Battery state prediction method and system based on whole vehicle control big data and double-mode network

Abstract

The application discloses a battery state prediction method and system based on whole vehicle control big data and a dual-mode network, constructs a historical big data set fusing battery charging and discharging time sequence signals, whole vehicle controller logic state encoding and dynamic scheduling labels, realizes joint modeling of battery operation states and vehicle control strategies through a time-varying coupling state space model in the dual-mode network, extracts dynamic scheduling influence characteristics and battery irreversible aging characteristics. A latent space orthogonal decoupling mechanism is introduced to separate transient scheduling disturbance and real aging information, and a causally isolated health state representation is obtained. In the prediction stage, a correction flow generation model based on ordinary differential equations is adopted to learn the state distribution evolution law, and the probability prediction of SOH and RUL is realized. Meanwhile, the model prediction precision, reliability and uncertainty quantification ability are improved by combining physical monotonicity constraints and multi-task collaborative constraints. The method can fully utilize whole vehicle control big data, and realize accurate prediction of power battery health state and remaining life under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of new energy vehicle technology, and relates to battery state prediction technology in the battery management system of hybrid electric vehicles. Specifically, it relates to a battery state prediction method and system based on vehicle control big data and a dual-modal network. Background Technology

[0002] As a crucial component of new energy vehicles, the accurate prediction of the health status and remaining lifespan of the power batteries in hybrid electric vehicles is essential for ensuring safe vehicle operation, optimizing energy management strategies, and reducing maintenance costs. However, under actual vehicle operating conditions, the dynamic aging characteristics of the battery are easily masked by the frequent scheduling of the vehicle controller's energy management strategy, making it difficult for traditional prediction methods to guarantee accuracy.

[0003] Currently, battery state prediction methods can be mainly classified into the following categories: I. Equivalent Circuit Model-Based Method: This method establishes an equivalent circuit model of the battery and uses circuit components such as resistors and capacitors to simulate the battery's electrochemical characteristics. However, the parameter identification of the equivalent circuit model depends on specific operating conditions. In complex operating conditions under HCU dynamic scheduling, the model parameters are difficult to identify accurately, leading to a decrease in prediction accuracy.

[0004] II. Electrochemical Model-Based Approach: This approach establishes a system of partial differential equations to describe the dynamic behavior of a battery based on its internal electrochemical reaction mechanism. While electrochemical models possess strong physical significance, their high complexity, computational cost, and difficulty in obtaining model parameters limit their application in practical automotive scenarios.

[0005] III. Data-Driven Machine Learning Methods: In recent years, with the development of deep learning technology, battery state prediction methods based on recurrent neural networks such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs) have been widely studied. However, these methods are mainly trained on data under ideal laboratory conditions, making it difficult to adapt to the high-frequency noise and complex load patterns generated by HCU dynamic scheduling in actual automotive applications.

[0006] Based on this, the existing technology has the following problems: 1. There are significant differences between ideal laboratory conditions and actual vehicle conditions. Traditional methods are mostly based on data from ideal conditions such as constant current charging and discharging for modeling. However, in actual vehicle conditions, the HCU dynamically schedules the charging and discharging power of the battery according to factors such as driving needs, battery status, and engine status, resulting in high-frequency fluctuations in battery load. Traditional methods are unable to accurately capture this dynamic characteristic.

[0007] 2. HCU dynamic scheduling masks the dynamic aging characteristics of batteries. Under the energy management strategy of HCU, the charging and discharging process of batteries is frequently interfered with, which causes the true aging characteristics of batteries to be masked in the complex load mode. Traditional feature extraction methods are difficult to extract the key features that have the most significant impact on aging from a large range of historical data.

[0008] 3. High-frequency noise interference feature extraction: The high-frequency load fluctuations generated by HCU dynamic scheduling manifest as high-frequency noise in time series data. Traditional recurrent neural networks are easily affected by noise interference when processing long time series data, resulting in inaccurate feature extraction.

[0009] 4. Traditional methods cannot quantify prediction uncertainty. Existing point prediction methods only output a single prediction value and cannot provide a confidence interval for the prediction result. This makes it difficult to assess the reliability of the prediction result and limits its application in practical decision-making.

[0010] 5. Single-task prediction is inefficient and information utilization is insufficient. Traditional methods usually model and predict SOH or RUL separately, ignoring the inherent relationship between the two, which makes the model unable to fully utilize the information in the data and results in low prediction efficiency.

[0011] Therefore, there is an urgent need for a battery state prediction method that can adapt to the dynamic scheduling conditions of HCU, effectively remove high-frequency noise, quantify prediction uncertainties, and achieve multi-task collaborative prediction. Summary of the Invention

[0012] The purpose of this invention is to provide a battery state prediction method and system based on vehicle control big data and a dual-modal network to solve the problems existing in the current system. This invention can effectively address the problems of low battery state prediction accuracy and inability to quantify prediction uncertainty in hybrid electric vehicles under HCU dynamic scheduling conditions.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: A battery state prediction method based on vehicle control big data and a dual-modal network includes the following steps: Step 1: Construct a battery history big data set under the dynamic scheduling of the vehicle controller. The battery history big data set includes battery charging and discharging timing signals, vehicle controller logic state codes, and dynamic scheduling state labels represented by the logic state codes. Step 2: Input the battery historical big data set into the feature extraction layer of the dual-modal network. The feature extraction layer adopts the time-varying coupled state-space model TV-SSM to couple and model the charging and discharging time sequence signal of the battery terminal with the logic state encoding of the vehicle controller, so as to obtain the local features representing the temporary relationship caused by the dynamic scheduling of the vehicle and the reverse aging features representing the irreversible degradation process of the battery. Step 3: Introduce a latent space orthogonal decoupling mechanism in the feature extraction layer. Based on feature reprojection, project the local features of the temporary relationship to a latent subspace orthogonal to the inverse aging features. Based on orthogonal constraints, suppress the mixing of dynamic scheduling state labels into the inverse aging features to form an aging state representation after causal isolation. Step 4: Input the aging state characterization after causal isolation into the prediction generation layer of the dual-modal network. The prediction generation layer adopts the modified flow generation model based on the ordinary differential equation ODE to learn the deterministic velocity field from the prior distribution to the target distribution of battery state, and integrates along the deterministic velocity field to generate the predicted distribution of battery health state SOH and remaining service life RUL. Step 5: Train the dual-modal network and introduce physical monotonicity constraints and multi-task collaborative constraints during training; wherein, the physical monotonicity constraints are used to constrain the battery health state (SOH) prediction value to meet the irreversible law of battery aging; the multi-task collaborative constraints are used to enable the battery health state (SOH) prediction task and the remaining useful life (RUL) prediction task to share the aging state representation after causal isolation. Step 6: Based on the bimodal network trained by the physical monotonicity constraint and multi-task collaborative constraint, perform statistical processing on the predicted distribution and output the predicted battery health state (SOH), the predicted remaining service life (RUL), and their uncertainty intervals.

[0014] Furthermore, the construction of the battery historical big data set under the dynamic scheduling of the vehicle controller includes: The battery charging and discharging timing signals of a hybrid electric vehicle under actual operating conditions are collected. The battery charging and discharging timing signals include at least one of the following: battery current, battery voltage, battery temperature, state of charge (SOC), and charging and discharging power. The logic state code of the vehicle controller is collected, and the logic state code of the vehicle controller includes at least one of the following: drive mode, brake energy recovery mode, engine start-stop state, charge-discharge power distribution state and thermal management state. A dynamic scheduling status label is generated based on the vehicle controller logic state code, and the dynamic scheduling status label is aligned with the battery charging and discharging timing signal at the corresponding time. The aligned data is normalized, outlier removed, missing value imputed, and divided into sliding window segments to form the battery history dataset.

[0015] Furthermore, the time-varying coupled state-space model TV-SSM in step 2 includes a state transition module, an input coupling module, and a time-varying parameter generation module; The state transition module is used to update the potential state of the battery. The input coupling module is used to map the battery charging and discharging timing signals and the vehicle controller logic state encoding to the same potential state space. The time-varying parameter generation module generates at least one time-varying parameter among the state transition matrix, input matrix, or gating coefficients based on the vehicle controller's logic state encoding, so that the state evolution of the TV-SSM adapts to the dynamic scheduling state of the vehicle controller.

[0016] Furthermore, the state evolution of the TV-SSM satisfies the following form:

[0017]

[0018] in, Represents the potential state at time t. This represents the battery charging / discharging timing input at time t. This represents the logic state code of the vehicle controller at time t. , , , These are the time-varying parameters generated by the time-varying parameter generation module. Representing output features; by introducing The terms and their time-varying parameters enable the transient response caused by the vehicle controller scheduling to be explicitly characterized in the state space.

[0019] Furthermore, the potential space orthogonal decoupling mechanism in step 3 includes an eigenvalue decomposition submodule, an eigenvalue reprojection submodule, and an orthogonal constraint submodule; The feature decomposition submodule decomposes the TV-SSM output features into temporary relational local features and inverse aging features; The feature reprojection submodule is used to reproject the local features of the temporary relationship to the orthogonal complement space of the inverse aging feature. The orthogonal complement space constitutes a potential subspace orthogonal to the inverse aging feature, and an orthogonal reprojection feature is obtained. The orthogonal constraint submodule is used to construct an orthogonal constraint loss based on the inner product or correlation between the temporary relationship local features and the reverse aging features, so as to suppress feature mixing between the temporary relationship local features and the reverse aging features.

[0020] Furthermore, the orthogonal reprojection feature satisfies the following form:

[0021] Among them, z d z represents a local feature of a temporary relation. a Indicates reverse aging characteristics, z d⊥ represents the orthogonal reprojection feature, and ε is a stability coefficient to prevent the denominator from being zero; by subtracting the projection component of the temporary relation local feature in the direction of the reverse aging feature, the mixing of the vehicle controller dynamic scheduling feature with the reverse aging feature is suppressed.

[0022] Furthermore, the modified flow generation model in step 4 includes a conditional coding module, a velocity field prediction module, and an ODE integration module; The conditional encoding module is used to encode the aging state characterization after causal isolation to obtain conditional features. The velocity field prediction module is used to predict based on prior variables. Target variable Continuous time variables and the conditional features Predicting the velocity field The ODE integration module is used to integrate from the prior distribution to the target distribution of battery state along the velocity field, generating predicted distributions of battery state of health (SOH) and remaining useful life (RUL). During the training of the bimodal network, the Rectified Flow generative model aims to minimize the flow matching loss, ensuring that the velocity field output by the velocity field prediction module fits the linear path velocity from the prior variable to the battery state target variable; the flow matching loss satisfies the following form:

[0023] in, This represents the prior variable obtained by sampling from the prior distribution. The target variable represents the battery state. Indicates by and The intermediate variable obtained by linear interpolation, where t is a continuous time variable between 0 and 1, and c represents the conditional feature. This represents the velocity field output by the velocity field prediction module. Let E represent the flow matching loss, and E represent the mathematical expectation.

[0024] Furthermore, the total training loss function of the bimodal network includes SOH prediction loss, RUL prediction loss, modified flow-flow matching loss, latent space orthogonal constraint loss, and SOH monotonicity constraint loss; Specifically, the SOH prediction loss is used to constrain the deviation between the predicted SOH value and the actual SOH value of the battery health state; the RUL prediction loss is used to constrain the deviation between the predicted RUL value and the actual RUL value of the remaining useful life; the modified flow matching loss is used to train the velocity field prediction module; the latent space orthogonal constraint loss is used to penalize the correlation between temporary relational local features and reverse aging features; and the SOH monotonicity constraint loss is used to penalize the non-physical increase of SOH at adjacent prediction times.

[0025] Furthermore, the uncertainty interval in step 6 is obtained by sampling the prior variables of the modified flow generation model multiple times and integrating along the same ODE velocity field to obtain multiple prediction samples, and then calculating the uncertainty intervals of SOH and RUL based on the quantiles of the multiple prediction samples. The uncertainty interval, along with the SOH and RUL prediction values, are sent to the vehicle controller via the vehicle communication interface for charging and discharging power allocation, braking energy recovery strategy adjustment, thermal management strategy adjustment, or battery maintenance reminders.

[0026] A battery state prediction system based on vehicle control big data and a dual-modal network includes: The first module is used to construct a battery historical big data set under the dynamic scheduling of the vehicle controller. The battery historical big data set includes battery charging and discharging timing signals, vehicle controller logic state codes, and dynamic scheduling state labels represented by the logic state codes. The second module is used to input the battery historical big data into the feature extraction layer of the dual-modal network. The feature extraction layer adopts the time-varying coupled state-space model TV-SSM to couple and model the battery charging and discharging time sequence signal with the vehicle controller logic state encoding to obtain local features that characterize the temporary relationship caused by the dynamic scheduling of the vehicle and reverse aging features that characterize the irreversible degradation process of the battery. The third module is used to introduce a latent space orthogonal decoupling mechanism in the feature extraction layer. Based on feature reprojection, the local features of the temporary relationship are projected to a latent subspace orthogonal to the inverse aging features. Based on orthogonal constraints, the mixing of dynamic scheduling state labels into the inverse aging features is suppressed, so as to form an aging state representation after causal isolation. The fourth module is used to input the aging state characterization after causal isolation into the prediction generation layer of the dual-modal network. The prediction generation layer adopts the modified flow generation model based on the ordinary differential equation ODE, learns the deterministic velocity field from the prior distribution to the target distribution of the battery state, and generates the predicted distribution of the battery health state SOH and remaining service life RUL by integrating along the deterministic velocity field. The fifth module is used to train the dual-modal network and introduce physical monotonicity constraints and multi-task collaborative constraints during training. The physical monotonicity constraints are used to constrain the predicted state of battery health (SOH) to meet the irreversible law of battery aging. The multi-task collaborative constraints are used to enable the SOH prediction task and the remaining useful life (RUL) prediction task to share the aging state representation after causal isolation. The sixth module is used to perform statistical processing on the predicted distribution based on the bimodal network trained by the physical monotonicity constraint and multi-task collaborative constraint, and output the predicted battery health state (SOH), the predicted remaining service life (RUL), and their uncertainty intervals.

[0027] Compared with the prior art, the present invention has the following beneficial technical effects: 1. This invention uses a time-varying coupled state-space model (TV-SSM) to jointly model the battery charging and discharging timing signals and the HCU (vehicle controller) logic state encoding. This enables explicit characterization of the impact of dynamic scheduling states such as pure electric, hybrid, engine direct drive, and regenerative braking on the battery response, thereby improving the feature expression capability under complex real-world vehicle conditions.

[0028] 2. This invention introduces a potential space orthogonal decoupling mechanism to separate the temporary local features caused by frequent energy scheduling of the vehicle controller from the reverse aging features related to irreversible degradation such as battery capacity decay and internal resistance growth. By suppressing the mixing between the two types of features through feature reprojection and orthogonal constraints, the impact of dynamic scheduling interference on the characterization of the real aging state is reduced.

[0029] 3. This invention employs a modified flow generation model based on the ordinary differential equation ODE to learn a deterministic velocity field from the prior distribution to the target distribution of the battery state, and generates predicted distributions of SOH and RUL by integrating along the deterministic velocity field, thereby improving the stability and reproducibility of the predicted distribution generation process.

[0030] 4. The Rectified Flow prediction generation layer of the present invention can be solved step-by-step or in a few steps by the ODE integration module, which reduces the amount of iterative computation in the prediction generation process and thus reduces the online prediction latency.

[0031] 5. This invention introduces physical monotonicity constraints to constrain the non-physical increase of SOH at adjacent prediction times, so that the predicted SOH value of battery health state conforms to the irreversible law of battery aging, thereby improving the physical rationality and reliability of the prediction results.

[0032] 6. This invention introduces multi-task collaborative constraints, enabling the State of Health (SOH) prediction task and the Remaining Useful Life (RUL) prediction task to share aging state characterization after causal isolation. Collaborative prediction is achieved by utilizing the correlation between SOH and RUL, and decision support is provided for optimizing the vehicle controller's energy management strategy by incorporating uncertainty intervals. Attached Figure Description The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0033] Figure 1 This is a flowchart of the overall method of the present invention; Figure 2 This is a diagram of the bimodal network architecture based on flow matching state space and orthogonal decoupling of the present invention; Figure 3 This is a detailed structural diagram of the TV-SSM and the orthogonally decoupled feature extraction layer of the latent space in this invention; Figure 4 This is a schematic diagram of the SOH prediction results and confidence intervals of the present invention; Figure 5 This is a schematic diagram of the RUL prediction results and confidence intervals of the present invention; Figure 6 This is a comparison chart of the prediction accuracy of the method of this invention and existing methods; Figure 7 This is a verification diagram of the uncertainty quantification effect of the present invention. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0035] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0036] Example 1 This invention provides a battery state prediction method based on vehicle control big data and a dual-modal network, comprising the following steps: Step 1: Construct a battery history big data set under the dynamic scheduling of the vehicle controller. The battery history big data set includes battery charging and discharging timing signals, vehicle controller logic state codes, and dynamic scheduling state labels represented by the logic state codes. Step 2: Input the battery historical big data set into the feature extraction layer of the dual-modal network. The feature extraction layer adopts the time-varying coupled state-space model TV-SSM to couple and model the battery charging and discharging time sequence signal with the vehicle controller logic state encoding to obtain local features that characterize the temporary relationship caused by the dynamic scheduling of the vehicle and reverse aging features that characterize the irreversible degradation process of the battery. Step 3: Introduce a latent space orthogonal decoupling mechanism in the feature extraction layer. Based on feature reprojection, project the local features of the temporary relationship to a latent subspace orthogonal to the inverse aging features. Based on orthogonal constraints, suppress the mixing of dynamic scheduling state labels into the inverse aging features to form an aging state representation after causal isolation. Step 4: Input the aging state characterization after causal isolation into the prediction generation layer of the dual-modal network. The prediction generation layer adopts the modified flow generation model based on the ordinary differential equation ODE to learn the deterministic velocity field from the prior distribution to the target distribution of battery state, and integrates along the deterministic velocity field to generate the predicted distribution of battery health state SOH and remaining service life RUL. Step 5: Train the dual-modal network, and introduce physical monotonicity constraints and multi-task collaborative constraints during the training process; wherein, the physical monotonicity constraints are used to constrain the battery health state (SOH) prediction value to meet the irreversible law of battery aging; the multi-task collaborative constraints are used to enable the battery health state (SOH) prediction task and the remaining useful life (RUL) prediction task to share the aging state representation after causal isolation. Step 6: Based on the bimodal network trained by the physical monotonicity constraint and multi-task collaborative constraint, perform statistical processing on the predicted distribution and output the predicted battery health state (SOH), the predicted remaining service life (RUL), and their uncertainty intervals.

[0037] Through the above steps, TV-SSM can explicitly model the time-varying coupling relationship between the dynamic scheduling state of the vehicle controller and the charging and discharging timing signal of the battery. The latent space orthogonal decoupling mechanism can separate the temporary local features caused by dynamic scheduling from the reverse aging features. The prediction generation layer can generate the predicted distribution of battery health state (SOH) and remaining service life (RUL) based on the aging state characterization after causal isolation, thereby reducing the interference of the vehicle controller's dynamic scheduling on battery state prediction.

[0038] Example 2 See Figure 1 The present invention presents a flowchart of its overall method, proposing a collaborative prediction method for battery state based on historical big data from a vehicle controller and a dual-modal network based on flow matching state space and orthogonal decoupling. The method first constructs a historical battery dataset under dynamic scheduling of the vehicle controller; then, it uses TV-SSM to couple and model the battery charging / discharging timing signals with the vehicle controller's logic state encoding to obtain temporary relationship local features and inverse aging features; next, it forms an aging state representation after causal isolation through a latent space orthogonal decoupling mechanism; subsequently, it uses a modified flow generation model based on ordinary differential equations (ODE) to generate predicted distributions of battery health state (SOH) and remaining useful life (RUL); then, it trains the dual-modal network through physical monotonicity constraints and multi-task collaborative constraints; finally, it outputs predicted SOH values, predicted RUL values, and their uncertainty intervals based on the predicted distributions.

[0039] The specific steps are as follows: Step 1: Construct a historical battery dataset under dynamic scheduling by the vehicle controller. Collect the battery charging and discharging timing signals of the hybrid vehicle under actual operating conditions. These signals include at least one of the following: battery current I(t), battery voltage V(t), battery temperature T(t), state of charge (SOC), and charging / discharging power. The sampling frequency is 1 Hz. Collect the vehicle controller logic state code S. HCU(t), whereby the vehicle controller logic state code is used to characterize the energy management strategy of the vehicle controller under different operating modes, including at least one of drive mode, regenerative braking mode, engine start-stop state, charge / discharge power distribution state, and thermal management state. The drive mode includes at least one of pure electric mode, series mode, parallel mode, and engine direct drive mode. A dynamic scheduling state label is generated based on the vehicle controller logic state code, and the dynamic scheduling state label is aligned with the corresponding battery charge / discharge timing signal. The aligned data is preprocessed, including data normalization, outlier removal, and missing value imputation. Data normalization uses the Min-Max normalization method, expressed as follows:

[0040] In the formula, x represents the original data, x norm For the normalized data, x min and x max These represent the minimum and maximum values ​​of the data, respectively. Outlier removal follows the 3σ principle, meaning data points differing from the mean by more than three times the standard deviation are removed. Missing value imputation uses linear interpolation. A sliding window method is used to divide the preprocessed data, with a window length of L = 1000 time steps and a sliding step size of Δ = 100 time steps, resulting in the sample sequence constituting the battery history dataset. Further, the battery history dataset is divided into a training sample set and a test sample set, with a ratio of 8:2.

[0041] Step 2: Design a feature extraction layer based on the orthogonal decoupling mechanism between TV-SSM and latent space.

[0042] The battery historical dataset is input into the feature extraction layer of the dual-modal network. The battery charging and discharging timing signal and the vehicle controller logic state encoding are coupled and modeled through TV-SSM to obtain local features that characterize the temporary relationship caused by the dynamic scheduling of the vehicle and reverse aging features that characterize the irreversible degradation process of the battery.

[0043] Step 3: Design a potential space orthogonal decoupling mechanism.

[0044] In this embodiment, the latent spatial orthogonal decoupling mechanism is positioned between the feature output of the time-varying coupled state-space model TV-SSM and the conditional input of the modified flow generation model based on ordinary differential equations (ODEs). This mechanism separates the temporary local features caused by the dynamic energy scheduling of the vehicle controller (HCU) from the reverse aging features related to the irreversible battery degradation process. The purpose of this latent spatial orthogonal decoupling mechanism is to ensure that subsequent predictions of the battery's state of health (SOH) and remaining useful life (RUL) are no longer directly dominated by short-term voltage, current, and temperature disturbances caused by frequent switching between pure electric mode, series mode, parallel mode, and engine direct drive mode. Instead, predictions are based on stable latent characteristics that are causally related to the battery aging process.

[0045] I. Physical Logic of Dynamic Disturbance Scheduling During actual operation, hybrid vehicles frequently switch between various operating modes, including pure electric drive, hybrid drive, engine direct drive, charging while driving, and regenerative braking, based on driving demands, engine efficiency range, battery state of charge (SOC), vehicle speed, regenerative braking requirements, and thermal management status. These mode switching patterns cause rapid changes in the direction and amplitude of the battery current, the voltage response, and the rate of temperature change, resulting in significant transient polarization.

[0046] The transient polarization phenomenon is mainly manifested as follows: within a short timescale, the battery terminal voltage shifts rapidly due to ohmic internal resistance, electrochemical polarization, and concentration polarization; under conditions such as energy recovery or engine intervention, the current signal exhibits high-frequency reversal or pulses; and when the thermal management strategy is switched, there is a lag between the battery temperature response and the actual aging state. These phenomena are not irreversible aging results caused by battery capacity decay or long-term increase in internal resistance, but rather reversible or short-term disturbances triggered by the HCU energy dispatch strategy.

[0047] If the aforementioned battery charging and discharging timing signals are directly input into the prediction generation layer of the dual-modal network, the prediction generation layer is easily affected by short-term voltage, current, and temperature disturbances caused by frequent energy scheduling by the vehicle controller (HCU), causing the battery state prediction results to deviate from the true trend corresponding to the irreversible battery degradation process. To reduce the impact of such dynamic scheduling disturbances on the prediction results, this embodiment uses a latent space orthogonal decoupling mechanism to separate the temporary local features caused by the dynamic scheduling of the vehicle controller (HCU) from the reverse aging features related to the irreversible battery degradation process, and suppresses the mixing between the two types of features through feature reprojection and orthogonal constraints.

[0048] II. Characteristic Decomposition Let the latent feature output by TV-SSM in the t-th time window be:

[0049] in, This indicates the battery charging and discharging timing signal, which includes at least one of the following: current, voltage, temperature, and SOC. The HCU logic state code represents at least one of the following: pure electric, hybrid, direct drive, regenerative braking, and driving charging. This indicates a potential feature that integrates battery response and HCU scheduling state.

[0050] In order to The transient polarization perturbation is separated from the actual aging information, and a polarization feature mapping function is set. Mapping function for aging characteristics :

[0051]

[0052] in, This represents the polarization candidate features caused by HCU dynamic scheduling. These represent candidate aging features associated with irreversible battery degradation. Furthermore, [the text abruptly ends here]. Zhang Cheng's potential subspace is denoted as polarization space. ,Will Zhang Cheng's potential subspace is denoted as the aging space. :

[0053]

[0054] in, Used to characterize temporary local features of relationships caused by dynamic scheduling such as mode switching, power allocation, and regenerative braking; It is used to characterize long-term irreversible aging features such as capacity decay, internal resistance growth, and decrease in available energy.

[0055] III. Feature Reprojection To avoid polarization space With aging space Feature aliasing occurs, and this implementation introduces a Schmitt orthogonalization constraint in the intermediate layer of the network. Specifically, aging candidate features are first used... As an aging basis vector, and then normalized:

[0056] in, Indicates aging space unit basis vectors, Describing the L2 norm, To prevent the stability coefficient from being zero.

[0057] Then, polarization candidate features are calculated. In aging baseline vector Projection components in the direction:

[0058] Where · represents the vector inner product.

[0059] Then from polarization candidate features Subtract its aging space Projecting along the direction yields polarization characteristics orthogonal to the aging space:

[0060] Right now:

[0061] thus, satisfy:

[0062] When the aging space consists of multiple basis vectors, let the aging space be... Then the projection of the polarization candidate features onto the aging space can be expressed as:

[0063] because For an orthogonal basis, we have Therefore, the above formula can be simplified to:

[0064] The corresponding orthogonal polarization characteristics are:

[0065] Through the above Schmidt orthogonalization process, the polarization space Medium and aging space Collinear or correlated components are stripped away, primarily retaining transient polarization perturbations caused by HCU dynamic scheduling. Meanwhile, aging characteristics are... The modified flow generation model is used as the main conditional feature input for subsequent SOH and RUL predictions, thereby enabling the prediction generation layer to focus on the real aging trend rather than short-term scheduling disturbances.

[0066] In another implementation, reprojection can also be performed in reverse, i.e., from aging candidate features. Subtracting its polarization space Projecting along the direction yields the depolarized aging characteristics:

[0067] in:

[0068] For polarization space The orthogonal basis.

[0069] IV. Orthogonal Constraints An orthogonal constraint loss is constructed based on the inner product or correlation between temporary relation local features and inverse aging features to suppress feature mixing between the temporary relation local features and the inverse aging features. In this embodiment, the orthogonal constraint loss is denoted as Lorth, and for a single sample, the orthogonal constraint loss can be defined as:

[0070] In specific implementation, when and When represented by the polarization feature matrix P and the aging feature matrix A of the batch samples, respectively, the orthogonality penalty term can be written as:

[0071] or:

[0072] in, , B represents the batch sample size. Denotes the Frobenius norm. This represents the polarization feature vector of the i-th sample. Let represent the aging feature vector of the i-th sample.

[0073] To avoid the vector magnitude affecting the orthogonality determination, a normalized inner product form can also be used:

[0074] The smaller this loss term, the closer the angle between polarization features and aging features is to 90 degrees, and the lower the correlation between the two types of features in the latent space. During training, by minimizing... The network is forced to classify the transient polarization noise generated by HCU dynamic scheduling into the polarization space. Irreversible changes such as capacity decay and internal resistance increase are attributed to the aging space. This allows for feature stripping.

[0075] V. Integration with the training process The aforementioned orthogonal constraint loss is used to constrain the correlation between local features of temporary relationships and reverse aging features, and can be used as part of the total training loss function in the subsequent bimodal network training process, together with SOH prediction loss, RUL prediction loss, modified flow matching loss and SOH monotonicity constraint loss for model training.

[0076] Step 4: Design a Rectified Flow Prediction Generation Layer Based on Ordinary Differential Equations (ODE). In this embodiment, the prediction generation layer adopts a Rectified Flow generation model based on Ordinary Differential Equations (ODE). The prediction generation layer receives the aging state characterization after causal isolation output by the time-varying coupled state-space model TV-SSM and the latent space orthogonal decoupling mechanism, and generates the predicted distributions of battery health state (SOH) and remaining useful life (RUL) by integrating along a deterministic velocity field under the conditional constraints of this characterization. The prediction generation layer includes a conditional encoding module, a velocity field prediction module, and an ODE integration module.

[0077] I. Predictive Generation Layer In this embodiment, the Rectified Flow generation model includes a conditional encoding module, a velocity field prediction module, and an ODE integration module. The conditional encoding module encodes the aging state characterization after causal isolation to obtain conditional feature c; the velocity field prediction module predicts the velocity field based on prior variables, target variables, continuous-time variables, and the conditional feature c; the ODE integration module integrates along the velocity field from the prior distribution to the target battery state distribution, generating predicted distributions for the battery state of health (SOH) and remaining useful life (RUL).

[0078] II. Linear mapping trajectory from standard Gaussian noise to target battery state distribution Let X0 represent the prior variable obtained by sampling from a standard Gaussian distribution:

[0079] Let X1 represent the target variable for battery state, where X1 includes the target value for State of Health (SOH), the target value for Residual Battery Usage (RUL), or a joint target vector composed of SOH and RUL:

[0080] The Rectified Flow constructs a flow within the interval [0,1] for continuous time t. arrive Linear interpolation path:

[0081] in, This represents the intermediate state at time t. When t=0, When t=1, From the above linear interpolation path, its time derivative can be obtained as:

[0082] Equivalent land can be written as:

[0083] This formula represents the ideal linear mapping trajectory from standard Gaussian noise to the target distribution of battery state. Since X1 is unknown during the inference phase, this implementation trains a velocity field network. To approximate this ideal speed, where c represents the causal isolation aging state characterization output by TV-SSM and orthogonal decoupling mechanism, i.e., the conditional characteristic.

[0084] The ODE form of the prediction generation layer is:

[0085] Where c represents the conditional feature. By integrating the above ODE from t=0 to t=1, the predicted battery state variables can be obtained:

[0086] III. Training Objectives for the Velocity Field During the training phase, a target real battery state is given. and samples obtained from the standard Gaussian distribution Construct intermediate variables at random time t:

[0087] The ideal speed is:

[0088] Velocity Field Network The training objective is to fit this ideal velocity, and the flow matching loss is defined as:

[0089] Where E represents the expected value, Let L2 norm be represented. This is achieved by minimizing... The predictive generation layer learns to map the standard Gaussian prior variables to the target distribution of battery state along an approximately straight-line trajectory under the constraint of the conditional feature c.

[0090] When SOH and RUL are used in a joint prediction method It can be represented as a two-dimensional or multi-dimensional joint state vector:

[0091] Correspondingly, the velocity field network outputs a velocity vector of the same dimension:

[0092] Therefore, Rectified Flow can simultaneously model the correlation between SOH and RUL and generate prediction results that satisfy the joint distribution relationship between the two.

[0093] IV. Conditional Constraint Methods To ensure that the prediction generation layer focuses on aging characteristics related to the irreversible degradation process of the battery and reduces the impact of dynamic scheduling disturbances in the vehicle control unit (HCU), this implementation uses the causally isolated aging state representation as the basis for conditional input and generates conditional feature c through a conditional coding network. The conditional feature c can be obtained in the following way: c = g(z a ) or: c = g(z a , z d ⊥) Among them, z a Indicates reverse aging characteristics, z d ⊥ represents the orthogonal reprojection feature, g represents the conditional coding network, and c represents the conditional feature of the input corrected flow generative model. Preferably, z a As the primary conditional input, the velocity field v θ It is mainly controlled by the characteristics related to irreversible battery degradation. In the specific network structure, the continuous time variable t can be encoded into e through a time embedding layer. t The conditional feature c is encoded into e through a conditional coding layer. c Then, with the intermediate state x t After stitching, input the data into the velocity field prediction module:

[0094] Here, MLP stands for Multilayer Perceptron. For scenarios requiring stronger expressive power, lightweight residual networks or gated networks can also be used as velocity field networks.

[0095] V. Reasoning Process During the inference phase, initial variables are first obtained by sampling from a standard Gaussian distribution. Then, conditional features c are obtained based on the battery-side timing signals and HCU logic state codes collected from the current vehicle. Subsequently, the ODE is solved using the trained velocity field network:

[0096] Integrating from t=0 to t=1 yields the predictions for SOH and RUL.

[0097] In low-computing-power automotive environments, the Euler solver can be used for approximate integration. If an N-step Euler solution is used, the updated formula is:

[0098] in,

[0099] When the Rectified Flow learns a near-straight-line mapping trajectory, single-step Euler inference can be used, i.e., N=1:

[0100] Alternatively, fewer steps of Euler inference can be used, such as N=2 to N=5, to achieve a balance between inference speed and prediction accuracy.

[0101] VI. Project Advantages Compared to traditional DDPM, the Rectified Flow prediction generation layer has significant engineering advantages in the automotive BMS microcontroller environment.

[0102] First, Rectified Flow does not require multiple rounds of random inverse denoising sampling. Instead, it achieves a direct mapping from standard Gaussian noise to the SOH / RUL target distribution through the ODE velocity field, thus significantly reducing the number of inference steps.

[0103] Secondly, the inference process of Rectified Flow can be completed using the Euler solver, and the computation mainly consists of forward network computation and simple additive updates, making it suitable for deployment in BMS microcontrollers with limited computing power, storage, and power consumption.

[0104] Furthermore, under single-step or very few-step inference conditions, Rectified Flow can reduce on-board online prediction latency, enabling SOH and RUL prediction results to be fed back to the vehicle controller more promptly for charging and discharging power allocation, regenerative braking strategy adjustment, thermal management strategy adjustment, and battery maintenance reminders.

[0105] Finally, since Rectified Flow is generated using deterministic ODE trajectories, its online inference results have better stability and reproducibility compared to models that rely on random inverse denoising, making it more suitable for battery state prediction tasks with high safety and real-time requirements during actual vehicle operation.

[0106] Through the aforementioned prediction generation layer, this implementation significantly reduces the inference complexity of the vehicle-side while maintaining generative distributed prediction capabilities, enabling SOH and RUL collaborative prediction based on big data of vehicle control to be applicable to engineering deployment scenarios with low computing power and low latency.

[0107] Step 5: Train the bimodal network and introduce physical monotonicity constraints and multi-task collaborative constraints. In this embodiment, the physical monotonicity constraint is used to ensure that the predicted state of health (SOH) of the battery meets the irreversible law of battery aging. Since the battery aging process usually manifests as a gradual decrease in SOH over time, when the predicted SOH value at adjacent prediction times shows a non-physical increase, the physical monotonicity constraint loss incurs a penalty. The SOH monotonicity constraint loss can be expressed as:

[0108] In the formula, N_seq is the length of the prediction sequence, SOH_t^pred is the predicted SOH value of the battery health state at time t, SOH_{t+1}^pred is the predicted SOH value of the battery health state at time t+1, and max(·) represents the maximum value function.

[0109] When the predicted SOH value exhibits a monotonicity violation, i.e. The loss function imposes a penalty, forcing the model to learn predictions that conform to physical laws.

[0110] In this embodiment, the total training loss function of the bimodal network includes SOH prediction loss, RUL prediction loss, corrected flow-flow matching loss, latent space orthogonality constraint loss, and SOH monotonicity constraint loss, and its expression is as follows: L total = λ1L SOH + λ2L RUL + λ3L RF + λ4L orth + λ5L mono In the formula, L SOH L represents the SOH prediction loss, used to constrain the deviation between the predicted SOH value and the actual SOH value of the battery health state; RUL L represents the predicted loss of the remaining useful life (RUL), used to constrain the deviation between the predicted RUL value and the actual RUL value; RF This indicates the correction of the flow-to-flow matching loss, used to train the velocity field prediction module; L orth L represents the latent space orthogonal constraint loss, used to constrain the correlation between local features of temporary relations and inverse aging features; mono The loss represents the monotonicity constraint loss of SOH, used to penalize non-physical increases in SOH at adjacent prediction times; λ1 to λ5 are the weighting coefficients of the corresponding loss terms. Multi-task collaborative constraints are used to enable the battery health state (SOH) prediction task and the remaining useful life (RUL) prediction task to share the causally isolated aging state characterization.

[0111] There is an intrinsic correlation between State of Health (SOH) and Remaining Life (RUL). SOH reflects the current health status of the battery, while RUL reflects the remaining lifespan of the battery. This invention constructs a multi-task collaborative learning framework, achieving collaborative prediction by sharing feature layers and task-specific output heads.

[0112] The shared feature layer consists of a TV-SSM feature extraction layer and a Rectified Flow prediction generation layer. The extracted shared features are used for SOH and RUL prediction tasks.

[0113] The specific output head for the SOH prediction task is a fully connected neural network, containing two fully connected layers. The first layer has a dimension of 512, and the second layer has a dimension of 1. It outputs the SOH prediction value, expressed as follows:

[0114] In the formula, The output of the shared feature layer; and Weights and biases for the SOH prediction head; The Sigmoid activation function maps the output to the [0,1] interval.

[0115] The specific output head for the RUL prediction task is a fully connected neural network containing two fully connected layers. The first layer has a dimension of 512, and the second layer has a dimension of 1. It outputs the RUL prediction value, as shown in the following expression:

[0116] In the formula, and Weights and biases for the RUL prediction head; It uses the ReLU activation function to ensure that the output is non-negative.

[0117] The multi-task loss balancing strategy employs an uncertainty-weighted method to automatically adjust the loss weights of each task, as shown in the following expression:

[0118] In the formula, , Learnable weight parameters for SOH and RUL tasks are automatically updated via backpropagation; , The loss functions for the SOH and RUL tasks are respectively, using mean squared error loss.

[0119] Step 6: Output the predicted State of Health (SOH) value, the predicted Remaining Life (RUL) value, and their uncertainty intervals; based on the bimodal network trained with the physical monotonicity constraint and multi-task collaborative constraint, perform statistical processing on the predicted distribution to output the predicted State of Health (SOH) value, the predicted Remaining Life (RUL) value, and their uncertainty intervals. Specifically, in this embodiment, multiple SOH and RUL prediction samples are obtained by sampling the prior variables of the Rectified Flow generation model multiple times and integrating along the same ODE velocity field. The predicted values ​​of SOH and RUL are calculated based on these multiple prediction samples, and the uncertainty intervals of SOH and RUL are calculated using the quantile estimation method. In one embodiment, the uncertainty interval can be expressed as: CI = [Q0. 025 , Q0. 975 ] In the formula, Q0. 025 And Q0. 975 These represent the 2.5% and 97.5% quantiles of the predicted sample, respectively. The uncertainty interval is used to characterize the degree of uncertainty of the prediction result; the narrower the uncertainty interval, the lower the uncertainty of the prediction result; the wider the uncertainty interval, the higher the uncertainty of the prediction result.

[0120] Model training and validation: The training dataset is derived from real-world vehicle operation data of a hybrid electric vehicle model, containing complete lifecycle data for 50 vehicles, with a total of approximately 10 million records. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0121] The model hyperparameters are set as follows: the TV-SSM feature extraction layer has 4 to 6 stacked layers, the latent state dimension is 256 to 512, the time-varying parameter generation module consists of two fully connected layers; the polarization space in the latent space orthogonal decoupling module... and aging space The dimensions of the basis vectors are all set to 128 to 256, and the orthogonalization stability coefficient is... Take 1×10^-6.

[0122] The Rectified Flow prediction generation layer uses a lightweight multilayer perceptron as the velocity field network, with inputs including intermediate states. Time embedding And causal isolation aging state characterization c; Euler inference steps are set to 1 to 5 steps. Single-step or few-step inference is preferred when using vehicle-mounted online inference to reduce the computational burden on the BMS microcontroller.

[0123] The optimizer used is the Adam optimizer, with an initial learning rate of 1×10^-4. A cosine learning rate scheduling strategy is employed, with the learning rate decaying cosinely as training progresses. The batch size is set to 32, and an early stopping strategy is used, stopping training when the validation set loss no longer decreases over 10 consecutive epochs. Evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and prediction interval coverage (PICP). RMSE and MAE reflect prediction accuracy, MAPE reflects relative error, and PICP reflects the reliability of the confidence interval.

[0124] Experimental results show that the method of the present invention has an RMSE of 1.2%, a MAE of 0.9%, a MAPE of 2.5%, and a PICP of 94.5% on the test set, which is significantly better than traditional LSTM, GRU and other methods.

[0125] Implementation effect analysis: Figure 4 and Figure 5 The prediction results and confidence intervals for SOH and RUL are shown respectively. As can be seen from the figure, the prediction curve of the method of the present invention matches the true curve well, the confidence interval can effectively cover the true value, and the width of the confidence interval is moderate, reflecting the uncertainty of the prediction result.

[0126] Figure 6 The figure shows a comparison of the prediction accuracy of the method of this invention with existing methods. As can be seen from the figure, the RMSE, MAE, and MAPE of the method of this invention are significantly lower than those of traditional methods such as LSTM, GRU, and Transformer, proving the effectiveness of the method of this invention.

[0127] Figure 7 The results demonstrate the effectiveness of uncertainty quantification. As shown in the figure, the confidence interval provided by the method of this invention can effectively cover the true value, with a coverage rate of 94.5%, which is close to the theoretical value of 95%, proving the reliability of uncertainty quantification.

[0128] Example 3 This invention provides a battery state prediction system based on vehicle control big data and a dual-modal network, comprising: The first module is used to construct a battery historical big data set under the dynamic scheduling of the vehicle controller. The battery historical big data set includes battery charging and discharging timing signals, vehicle controller logic state codes, and dynamic scheduling state labels represented by the logic state codes. The second module is used to input the battery historical big data into the feature extraction layer of the dual-modal network. The feature extraction layer adopts the time-varying coupled state-space model TV-SSM to couple and model the battery charging and discharging time sequence signal with the vehicle controller logic state encoding to obtain local features that characterize the temporary relationship caused by the dynamic scheduling of the vehicle and reverse aging features that characterize the irreversible degradation process of the battery. The third module is used to introduce a latent space orthogonal decoupling mechanism in the feature extraction layer. Based on feature reprojection, the local features of the temporary relationship are projected to a latent subspace orthogonal to the inverse aging features. Based on orthogonal constraints, the mixing of dynamic scheduling state labels into the inverse aging features is suppressed, so as to form an aging state representation after causal isolation. The fourth module is used to input the aging state characterization after causal isolation into the prediction generation layer of the dual-modal network. The prediction generation layer adopts the modified flow generation model based on the ordinary differential equation ODE, learns the deterministic velocity field from the prior distribution to the target distribution of the battery state, and generates the predicted distribution of the battery health state SOH and remaining service life RUL by integrating along the deterministic velocity field. The fifth module is used to train the dual-modal network and introduce physical monotonicity constraints and multi-task collaborative constraints during training. The physical monotonicity constraints are used to constrain the predicted state of battery health (SOH) to meet the irreversible law of battery aging. The multi-task collaborative constraints are used to enable the SOH prediction task and the remaining useful life (RUL) prediction task to share the aging state representation after causal isolation. The sixth module is used to perform statistical processing on the predicted distribution based on the bimodal network trained by the physical monotonicity constraint and multi-task collaborative constraint, and output the predicted battery health state (SOH), the predicted remaining service life (RUL), and their uncertainty intervals.

[0129] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0130] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0131] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0132] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0133] 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 its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A battery state prediction method based on vehicle control big data and a dual-modal network, characterized in that, Includes the following steps: Step 1: Construct a battery history big data set under the dynamic scheduling of the vehicle controller. The battery history big data set includes battery charging and discharging timing signals, vehicle controller logic state codes, and dynamic scheduling state labels represented by the logic state codes. Step 2: Input the battery historical big data set into the feature extraction layer of the dual-modal network. The feature extraction layer adopts the time-varying coupled state-space model TV-SSM to couple and model the battery charging and discharging time sequence signal with the vehicle controller logic state encoding to obtain local features that characterize the temporary relationship caused by the dynamic scheduling of the vehicle and reverse aging features that characterize the irreversible degradation process of the battery. Step 3: Introduce a latent space orthogonal decoupling mechanism in the feature extraction layer. Based on feature reprojection, project the local features of the temporary relationship to a latent subspace orthogonal to the inverse aging features. Based on orthogonal constraints, suppress the mixing of dynamic scheduling state labels into the inverse aging features to form an aging state representation after causal isolation. Step 4: Input the aging state characterization after causal isolation into the prediction generation layer of the dual-modal network. The prediction generation layer adopts the modified flow generation model based on the ordinary differential equation ODE to learn the deterministic velocity field from the prior distribution to the target distribution of battery state, and integrates along the deterministic velocity field to generate the predicted distribution of battery health state SOH and remaining service life RUL. Step 5: Train the dual-modal network and introduce physical monotonicity constraints and multi-task collaborative constraints during training; wherein, the physical monotonicity constraints are used to constrain the battery health state (SOH) prediction value to meet the irreversible law of battery aging; the multi-task collaborative constraints are used to enable the battery health state (SOH) prediction task and the remaining useful life (RUL) prediction task to share the aging state representation after causal isolation. Step 6: Based on the bimodal network trained by the physical monotonicity constraint and multi-task collaborative constraint, perform statistical processing on the predicted distribution and output the predicted battery health state (SOH), the predicted remaining service life (RUL), and their uncertainty intervals.

2. The battery state prediction method based on vehicle control big data and dual-modal network according to claim 1, characterized in that, The construction of the battery historical big data set under the dynamic scheduling of the vehicle controller includes: The battery charging and discharging timing signals of a hybrid electric vehicle under actual operating conditions are collected. The battery charging and discharging timing signals include at least one of the following: battery current, battery voltage, battery temperature, state of charge (SOC), and charging and discharging power. The logic state code of the vehicle controller is collected, and the logic state code of the vehicle controller includes at least one of the following: drive mode, brake energy recovery mode, engine start-stop state, charge-discharge power distribution state and thermal management state. A dynamic scheduling status label is generated based on the vehicle controller logic state code, and the dynamic scheduling status label is aligned with the battery charging and discharging timing signal at the corresponding time. The aligned data is normalized, outlier removed, missing value imputed, and divided into sliding window segments to form the battery history dataset.

3. The battery state prediction method based on vehicle control big data and dual-modal network according to claim 1, characterized in that, The time-varying coupled state-space model TV-SSM in step 2 includes a state transition module, an input coupling module, and a time-varying parameter generation module; The state transition module is used to update the potential state of the battery. The input coupling module is used to map the battery charging and discharging timing signals and the vehicle controller logic state encoding to the same potential state space. The time-varying parameter generation module generates at least one time-varying parameter among the state transition matrix, input matrix, or gating coefficients based on the vehicle controller's logic state encoding, so that the state evolution of the TV-SSM adapts to the dynamic scheduling state of the vehicle controller.

4. The battery state prediction method based on vehicle control big data and dual-modal network according to claim 3, characterized in that, The state evolution of the TV-SSM satisfies the following form: in, Represents the potential state at time t. This represents the battery charging / discharging timing input at time t. This represents the logic state code of the vehicle controller at time t. , , , These are the time-varying parameters generated by the time-varying parameter generation module. Representing output features; by introducing The terms and their time-varying parameters enable the transient response caused by the vehicle controller scheduling to be explicitly characterized in the state space.

5. The battery state prediction method based on vehicle control big data and dual-modal network according to claim 1, characterized in that, The latent space orthogonal decoupling mechanism in step 3 includes an eigenvalue decomposition submodule, an eigenvalue reprojection submodule, and an orthogonal constraint submodule; The feature decomposition submodule decomposes the TV-SSM output features into temporary relational local features and inverse aging features; The feature reprojection submodule is used to reproject the local features of the temporary relationship to the orthogonal complement space of the inverse aging feature. The orthogonal complement space constitutes a potential subspace orthogonal to the inverse aging feature, and an orthogonal reprojection feature is obtained. The orthogonal constraint submodule is used to construct an orthogonal constraint loss based on the inner product or correlation between the temporary relationship local features and the reverse aging features, so as to suppress feature mixing between the temporary relationship local features and the reverse aging features.

6. The battery state prediction method based on vehicle control big data and dual-modal network according to claim 1, characterized in that, The orthogonal reprojection feature satisfies the following form: Among them, z d z represents a local feature of a temporary relation. a Indicates reverse aging characteristics, z d ⊥ represents the orthogonal reprojection feature, and ε is a stability coefficient to prevent the denominator from being zero; by subtracting the projection component of the temporary relation local feature in the direction of the reverse aging feature, the mixing of the vehicle controller dynamic scheduling feature with the reverse aging feature is suppressed.

7. The battery state prediction method based on vehicle control big data and a dual-modal network according to claim 1, characterized in that, The Rectified Flow generation model in step 4 includes a conditional coding module, a velocity field prediction module, and an ODE integration module; The conditional encoding module is used to encode the aging state characterization after causal isolation to obtain conditional features. The velocity field prediction module is used to predict based on prior variables. Target variable Continuous time variables and the conditional features Predicting the velocity field The ODE integration module is used to integrate from the prior distribution to the target distribution of battery state along the velocity field, generating predicted distributions of battery state of health (SOH) and remaining useful life (RUL). During the training of the bimodal network, the Rectified Flow generative model aims to minimize the flow matching loss, ensuring that the velocity field output by the velocity field prediction module fits the linear path velocity from the prior variable to the battery state target variable; the flow matching loss satisfies the following form: in, This represents the prior variable obtained by sampling from the prior distribution. The target variable represents the battery state. Indicates by and The intermediate variable obtained by linear interpolation, where t is a continuous time variable between 0 and 1, and c represents the conditional feature. This represents the velocity field output by the velocity field prediction module. Let E represent the flow matching loss, and E represent the mathematical expectation.

8. The battery state prediction method based on vehicle control big data and dual-modal network according to claim 1, characterized in that, The total training loss function of the bimodal network includes SOH prediction loss, RUL prediction loss, modified flow-flow matching loss, latent space orthogonal constraint loss, and SOH monotonicity constraint loss. Specifically, the SOH prediction loss is used to constrain the deviation between the predicted SOH value and the actual SOH value of the battery health state; the RUL prediction loss is used to constrain the deviation between the predicted RUL value and the actual RUL value of the remaining useful life; the modified flow matching loss is used to train the velocity field prediction module; the latent space orthogonal constraint loss is used to penalize the correlation between temporary relational local features and reverse aging features; and the SOH monotonicity constraint loss is used to penalize the non-physical increase of SOH at adjacent prediction times.

9. The battery state prediction method based on vehicle control big data and dual-modal network according to claim 1, characterized in that, The uncertainty interval in step 6 is obtained by sampling the prior variables of the modified flow generation model multiple times and integrating along the same ODE velocity field to obtain multiple prediction samples, and then calculating the uncertainty intervals of SOH and RUL based on the quantiles of the multiple prediction samples. The uncertainty interval, along with the SOH and RUL prediction values, are sent to the vehicle controller via the vehicle communication interface for charging and discharging power allocation, braking energy recovery strategy adjustment, thermal management strategy adjustment, or battery maintenance reminders.

10. A battery state prediction system based on vehicle control big data and a dual-modal network, characterized in that, include: The first module is used to construct a battery historical big data set under the dynamic scheduling of the vehicle controller. The battery historical big data set includes battery charging and discharging timing signals, vehicle controller logic state codes, and dynamic scheduling state labels represented by the logic state codes. The second module is used to input the battery historical big data into the feature extraction layer of the dual-modal network. The feature extraction layer adopts the time-varying coupled state-space model TV-SSM to couple and model the battery charging and discharging time sequence signal with the vehicle controller logic state encoding to obtain local features that characterize the temporary relationship caused by the dynamic scheduling of the vehicle and reverse aging features that characterize the irreversible degradation process of the battery. The third module is used to introduce a latent space orthogonal decoupling mechanism in the feature extraction layer. Based on feature reprojection, the local features of the temporary relationship are projected to a latent subspace orthogonal to the inverse aging features. Based on orthogonal constraints, the mixing of dynamic scheduling state labels into the inverse aging features is suppressed, so as to form an aging state representation after causal isolation. The fourth module is used to input the aging state characterization after causal isolation into the prediction generation layer of the dual-modal network. The prediction generation layer adopts the modified flow generation model based on the ordinary differential equation ODE, learns the deterministic velocity field from the prior distribution to the target distribution of the battery state, and generates the predicted distribution of the battery health state SOH and remaining service life RUL by integrating along the deterministic velocity field. The fifth module is used to train the dual-modal network and introduce physical monotonicity constraints and multi-task collaborative constraints during training. The physical monotonicity constraints are used to constrain the predicted state of battery health (SOH) to meet the irreversible law of battery aging. The multi-task collaborative constraints are used to enable the SOH prediction task and the remaining useful life (RUL) prediction task to share the aging state representation after causal isolation. The sixth module is used to perform statistical processing on the predicted distribution based on the bimodal network trained by the physical monotonicity constraint and multi-task collaborative constraint, and output the predicted battery health state (SOH), the predicted remaining service life (RUL), and their uncertainty intervals.

Images