross-interval conditional reconstruction network model. When generating conditional reconstruction routing instructions, input the data corresponding to the source information interval into the cross-interval conditional reconstruction network model. Use the cross-interval conditional reconstruction network model to generate the reconstruction health features corresponding to the optimal information interval OIS. S4: Align the format of the real health features or reconstructed health features corresponding to the extracted optimal information interval OIS, and feed them into the same SOH fusion prediction model. Calculate the reliability coefficient for each health feature dimension, construct the quality coupling adaptive kernel function, the routing consistency kernel, and the heteroscedasticity noise term based on reconstruction uncertainty, so that the SOH fusion prediction model adaptively outputs the mean SOH prediction of the target battery.

[0007] Based on the above technical solution, preferably, step S1 includes: firstly, acquiring the historical charging data and corresponding SOH tags of the target battery. The historical charging data includes the voltage-capacity sequence during the charging process, as well as current, charging time, charging temperature, SOC, charging rate, and ambient temperature operating condition data. The SOH tag is determined based on the ratio of the current maximum available capacity to the rated capacity or initial capacity. The charging voltage range of the target battery is divided into multiple candidate voltage windows. For each candidate voltage window, corresponding health features are extracted. Health features include, but are not limited to, basic capacity statistical features, curve shape features, electrochemical related features, sequence features, and operating condition fusion features. Based on the health features, a feature vector is constructed for each candidate voltage window. k The corresponding voltage range is [ V k , V k + △V ], constructing a health feature vector using each candidate voltage window X k As input, a SOH evaluation model is established, with the health feature vector corresponding to each candidate voltage window as input, and the SOH estimation error corresponding to each candidate voltage window is calculated. The candidate voltage window whose SOH estimation error satisfies the preset optimization criterion is determined as the optimal information interval (OIS).

[0008] Preferably, the preset optimization criteria include any of the following situations: 1) the SOH estimation error reaches a global minimum; 2) the SOH estimation error reaches a local minimum and is lower than a preset error threshold; 3) the SOH estimation error is located in a continuous low error plateau region; 4) the comprehensive objective function composed of SOH estimation error, candidate voltage window data availability, and online computation cost is optimal; the candidate voltage window that meets the preset optimization criteria is determined as the optimal information interval (OIS); when multiple candidate voltage windows simultaneously meet the conditions, the candidate voltage window with the smallest SOH estimation error, the highest data availability, or the lowest online computation cost is preferentially selected as the optimal information interval (OIS).

[0009] Preferably, in step S2, the current fragmented charging segment is compared with the optimal information interval (OIS) and the source information interval to determine their coverage relationship, wherein the coverage relationship conforms to any of the following rules: 1) the voltage overlap ratio is greater than or equal to a preset coverage threshold; 2) the number of effective sampling points in the optimal information interval (OIS) is greater than or equal to a preset sampling point threshold; 3) the performance of the optimal information interval (OIS) after interpolation and resampling meets a preset quality condition; 4) the voltage monotonicity, current stability, temperature range, and abnormal point ratio in the optimal information interval (OIS) meet preset conditions.

[0010] Preferably, the cross-interval conditional reconstruction network model described in step S3 includes an input layer, a multi-level one-dimensional convolutional feature extraction layer, a quality-gated degenerate pooling layer, and an output layer arranged sequentially; when generating conditional reconstruction routing instructions, the voltage-capacity data corresponding to the source information interval LII is resampled at equal intervals to a length of L in One-dimensional capacity sequence Based on a one-dimensional capacity sequence and further constructing an input tensor X in Condition vector corresponding to a one-dimensional capacity sequence c ; one-dimensional tensor X in The input is fed into the input layer; the multi-level one-dimensional convolutional feature extraction layer includes at least two cascaded one-dimensional convolutional blocks, each of which sequentially includes a one-dimensional convolutional layer, a batch normalization layer, and a nonlinear activation layer. The size of the convolutional kernel and the number of channels of the convolutional kernel in the one-dimensional convolutional layers of different one-dimensional convolutional blocks increase layer by layer, with a stride of 1 for each layer. This is used to capture multi-scale aging features from local polarization to global capacity decay and output multi-channel feature maps; the batch normalization layer is used to accelerate network convergence and alleviate internal covariate shift; the nonlinear activation layer is used to enhance the nonlinearity of the network for multi-scale aging features. The system enhances the representation capabilities of features. After processing through multi-level one-dimensional convolutional feature extraction layers, it performs layer-by-layer high-dimensional mapping and receptive field expansion to output multi-channel features. The quality-gated degenerate pooling layer combines multi-channel features and conditional vectors to obtain quality-gated weights at different positions. Based on the quality-gated weights, degenerate pooling features are obtained for different channels. These degenerate pooling features are fed into a fully connected network, which outputs two types of prediction results in the form of a dual-branch reconstruction head. The first type of prediction result is the reconstruction curve, from which reconstruction health characteristics are obtained. The second type of prediction result is the reconstruction uncertainty.

[0011] Preferably, the loss function of the cross-interval conditional reconstruction network model is a multi-objective loss function, which is obtained by weighting the mean square error between the optimal information interval (OIS) reconstruction curve and the true optimal information interval (OIS) curve, the mean square error between the reconstructed health feature and the true health feature, the curve constraint loss, the SOH loss corresponding to the reconstructed health feature input SOH fusion prediction model, and the uncertainty calibration loss.

[0012] Preferably, step S4 includes: the SOH fusion prediction model includes a quality coupling adaptive kernel function and a heteroscedasticity noise term; and an input feature vector is constructed using the true health features obtained by directly estimating the routing instructions or the reconstructed health features obtained by conditionally reconstructing the routing instructions. x , input feature vector x The m The reliability coefficient of a dimensional feature is defined as follows: r mThe quality-coupled adaptive kernel function includes an exponential term and a routing consistency kernel term. The exponential term is based on two different input feature vectors. The m The first dimension, the first m The length dimension of the first dimension, the first m Definition of dimensional characteristic reliability coefficient r m The routing consistency kernel term, obtained from the signal equation, is represented by either an explicit indication form using the Kronecker function or an exponentially decaying form based on segment quality; the heteroscedasticity noise term is derived by applying the first... n The result of the weighted summation of the mean of online segment quality and reconstruction uncertainty of each input characteristic vector is added to the variance of the baseline noise; the SOH fusion prediction model is used to output the mean of SOH prediction.

[0013] Based on the above technical solutions, preferably, the SOH evaluation model is one or a combination of different models selected from Gaussian process regression model, support vector regression model, random forest model, gradient boosting model, and neural network model.

[0014] Based on the above technical solutions, preferably, the SOH fusion prediction model is any one of the following: Gaussian process regression model ARD-GPR with automatic correlation judgment, Bayesian regression with adaptive adjustment capability of feature weights, sparse Gaussian process, attention network, uncertainty neural network, or integrated uncertainty model.

[0015] On the other hand, the present invention provides a battery SOH adaptive estimation method and system based on conditional reconstruction, comprising: The interval optimization module is used to acquire target battery data to construct candidate voltage windows, acquire the real health characteristics corresponding to each candidate health window, construct a characteristic vector based on the real health characteristics and send it into the SOH evaluation model, and take the candidate voltage window whose SOH estimation error meets the preset optimization criteria as the optimal information interval (OIS). The source information interval determination module is used to determine the source information interval based on collectability, consistency, noise level, and mapping relationship with the optimal information interval (OIS); it determines the coverage relationship between the current fragmented charging segment, the optimal information interval (OIS), and the source information interval, and generates a direct estimation routing instruction or a conditional reconstruction routing instruction; for the direct estimation routing instruction, it extracts the true health characteristics of the optimal information interval (OIS); The reconstruction health feature acquisition module is used to pre-build and train the cross-interval conditional reconstruction network model. When generating conditional reconstruction routing instructions, the data corresponding to the source information interval is input into the cross-interval conditional reconstruction network model. The cross-interval conditional reconstruction network model is used to generate the reconstruction curve and reconstruction uncertainty corresponding to the optimal information interval OIS, and generate the reconstruction health features of OIS. The SOH fusion prediction module is used to adaptively reduce the weight of the true health features or reconstructed health features of the optimal information interval OIS based on the improved Gaussian process regression model ARD-GPR model with automatic correlation judgment and output the SOH estimate. It also updates the OIS, source information interval, routing threshold or model parameters according to the target battery's early cycle data, subsequent running data or cloud-updated data.

[0016] This invention provides a battery SOH adaptive estimation method and system based on conditional reconstruction, comprising: , and has the following advantages compared to the prior art: 1. By scanning candidate windows and establishing a SOH evaluation model, combined with SOH error feedback, the optimal information interval (OIS) is selected through preset optimization criteria. The source information interval is introduced as the reconstruction anchor point, and the coverage relationship between the current fragmented charging segment and OIS and source information interval (LII) is dynamically judged. Two paths are selected: direct estimation or conditional reconstruction, which improves the problem of incomplete and unpredictable charging segments in practical applications. For charging segments that have already covered OIS, the true health characteristics are directly obtained. For segments that need to be reconstructed, the reconstructed health characteristics are obtained through cross-interval conditional reconstruction network. By calculating the reliability coefficient for each health characteristic dimension and constructing a complex kernel function containing routing consistency kernel and heteroscedasticity noise term, the model can automatically identify and weight features from different sources and with different reliability levels. This adaptive mechanism greatly improves the stability and accuracy of the final SOH prediction.

[0017] 2. It clarifies that historical data must include voltage-capacity sequences and data under various operating conditions, and defines multiple health characteristic types, which is conducive to uncovering the multidimensional intrinsic laws of battery degradation. It describes in detail the complete closed-loop process from data acquisition, window division, feature extraction to model evaluation to obtain the optimal information interval OIS, ensuring the objectivity and scientific nature of the OIS screening process.

[0018] 3. The cross-interval conditional reconstruction network model adopts multi-level one-dimensional convolution, with its kernel size and channel number increasing layer by layer. This design can efficiently expand from short-term local fluctuations to long-term global trends, perfectly matching the multi-scale characteristics of the battery charging curve. The quality-gated degradation pooling layer combined with conditional vectors realizes dynamic weighting of different locations and quality features, improving the reconstruction quality. The dual-branch output provides key confidence information for subsequent fusion prediction, which is the basis for achieving adaptive fusion.

[0019] 4. The SOH fusion prediction model defines a reliability coefficient for each feature dimension and constructs a quality-coupled adaptive kernel function based on it. This can distinguish the importance of different features and give differentiated trust to features from different sources. The routing consistency kernel considers whether the data comes from direct estimation or conditional reconstruction path. The heteroscedasticity noise term treats the reconstruction uncertainty as part of the noise, which can effectively improve the robustness and interpretability of the final prediction.

[0020] 5. Possesses robust physical noise reduction and accuracy surpassing capabilities. Real-world testing data shows that, on the Oxford dataset under dynamic driving conditions, the smooth features reconstructed using this invention achieve a prediction RMSE as low as 0.83%, not only compensating for data gaps but also exceeding the prediction accuracy of 0.91% using the original noisy real features. It provides a safety net under extreme conditions. In extreme situations such as high-current accelerated aging and severe polarization shifts, like the NASA dataset, traditional direct mapping methods suffer from feature loss, leading to an error collapse to 12.59%. This invention, through conditional reconstruction combined with the leak-free ARD-GPR weight dilution algorithm, overcomes extreme information gaps, forcibly pulling the prediction error back to 5.75%, effectively preventing the divergence and loss of control in SOH estimation, and establishing a high-robustness technical barrier for dealing with real-world complex conditions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall process of the battery SOH adaptive estimation method and system based on conditional reconstruction according to the present invention. Figure 2 This is a schematic diagram illustrating the root mean square error of a battery SOH adaptive estimation method and system based on conditional reconstruction, using different datasets, according to the present invention. Figure 3 This invention presents a battery SOH adaptive estimation method and system based on conditional reconstruction, showing the SOH estimation curves and actual measured SOH curves under different cycle numbers and datasets. Figure 4 This is a schematic diagram of the root mean square error of SOH estimation for different voltage ranges and different datasets in the battery SOH adaptive estimation method and system based on conditional reconstruction according to the present invention. Figure 5This is a schematic diagram of the cross-interval conditional reconstruction network model structure of a battery SOH adaptive estimation method and system based on conditional reconstruction according to the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] Existing SOH estimation methods have the following problems: 1. Target interval selection is biased. Existing methods usually select the target reconstruction interval based on correlation coefficients or empirical rules, easily assuming that the high-voltage interval contains more aging information. However, under different material systems, battery structures, and aging conditions, the distribution of aging information along the voltage axis is not fixed or monotonic, and the correlation interval or high-voltage interval is not necessarily the optimal interval with the smallest SOH estimation error; 2. There is a lack of online conditional routing mechanisms. Some existing reconstruction methods often perform reconstruction directly after obtaining local segments without first determining whether the real-time acquired segments have covered the key aging information interval. If the real-time segment already covers the effective optimal information interval, continuing to call the deep reconstruction model will cause unnecessary computational overhead and may also introduce additional reconstruction errors, which is not conducive to deployment in vehicle-side BMS with limited computing power and storage resources; 3. Lack of handling for differences in the quality of reconstructed features. Existing methods usually input the reconstructed health features as a whole into the back-end regression model without adaptively reducing the weights of low-correlation components, high-noise components, or components with large reconstruction errors in different feature dimensions. Direct fusion can easily lead to amplification of SOH estimation errors; 4. Insufficient generalization ability across material systems and across aging paths. Under different material systems and different aging conditions, the optimal information interval may shift between low-pressure, medium-pressure, or high-pressure intervals. Fixed target interval reconstruction strategies are difficult to adapt to such nonlinear changes.

[0025] In view of this, such as Figure 1 As shown, on one hand, the present invention provides a battery SOH adaptive estimation method based on conditional reconstruction, comprising the following steps: S1: Obtain target battery data to construct candidate voltage windows, obtain the real health characteristics corresponding to each candidate health window, construct a characteristic vector based on the real health characteristics and send it into the SOH evaluation model, and take the candidate voltage window whose SOH estimation error meets the preset optimization criteria as the optimal information interval (OIS).

[0026] The specific steps are as follows: First, acquire the historical charging data and corresponding State of Health (SOH) tags of the target battery. The historical charging data includes the voltage-capacity sequence during charging, as well as data on current, charging time, charging temperature, State of Charge (SOC), charging rate, and ambient temperature. The SOH tag is determined based on the ratio of the current maximum available capacity to the rated capacity or initial capacity. The charging voltage range of the target battery is divided into multiple candidate voltage windows. Each candidate voltage window has a fixed width, sliding step size, and adaptive voltage platform. The width of the candidate voltage window is 0.01V-0.20V, preferably 0.05V. Candidate voltage windows can be continuous windows or combinations of multiple discontinuous windows. The method of dividing candidate voltage windows is not limited to a single voltage axis variable; it can also combine SOC, time, or capacity data on top of the voltage axis. For each candidate voltage window, extract the corresponding health features and construct a feature vector based on these health features. k The corresponding voltage range is [ V k , V k + △V ], constructing a health feature vector using each candidate voltage window X k As input, Formula 1, where Indicates the maximum and minimum capacity. Indicates the average capacity value. This represents the capacity increment within a voltage range. The standard deviation of the capacity data is represented; a SOH evaluation model is established, with the health feature vector corresponding to each candidate voltage window as input, and the SOH estimation error corresponding to each candidate voltage window is calculated. The candidate voltage window whose SOH estimation error satisfies the preset optimization criterion is determined as the optimal information interval (OIS).

[0027] Health characteristics include, but are not limited to: 1. Basic capacity statistical characteristics: maximum capacity, minimum capacity, average capacity, capacity increment, and capacity standard deviation; 2. Curve shape characteristics: local slope, curvature, area, and plateau length; 3. Electrochemical characteristics: incremental capacity characteristics, differential voltage characteristics, phase transition peak position or peak width; 4. Sequence characteristics: Voltage-capacity sequence, current-time sequence, or capacity-time sequence after interpolation or resampling; 5. Operating condition fusion characteristics: at least one of temperature, magnification, SOC start and end points, and segment duration.

[0028] The preset optimization criteria include any of the following scenarios: 1) The SOH estimation error reaches a global minimum; 2) The SOH estimation error reaches a local minimum and is below a preset error threshold; 3) The SOH estimation error is located in a continuous low-error plateau region; 4) The comprehensive objective function composed of the SOH estimation error, the availability of candidate voltage window data, and the online computation cost is optimal. Candidate voltage windows that satisfy the preset optimization criteria are determined as the optimal information interval (OIS). When multiple candidate voltage windows simultaneously satisfy the conditions, the candidate voltage window with the smallest SOH estimation error, the highest data availability, or the lowest online computation cost is preferentially selected as the optimal information interval (OIS). Formula 2, where For the first k The root mean square error of the prediction of SOH using candidate voltage windows. This indicates that the objective function Index of the candidate voltage window when the minimum value is reached k .

[0029] A local minimum refers to a situation where the SOH estimation error Ek of the k-th candidate voltage window is not greater than the SOH estimation error of its adjacent candidate windows, for example, satisfying Ek≤Ek-1 and Ek≤Ek+1. The preset error threshold can be an SOH estimation RMSE of no more than 0.5% to 5%, preferably no more than 2%; or Ek can be no more than 1.05 to 1.30 times the global minimum error Emin, preferably 1.15 times.

[0030] An error plateau region refers to a region where the SOH estimation error for at least two consecutive candidate voltage windows is below a plateau error threshold, and the error variation between adjacent windows is less than a preset fluctuation threshold. The plateau error threshold can be 1.05 to 1.30 times the global minimum error Emin, preferably 1.20 times; the preset fluctuation threshold can be the relative change rate of error between adjacent windows being no more than 5% to 20%.

[0031] The overall objective function is a custom weighted normalized objective function, which can be expressed as: Jk = α·Norm(Ek) + β·Norm(1-Ak) + γ·Norm(Ck), where Ek represents the SOH estimation error, Ak represents the data availability of the candidate window, Ck represents the online computation cost, and α, β, and γ are weighting coefficients with α+β+γ=1. The candidate voltage window with the smallest Jk is determined as the optimal information interval. Norm(·) is used for norm operations.

[0032] In this implementation, the SOH evaluation model is one or a combination of different models selected from Gaussian process regression, support vector regression, random forest, gradient boosting, and neural network models. As a preferred implementation, Gaussian process regression is used as the SOH evaluation model, and the root mean square error (RMSE) for each window is calculated through leave-one-cell cross-validation or cross-cell validation.

[0033] S2: Online adaptive data scheduling and route allocation; During the actual operation of the target battery, the BMS acquires the current fragmented charging segment, determines the effective voltage range of the charging segment, pre-sets the source information interval LII for the fragmented charging segment, and judges the coverage relationship between the current fragmented charging segment and the optimal information interval OIS and the source information interval. When the current fragmented charging segment covers the optimal information interval OIS, a direct estimation route instruction is generated; when the current fragmented charging segment does not cover the optimal information interval OIS but covers the source information interval LII, a conditional reconstruction route instruction is generated; otherwise, no route instruction is generated.

[0034] In step S2, the current fragmented charging segment is compared with the optimal information interval (OIS) and the source information interval to determine their coverage relationship. The coverage relationship is determined by any of the following rules: 1) The voltage overlap ratio is greater than or equal to the preset coverage threshold; 2) The number of effective sampling points in the optimal information interval (OIS) is greater than or equal to the preset sampling point threshold; 3) The performance of the optimal information interval (OIS) after interpolation and resampling meets the preset quality conditions; 4) The voltage monotonicity, current stability, temperature range, and abnormal point ratio in the optimal information interval (OIS) meet the preset conditions.

[0035] Specifically, when the voltage overlap ratio is greater than or equal to the preset coverage threshold, or the number of valid sampling points in the optimal information interval (OIS) is greater than or equal to the preset sampling point threshold, or the features of the optimal information interval (OIS) after interpolation and resampling meet the preset quality conditions, a direct estimation routing instruction is generated.

[0036] When a fragmented charging segment does not cover the OIS but covers the source information range, a conditional reconfiguration routing instruction is generated.

[0037] When a fragmented charging segment fails to cover either the OIS or the source information range, a low confidence prompt can be output, the system can wait for subsequent segments, invoke the historical segment fusion strategy, or request the cloud model to perform supplementary estimation.

[0038] When generating a direct estimation routing command, the system extracts the true health features corresponding to the OIS from fragmented charging segments and inputs these features into the SOH fusion prediction model in step S4. This path eliminates the need to call the cross-interval conditional reconstruction model, reducing online computational overhead and the introduction of reconstruction errors, making it suitable for real-time operation of the vehicle-side BMS. When generating a conditional reconstruction routing command, the data corresponding to the source information interval is input into the cross-interval conditional reconstruction model in step S3 to generate the reconstruction health features corresponding to the OIS of the optimal information interval.

[0039] S3: Pre-establish and train a cross-interval conditional reconstruction network model. When generating conditional reconstruction routing instructions, input the data corresponding to the source information interval into the cross-interval conditional reconstruction network model. Use the cross-interval conditional reconstruction network model to generate the reconstruction health features corresponding to the optimal information interval OIS.

[0040] The cross-interval conditional reconstruction network model mentioned here is a 1D-CNN-GAP network structure, including a sequentially arranged input layer, multi-level one-dimensional convolutional feature extraction layers, quality-gated degenerate pooling layers, and output layer. When generating conditional reconstruction routing instructions, the voltage-capacity data corresponding to the source information interval LII is resampled at equal intervals to a length of... L in One-dimensional capacity sequence Formula 3, based on a one-dimensional capacity sequence and further constructing the input tensor X in Condition vector corresponding to a one-dimensional capacity sequence c , Formula 4, where The voltage overlap coverage ratio of the source information interval. The number of effective sampling points, For the voltage span of the source information interval, The average current, The average temperature. and For the source information interval LII, the start SOC and end SOC are respectively. This is a type identifier for routing instructions, including two types: direct estimation routing instructions and conditional reconstruction routing instructions.

[0041] A one-dimensional tensor is fed into the input layer; the multi-level one-dimensional convolutional feature extraction layer includes at least two cascaded one-dimensional convolutional blocks, each of which sequentially includes a one-dimensional convolutional layer, a batch normalization layer, and a non-linear activation layer. The kernel size and number of channels of the one-dimensional convolutional layers in different one-dimensional convolutional blocks increase progressively, with a stride of 1 for each layer. This is used to capture multi-scale aging features from local polarization to global capacity decay and output multi-channel feature maps; the batch normalization layer is used to accelerate network convergence and alleviate internal covariate shift; the non-linear activation layer is used to enhance the network's response to... The nonlinear representation capability of multi-scale aging features is demonstrated. After processing through multi-level one-dimensional convolutional feature extraction layers, layer-by-layer high-dimensional mapping and receptive field expansion are performed to output multi-channel features. A quality-gated degenerate pooling layer combines multi-channel features and conditional vectors to obtain quality-gated weights at different locations. Based on these weights, degenerate pooling features for different channels are obtained. These features are then fed into a fully connected network, outputting two types of prediction results in the form of a dual-branch reconstruction head. The first type of prediction is the reconstruction curve; the second type is the reconstruction uncertainty. Reconstructed health features are obtained based on the reconstruction curve.

[0042] The training process of the conditional reconstruction model includes: acquiring source information interval data and optimal information interval data of the source domain battery in the same charging cycle or adjacent health states, performing outlier removal, equal-interval resampling, and normalization processing respectively, and constructing training sample pairs. Each training sample includes a one-dimensional input tensor corresponding to the source information interval, a condition vector, and a target reconstruction curve or target health feature corresponding to the optimal information interval.

[0043] During training, a one-dimensional input tensor and a conditional vector of the source information interval are input into the conditional reconstruction model. The model outputs a reconstruction curve and reconstruction uncertainty. A reconstruction loss is constructed based on the error between the reconstruction curve and the true curve in the optimal information interval, and an uncertainty constraint loss is constructed based on the reconstruction uncertainty. The training objective function can be expressed as: L = λ1·Lrec + λ2·Lunc + λ3·Lsmooth; where Lrec is the mean square error or mean absolute error between the reconstruction curve and the true curve, Lunc is the negative log-likelihood loss based on prediction uncertainty, Lsmooth is the first-order or second-order difference smoothing constraint of the reconstruction curve, and λ1, λ2, and λ3 are loss weight coefficients. Preferably, λ1 is between 0.5 and 1.0, λ2 is between 0.05 and 0.5, and λ3 is between 0.001 and 0.1.

[0044] The hyperparameters of the conditional reconstruction model include: 2 to 4 one-dimensional convolutional blocks; 3 to 9 convolutional kernels, preferably 3, 5, or 7; the number of convolutional channels is set layer by layer to 16, 32, 64 or 32, 64, 128; the stride is 1; the batch normalization layer is placed after the convolutional layer; the non-linear activation function is ReLU, LeakyReLU, or GELU; the dropout rate is 0.1 to 0.5, preferably 0.2; the number of fully connected layers is 1 to 3; the training batch size is 16 to 128; the learning rate is 1e-4 to 1e-2, preferably 1e-3; the optimizer is Adam, AdamW, or SGD; the number of training epochs is 50 to 500, and an early stopping strategy can be adopted, stopping training when the validation set loss no longer decreases within 10 to 30 consecutive epochs.

[0045] The second type of prediction result is the reconstruction uncertainty, which can be constrained to a positive value using a Softplus function or an exponential function. This uncertainty characterizes the reliability of the reconstruction curve or the reconstructed health feature. The reconstruction uncertainty can then be used as input to the backend SOH fusion prediction model, or used to weight the reconstructed health feature with confidence.

[0046] After constructing the input data, feature extraction is performed using a network body containing multi-scale one-dimensional convolutional branches. Different branches use different kernel lengths to extract local polarization perturbations, plateau shape changes, and capacity degradation trends, respectively. Preferably, the kernel size can be at least two of 3, 5, and 7; the number of channels can be set to 16, 32, and 64, increasing layer by layer; the convolution stride is set to 1, and each convolutional layer is followed by a cascaded batch normalization layer, i.e., a BN layer and a LeakyReLU activation layer. b The output of each convolutional branch is represented as: Formula 5, to overcome the feature loss problem of ordinary global average pooling, introduces a quality-gated degenerate pooling layer. This layer abandons the indiscriminate averaging of sequence positions and instead dynamically assigns reliability weights to different positions based on local features, segment quality, and conditional vectors. , Formula 6, Formula 7, where H i Indicates the first i Multi-channel features at each location This represents a trainable quality scoring function. For the first i Quality gating weights generated at each location, It is the first i Channels at each location c Corresponding features It is a passage c It is the degenerate pooling feature of the final output.L out This is the upper limit of the location. This design can effectively reduce the negative interference of anomalous sampling points, local polarization spikes, and low-coverage locations on the overall reconstructed features.

[0047] After being processed by a quality-gated degenerate pooling layer, the degenerate pooling features are fed into a fully connected network, outputting two types of prediction results in the form of a dual-branch reconstruction head. One is the OIS target interval reconstruction curve. Formula 8; another is the reconstruction uncertainty. , Formula 9, F rec and F unc These are the network mapping function notations, respectively.

[0048] The loss function of the cross-interval conditional reconstruction network model is a multi-objective loss function, which is obtained by weighting the mean squared error between the optimal information interval (OIS) reconstruction curve and the true optimal information interval (OIS) curve, the mean squared error between the reconstructed health feature and the true health feature, the curve constraint loss, the SOH loss corresponding to the coupling of the reconstructed health feature input SOH fusion prediction model, and the uncertainty calibration loss. Let the optimal information interval (OIS) reconstruction curve be... The reconstruction uncertainty is The reconstructed health feature vector extracted based on the OIS reconstruction curve of the optimal information interval is: Formula 10, where Represents the reconstructed values ​​for maximum and minimum capacity. The reconstructed value represents the average capacity value. This represents the reconstructed value of the capacity increment within the voltage range. Reconstructed values ​​representing the standard deviation of the capacity data. and Let represent the slope and area of ​​the OIS reconstruction curve representing the optimal information interval; the expression for the multi-objective loss function is: Formula 11, where , The curves representing the true optimal information interval (OIS) and their health characteristics are shown below. Represents healthy characteristics of restructuring. To calculate the mean square error, The monotonicity constraint of the OIS reconstruction curve for the optimal information interval and the constraint loss of the slope or area of ​​the incremental capacity curve are considered. , Formula 12, This represents two adjacent points on the OIS reconstruction curve representing the optimal information interval. The slope, The reconstructed first i The local capacity increment of the first interval and the first intervali The actual local capacity increment of each interval; Input the reconstructed health features corresponding to the reconstructed curve into the SOH evaluation model to assess the task loss; For uncertainty calibration loss; This is a weighted term, and its value range is... , , , , .

[0049] The conditional reconstruction network model in this embodiment is 1D-CNN-GAP, but it is not a simple combination of ordinary 1D-CNN and GAP layers. Instead, it embeds OIS optimization results, fragment quality evaluation, source interval coverage quality, and SOH estimation tasks into the network input, pooling weights, and loss function. The conditional reconstruction network model is not limited to 1D-CNN-GAP; it can also use temporal convolutional networks, Transformers, gated recurrent networks, autoencoders, or physically constrained neural networks. However, it preferably includes global pooling, attention aggregation, or other feature aggregation structures that can suppress local noise.

[0050] S4: Align the format of the real health features or reconstructed health features corresponding to the extracted optimal information interval OIS, and feed them into the same SOH fusion prediction model. Calculate the reliability coefficient for each health feature dimension, construct the quality coupling adaptive kernel function, the routing consistency kernel, and the heteroscedasticity noise term based on reconstruction uncertainty, so that the SOH fusion prediction model adaptively outputs the mean SOH prediction of the target battery.

[0051] Regardless of whether the input is the true health features obtained from direct route estimation or the reconstructed health features generated from conditional route reconstruction, it is uniformly input into the SOH fusion prediction model. This scheme introduces an adaptive weighted kernel function coupled with route quality, organically integrating route quality, feature reliability, and reconstruction uncertainty into the kernel function and noise term.

[0052] Specifically, the SOH fusion prediction model includes a quality-coupled adaptive kernel function and a heteroscedasticity noise term; an input feature vector is constructed using the true health features obtained by directly estimating routing instructions or the reconstructed health features obtained by conditionally reconstructing routing instructions. x , will input the first feature vector m The reliability coefficient of a dimensional feature is defined as follows: r m The quality-coupled adaptive kernel function includes an exponential term and a routing consistency kernel term. The exponential term is based on the first of two different input feature vectors. m The first dimension, the first m The length dimension of the first dimension, the firstm Definition of dimensional characteristic reliability coefficient r m The characteristic reliability coefficient is obtained from the signal equation. r m The adaptive calculation formula is: , Formula 13, For online clip quality, For the first m Dimensional reconstruction uncertainty, dimension sequence number , The upper limit of the dimension, For the first m 3D reconstruction error, To learn regression parameters or preset hyperparameters; when using direct route estimation, the reliability coefficient is mainly determined by the OIS coverage ratio, the number of sampling points, and the operational stability. When using conditional route reconstruction, the reliability coefficient needs to be further combined with the reconstruction uncertainty. Make corrections. The mass coupling adaptive kernel function is: , Formula 14, Represents the signal variance. Representing two different input feature vectors The m One dimension, For the first m The length dimension of each dimension It is a non-zero constant. For the routing consistency kernel term, it can be represented using either the explicit indication form of the Kronecker function or an exponential decay form based on fragment quality; the expression for the explicit indication form of the Kronecker function is: Formula 13, where To verify the m-th dimension reconstruction error obtained in the verification phase, These are learnable regression parameters or preset hyperparameters. For online fragment quality; the expression for exponential decay based on fragment quality is: Formula 14, where and To adaptively adjust the weights, These are two different samples. n , j The online fragment quality. The quality-coupled adaptive kernel function is used to describe the correlation of Gaussian processes between different locations.

[0053] The heteroscedasticity noise term is the first term... n The heteroscedasticity noise term is obtained by weighted summation of the mean values ​​of online segment quality and reconstruction uncertainty of each input characteristic vector, and then added to the baseline noise variance. , Formula 17, The baseline noise variance, This is the noise adjustment factor. For the first n The mean uncertainty of each sample is reconstructed; the SOH fusion prediction model uses the input feature vector to output the mean SOH prediction, which is a common technique in this field and will not be elaborated here.

[0054] The SOH fusion prediction model used here is the Gaussian process regression model ARD-GPR with automatic correlation judgment. The core difference between this model and the standard ARD-GPR is that the traditional ARD-GPR only relies on the length scale of the spatial dimension to automatically determine the correlation of data. In contrast, this scheme further embeds online segment coverage quality, conditional reconstruction uncertainty, and routing type into the kernel function and noise term, thereby achieving adaptive perception and dynamic weighting of low-quality reconstruction features, high-noise real features, and low-coverage segments, which significantly improves the robustness of SOH evaluation under all operating conditions.

[0055] In the above embodiments, the SOH fusion prediction model can be any of the following: in addition to the Gaussian process regression model ARD-GPR with automatic correlation judgment, it can also be Bayesian regression, sparse Gaussian process, attention network, uncertainty neural network or integrated uncertainty model with adaptive adjustment capability of feature weights.

[0056] On the other hand, the present invention provides a battery SOH adaptive estimation method and system based on conditional reconstruction, comprising: The interval optimization module is used to acquire target battery data to construct candidate voltage windows, acquire the real health characteristics corresponding to each candidate health window, construct a characteristic vector based on the real health characteristics and send it into the SOH evaluation model, and take the candidate voltage window whose SOH estimation error meets the preset optimization criteria as the optimal information interval (OIS). The source information interval determination module is used to determine the source information interval based on collectability, consistency, noise level, and mapping relationship with the optimal information interval (OIS); it determines the coverage relationship between the current fragmented charging segment, the optimal information interval (OIS), and the source information interval, and generates a direct estimation routing instruction or a conditional reconstruction routing instruction; for the direct estimation routing instruction, it extracts the true health characteristics of the optimal information interval (OIS); The reconstruction health feature acquisition module is used to pre-build and train the cross-interval conditional reconstruction network model. When generating conditional reconstruction routing instructions, the data corresponding to the source information interval is input into the cross-interval conditional reconstruction network model. The cross-interval conditional reconstruction network model is used to generate the reconstruction curve and reconstruction uncertainty corresponding to the optimal information interval OIS, and generate the reconstruction health features of OIS. The SOH fusion prediction module is used to adaptively reduce the weight of the true health features or reconstructed health features of the optimal information interval OIS based on the improved Gaussian process regression model ARD-GPR model with automatic correlation judgment and output the SOH estimate. It also updates the OIS, source information interval, routing threshold or model parameters according to the target battery's early cycle data, subsequent running data or cloud-updated data.

[0057] To verify the technical effects of the present invention, the following embodiments are compared.

[0058] Example 1: For an NCM battery system, full-lifecycle charging data of multiple cells in the same system are acquired offline, and a global scan is performed according to a voltage window of 0.05V. For each candidate window, health features such as maximum capacity, minimum capacity, average capacity, capacity increment, and capacity standard deviation are extracted, and a leave-one-out cross-validation model is used to obtain the SOH estimation error for each window. The scan results show that the OIS of this system is between 3.80V and 3.85V. In the online phase, when only the source information range of 3.20V to 3.35V is collected for the target battery, the system determines that the current fragmented charging segment does not cover the OIS, thus triggering conditional route reconstruction. The system resamples the voltage-capacity data in the 3.20V to 3.35V range into a fixed-length sequence and inputs it into a 1D-CNN-GAP cross-range conditional reconstruction model. After extracting local curve aging features through a one-dimensional convolutional layer, and then aggregating and suppressing local noise through a global average pooling layer, the model outputs the reconstruction health features corresponding to the 3.80V to 3.85V range. Subsequently, the system inputs the reconstructed health characteristics into the ARD-GPR model calibrated after 10 effective cycles of the target battery. Experimental results show that the SOH error of direct estimation of low-voltage fragmentation segments is approximately 2.36%. After using the conditional reconstruction and ARD-GPR of this invention, the SOH estimation error is reduced to approximately 2.22%, which is close to the theoretical error of 1.28% corresponding to the true OIS characteristics.

[0059] Example 2: For an LCO battery system under dynamic driving conditions, offline scanning revealed that the OIS (Optical Information Standard) ranged from 3.50V to 3.55V. This result indicates that the optimal information range is not fixed in the high-voltage range but varies with the material system and operating conditions. For real-world segments containing dynamic polarization noise, the RMSE (Recovery Mean Squared Error) for SOH estimation using real high-noise features is approximately 0.91%. After reconstructing the features using the 1D-CNN-GAP method of this invention and fusing them with ARD-GPR, the RMSE is reduced to approximately 0.83%. This result demonstrates that the global average pooling layer and the adaptive feature weighting mechanism can reduce the impact of local noise and polarization perturbations on the estimation results.

[0060] Example 3: For LCO battery systems under strong accelerated aging conditions, offline scanning revealed that the OIS shifted to 4.10V to 4.15V. If only low-information ranges are available in fragmented charging scenarios, the direct estimation error can rise to approximately 12.59%. After applying the conditional reconstruction and ARD-GPR feature weighting of this invention, the RMSE decreased to approximately 5.75%, close to the 5.55% corresponding to the true high-information range features. This result demonstrates that this invention can suppress the amplification of SOH estimation errors in scenarios with severe polarization and information discontinuity.

[0061] Figure 2 From left to right, the datasets used are from NASA, Oxford University, and Xi'an Jiaotong University (XJTU). The horizontal axis corresponds to different datasets, and the vertical axis represents the root mean square error (RMSE). In the legend, "low" represents the RMSE of SOH estimation using features from the low-information interval, "rp" represents the RMSE of SOH estimation using the reconstructed features after reconstructing the low-information interval to the high-information interval, and "high" represents the RMSE of SOH estimation using features from the high-information interval.

[0062] Figure 3 The horizontal axis of each graph represents the number of iterations, and the vertical axis represents the selected dataset name and SOH value. In the legend, "low" represents the curve for SOH estimation using features from the low-information interval, "rp" represents the curve for SOH estimation using the reconstructed features after reconstructing the low-information interval to the high-information interval, "high" represents the curve for SOH estimation using features from the high-information interval, and "measure" represents the actual measured SOH curve.

[0063] Figure 4 The horizontal axis of each figure represents different voltage ranges, and the vertical axis represents the selected dataset name and root mean square error (RMSE). As shown in the figures, features reconstructed from low-information ranges to high-information ranges are used for SOH estimation significantly better than those directly estimated from the low-information ranges themselves.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A battery SOH adaptive estimation method based on conditional reconstruction, characterized in that, Includes the following steps: S1: Obtain target battery data to construct candidate voltage windows, obtain the real health characteristics corresponding to each candidate health window, construct a characteristic vector based on the real health characteristics and send it into the SOH evaluation model, and take the candidate voltage window whose SOH estimation error meets the preset optimization criteria as the optimal information interval (OIS). S2: During the actual operation of the target battery, the BMS acquires the current fragmented charging segment, determines the effective voltage range of the charging segment, pre-sets the source information interval LII for the fragmented charging segment, and judges the coverage relationship between the current fragmented charging segment and the optimal information interval OIS and the source information interval. When the current fragmented charging segment covers the optimal information interval OIS, a direct estimation routing instruction is generated; when the current fragmented charging segment does not cover the optimal information interval OIS but covers the source information interval LII, a conditional reconstruction routing instruction is generated; otherwise, no routing instruction is generated. S3: Pre-establish and train the cross-interval conditional reconstruction network model. When generating conditional reconstruction routing instructions, input the data corresponding to the source information interval into the cross-interval conditional reconstruction network model. Use the cross-interval conditional reconstruction network model to generate the reconstruction health features corresponding to the optimal information interval OIS. S4: Align the format of the real health features or reconstructed health features corresponding to the extracted optimal information interval OIS, and feed them into the same SOH fusion prediction model. Calculate the reliability coefficient for each health feature dimension, construct the quality coupling adaptive kernel function, the routing consistency kernel, and the heteroscedasticity noise term based on reconstruction uncertainty, so that the SOH fusion prediction model adaptively outputs the mean SOH prediction of the target battery.

2. The adaptive estimation method for battery SOH based on conditional reconstruction according to claim 1, characterized in that, Step S1 involves: First, acquiring the historical charging data and corresponding SOH tags of the target battery. The historical charging data includes the voltage-capacity sequence during the charging process, as well as current, charging time, charging temperature, SOC, charging rate, and ambient temperature operating condition data. The SOH tag is determined based on the ratio of the current maximum available capacity to the rated capacity or initial capacity. The charging voltage range of the target battery is divided into multiple candidate voltage windows. For each candidate voltage window, corresponding health features are extracted. Health features include, but are not limited to, basic capacity statistical features, curve shape features, electrochemical related features, sequence features, and operating condition fusion features. Feature vectors are constructed based on health features, for each candidate voltage window. k The corresponding voltage range is [ V k , V k + △V ], constructing a health feature vector using each candidate voltage window X k As input; A SOH evaluation model is established, with the health feature vector corresponding to a candidate voltage window as input, and the SOH estimation error corresponding to each candidate voltage window is calculated. The candidate voltage window whose SOH estimation error satisfies the preset optimization criterion is determined as the optimal information interval (OIS).

3. The adaptive estimation method for battery SOH based on conditional reconstruction according to claim 2, characterized in that, The preset optimization criteria include any of the following situations: 1) The SOH estimation error reaches a global minimum; 2) The SOH estimation error reaches a local minimum and is lower than the preset error threshold; 3) The SOH estimation error is located in a continuous low error plateau region. 4) The comprehensive objective function consisting of SOH estimation error, availability of candidate voltage window data, and online computation cost is optimized; the candidate voltage window that meets the preset optimization criteria is determined as the optimal information interval (OIS); when multiple candidate voltage windows meet the conditions at the same time, the candidate voltage window with the smallest SOH estimation error, the highest data availability, or the lowest online computation cost is selected as the optimal information interval (OIS).

4. The adaptive estimation method for battery SOH based on conditional reconstruction according to claim 2, characterized in that, In step S2, the current fragmented charging segment is compared with the optimal information interval (OIS) and the source information interval to determine their coverage relationship. The coverage relationship is determined by any of the following rules: 1) The voltage overlap ratio is greater than or equal to a preset coverage threshold; 2) The number of effective sampling points in the optimal information interval (OIS) is greater than or equal to a preset sampling point threshold; 3) The performance of the optimal information interval (OIS) after interpolation and resampling meets a preset quality condition; 4) The voltage monotonicity, current stability, temperature range, and abnormal point ratio in the optimal information interval (OIS) meet preset conditions.

5. The adaptive estimation method for battery SOH based on conditional reconstruction according to claim 4, characterized in that, The cross-interval conditional reconstruction network model described in step S3 includes an input layer, a multi-level one-dimensional convolutional feature extraction layer, a quality-gated degenerate pooling layer, and an output layer arranged sequentially. When generating conditional reconstruction routing instructions, the voltage-capacity data corresponding to the source information interval LII is resampled at equal intervals to a length of [missing information]. L in One-dimensional capacity sequence Based on a one-dimensional capacity sequence and further constructing an input tensor X in Condition vector corresponding to a one-dimensional capacity sequence c ; one-dimensional tensor X in It is fed into the input layer as input; The multi-level one-dimensional convolutional feature extraction layer includes at least two cascaded one-dimensional convolutional blocks. Each one-dimensional convolutional block sequentially includes a one-dimensional convolutional layer, a batch normalization layer, and a nonlinear activation layer. The size of the convolutional kernel and the number of channels of the convolutional kernel in the one-dimensional convolutional layers of different one-dimensional convolutional blocks increase layer by layer, with a stride of 1 for each layer. This is used to capture multi-scale aging features from local polarization to global capacity decay and output multi-channel feature maps. The batch normalization layer is used to accelerate network convergence and alleviate internal covariate shift. The nonlinear activation layer is used to enhance the network's nonlinear representation ability of multi-scale aging features. After processing by the multi-level one-dimensional convolutional feature extraction layer, layer-by-layer high-dimensional mapping and receptive field expansion are performed to output multi-channel features. The quality-gated degenerate pooling layer is used to combine multi-channel features and conditional vectors to obtain quality-gated weights at different positions. Based on the quality-gated weights, degenerate pooling features of different channel outputs are obtained. The degenerate pooling features are fed into the fully connected network and output two types of prediction results in the form of a dual-branch reconstruction head. The first type of prediction result is the reconstruction curve, and the reconstruction health characteristics are obtained based on the reconstruction curve. The second type of prediction result is the reconstruction uncertainty.

6. The adaptive estimation method for battery SOH based on conditional reconstruction according to claim 5, characterized in that, The loss function of the cross-interval conditional reconstruction network model is a multi-objective loss function, which is obtained by weighting the mean square error between the optimal information interval (OIS) reconstruction curve and the true optimal information interval (OIS) curve, the mean square error between the reconstructed health feature and the true health feature, the curve constraint loss, the SOH loss corresponding to the reconstructed health feature input SOH fusion prediction model, and the uncertainty calibration loss.

7. The adaptive estimation method for battery SOH based on conditional reconstruction according to claim 5, characterized in that, Step S4 involves the following: The SOH fusion prediction model includes a quality coupling adaptive kernel function and a heteroscedasticity noise term; an input feature vector is constructed using the true health features obtained by directly estimating the routing instructions or the reconstructed health features obtained by conditionally reconstructing the routing instructions. x , will input the first feature vector m The reliability coefficient of a dimensional feature is defined as follows: r m ; The quality-coupled adaptive kernel function includes an exponential term and a routing consistency kernel term. The exponential term is based on two different input feature vectors. The m The first dimension, the first m The length dimension of the first dimension, the first m Definition of dimensional characteristic reliability coefficient r m The routing consistency kernel term, obtained from the signal equation, is represented by either an explicit indication form using the Kronecker function or an exponentially decaying form based on segment quality; the heteroscedasticity noise term is derived by applying the first... n The result is obtained by weighted summation of the mean values ​​of online segment quality and reconstruction uncertainty of each input characteristic vector, and then adding it to the baseline noise variance. The SOH fusion prediction model is used to output the SOH prediction mean.

8. The adaptive estimation method for battery SOH based on conditional reconstruction according to claim 1, characterized in that, The SOH evaluation model is one or a combination of different models, including Gaussian process regression model, support vector regression model, random forest model, gradient boosting model, and neural network model.

9. The adaptive estimation method for battery SOH based on conditional reconstruction according to claim 1, characterized in that, The SOH fusion prediction model is any one of the following: Gaussian process regression model ARD-GPR with automatic correlation judgment, Bayesian regression with adaptive adjustment capability of feature weights, sparse Gaussian process, attention network, uncertainty neural network, or integrated uncertainty model.

10. A battery SOH adaptive estimation system based on conditional reconstruction, used to implement the method of any one of claims 1-9, characterized in that, include: The interval optimization module is used to acquire target battery data to construct candidate voltage windows, acquire the real health characteristics corresponding to each candidate health window, construct a characteristic vector based on the real health characteristics and send it into the SOH evaluation model, and take the candidate voltage window whose SOH estimation error meets the preset optimization criteria as the optimal information interval (OIS). The source information interval determination module is used to determine the source information interval based on collectability, consistency, noise level, and mapping relationship with the optimal information interval (OIS); it determines the coverage relationship between the current fragmented charging segment, the optimal information interval (OIS), and the source information interval, and generates a direct estimation routing instruction or a conditional reconstruction routing instruction; for the direct estimation routing instruction, it extracts the true health characteristics of the optimal information interval (OIS); The reconstruction health feature acquisition module is used to pre-build and train the cross-interval conditional reconstruction network model. When generating conditional reconstruction routing instructions, the data corresponding to the source information interval is input into the cross-interval conditional reconstruction network model. The cross-interval conditional reconstruction network model is used to generate the reconstruction curve and reconstruction uncertainty corresponding to the optimal information interval OIS, and generate the reconstruction health features of OIS. The SOH fusion prediction module is used to adaptively reduce the weight of the true health features or reconstructed health features of the optimal information interval OIS based on the improved Gaussian process regression model ARD-GPR model with automatic correlation judgment and output the SOH estimate. It also updates the OIS, source information interval, routing threshold or model parameters according to the target battery's early cycle data, subsequent running data or cloud-updated data.

Images