Energy storage lithium battery state privacy protection perception method and system

Abstract

The application discloses a kind of energy storage lithium battery state privacy protection perception method and system, will original battery operating data save local privacy security domain;Using several battery physical models, the working condition feature vector extracted in battery operating data parallel parameter identification, to generate the physical information digital twin of battery according to parameter observability and construct mixed physical agent model, generate standardized virtual data using physical information digital twin, cloud end is aggregated by multiple edge computing terminal standardized virtual data training global deep learning model, for battery state perception.Cloud end server will be trained model and updated gate network parameters be issued back to edge computing terminal, to complete the closed-loop optimization of "perception-learning-evolution".The present application can simultaneously solve the problems of incomplete privacy protection, shallow physical model fusion level, heavy model training burden and large communication overhead in the process of energy storage lithium battery situation awareness.

Description

Technical Field

[0001] This invention relates to a battery state sensing method and system, belonging to the field of energy storage lithium battery state sensing, and particularly to an energy storage lithium battery state privacy protection sensing method and system. Background Technology

[0002] With the increasing prevalence of electric vehicles and large-scale energy storage power stations, the requirements for the intelligence level of battery management systems (BMS) are rising, with accurate assessment of key indicators such as state of health (SOH) and remaining useful life (RUL) being a core challenge. In recent years, data-driven methods based on deep learning have shown great potential in battery state assessment due to their powerful nonlinear fitting capabilities, becoming a research hotspot in both academia and industry. Through sensing and communication devices deployed in vehicles or energy storage stations, massive amounts of battery operating data can be collected and transmitted to the cloud, providing a data foundation for training high-precision deep learning models.

[0003] However, training high-performance deep learning models relies on massive, diverse data covering the entire battery lifecycle. This data not only contains sensitive user privacy information but also constitutes core trade secrets for battery manufacturers and operators. Therefore, directly sharing raw data for centralized model training faces serious "data silos" and privacy security issues. To alleviate these problems, some research has begun to explore privacy-preserving modeling frameworks such as federated learning, enabling collaborative updates of model parameters without uploading raw data. However, existing research still has some shortcomings:

[0004] 1. Incomplete privacy protection and risk of gradient leakage: While federated learning avoids the transmission of raw data, the shared model parameters or gradient information are not absolutely secure. Malicious parties could still use gradient inversion attacks to deduce information from the shared gradients, leading to privacy leaks. This risk is unacceptable for highly sensitive battery data.

[0005] 2. Shallow fusion of physical models and poor adaptability: Existing solutions typically select a fixed physical model and use it as a feature extractor. This approach ignores the dynamic changes in the dominant physicochemical processes of a battery at different life stages and under different operating conditions. Any single, fixed physical model is difficult to achieve optimal characterization across the entire life cycle and all operating conditions, resulting in limited adaptability and accuracy of the model in complex and variable scenarios.

[0006] 3. Heavy model training burden and insufficient physical interpretability: Each client model still needs to directly process raw, non-standard charging and discharging data. This not only places high demands on the computing power of edge computing terminals, but also means that the data-driven model is essentially a black box, learning complex electrochemical laws from scratch, resulting in low training efficiency. Furthermore, its internal decision-making logic lacks clear physical meaning, making it difficult to guarantee the model's credibility and robustness.

[0007] 4. High communication overhead and lack of efficient knowledge representation: The framework requires frequent transmission of model parameters between the client and server, resulting in significant bandwidth pressure and latency. Instead of a more concise and efficient physical representation of battery states, the framework transmits model parameters, leading to low information density. Summary of the Invention

[0008] Purpose of the invention: The purpose of this invention is to provide a method and system for privacy protection perception of the state of energy storage lithium batteries, which can solve problems such as incomplete privacy protection, shallow physical model fusion layer, heavy model training burden, and large communication overhead in the process of energy storage lithium battery situation perception.

[0009] Technical solution: The present invention provides a method for sensing the state privacy of energy storage lithium batteries, comprising:

[0010] Battery operation data is collected in real time from each edge computing terminal and saved to a local privacy and security domain;

[0011] The operating condition feature vector is extracted from the local battery operating data, and parallel parameter identification is performed based on several battery physical models. Then, a hybrid physical proxy model is constructed based on parameter observability.

[0012] A digital twin of the battery's physical information is generated based on the hybrid physical proxy model;

[0013] Standardized virtual data is generated using physical information digital twins and uploaded to a cloud server;

[0014] The cloud server aggregates the standardized virtual data from multiple edge computing terminals and trains a global deep learning model for battery status awareness based on this data.

[0015] The trained global deep learning model is distributed to edge computing terminals, and each edge computing terminal uses local data to perform battery status perception and prediction.

[0016] Optionally, the method further includes: pre-installing a physical proxy model library and a gated neural network on the edge computing terminal to construct a basic computing framework for multi-model fusion; wherein, the physical proxy model library contains several battery physical models for describing battery characteristics from different perspectives; the gated neural network is used to dynamically assign weights to each battery physical model based on the battery's operating conditions to construct a basic computing framework for multi-model fusion.

[0017] Optionally, the types of battery physical models include equivalent circuit models, electrochemical mechanism models, and multiphysics coupling models.

[0018] Optionally, the plurality of battery physical models preferably include macroscopic equivalent circuit models and simplified microscopic electrochemical mechanism models. The macroscopic equivalent circuit models include first-order and second-order Thevenin models and PNGV models, and the simplified microscopic electrochemical mechanism models include single-particle models and their reduced-order models, reduced-order P2D models, and thermal coupling models.

[0019] The local privacy security domain refers to the encrypted partition of the edge computing terminal, which contains a local privacy database to ensure that sensitive raw usage trajectory data does not leave the domain.

[0020] Optionally, the extraction of the operating condition feature vector from the local battery operating data includes:

[0021] Extract a current sequence of length L from the local battery operation data during the constant voltage charging process, and calculate the multi-scale current excitation entropy H. exc This is used to characterize the degree of activation of the internal electrochemical reactions of the battery under current operating conditions.

[0022]

[0023] in, This represents the normalized energy percentage of the m-th frequency component after frequency domain transformation of the current sequence. This represents the total number of frequency components. This represents the maximum step amplitude of the current within the window. Rated current; This is the adjustment coefficient;

[0024] Construct the working condition feature vector as follows ,in, This represents the average battery temperature. This is an estimate of the battery's state of charge. This represents the voltage recovery rate during the battery relaxation phase.

[0025] Preferably, the value of L is in the range of 50-100.

[0026] Preferred, The value range is 0.5-2.0.

[0027] Preferably, the parallel parameter identification is preferably extended Kalman filtering or recursive least squares with a forgetting factor.

[0028] Optionally, the hybrid physical proxy model is the optimal physical proxy model at the current moment, synthesized based on the dynamic weights of observability and the optimal parameter set obtained by the parallel parameter identification, as shown in the following formula:

[0029]

[0030] in, For dynamic weights, To utilize local battery operating data to solve for the optimal parameter set of each battery physical model at time k, where N is the number of battery physical models, Let i be the physical model of the i-th battery.

[0031] Optionally, the dynamic weights are obtained according to the following steps: the operating condition feature vector x is input into a gating network, and the network calculates the comprehensive score of each battery physical model at time k. , Let be the total number of battery physics models, and then use the Softmax function to generate the dynamic weights of the i-th battery physics model. :

[0032] .

[0033] Optionally, the comprehensive score represents the contribution of the i-th battery physics model under the current operating conditions:

[0034]

[0035] in, This represents the degree of matching and reliability of the i-th battery physical model relative to the current operating condition at time k; These are the weight parameters currently learned by the gated neural network; Let k be the feature vector of the working condition at time k; These are the weighting coefficients for the observability of model parameters; is the Jacobian matrix of the output voltage of the i-th battery physical model with respect to the set of parameters to be identified, which reflects the degree of influence of changes in model parameters on the output voltage; is the Frobenius norm; cond(⋅) is the condition number of the matrix. The larger this ratio is, the easier it is to accurately identify the parameters of the physical model under the current operating conditions; This is a penalty coefficient for calculating computational complexity; Let be the computational complexity index for the i-th model.

[0036] Optionally, the generation of standardized virtual data using a physical information digital twin includes:

[0037] Based on the actual physical parameters identified locally A battery physical model is constructed sequentially, and a standard virtual current I is applied to the input terminal of each model. std (t) drives the physical information digital twin to perform simulation and generate a virtual response voltage generated by the battery under public operating conditions. :

[0038]

[0039] in, and These are the set of physical parameters and the weight vector for real-time calibration at time k, respectively, encapsulated within the object of the physical information digital twin. To use local battery operating data to solve for the optimal parameters of the i-th battery physical model at time k; Let N be the i-th battery physical model, and N be the number of battery physical models.

[0040] virtual data packets Upload to the cloud server.

[0041] Optionally, the cloud server constructs a Transformer-based global deep learning model, with the input being a virtual voltage sequence. and physical parameter sequence The goal is to predict the health status and remaining lifespan of batteries.

[0042] The cloud-based global deep learning model and its loss function are as follows:

[0043]

[0044]

[0045] in, For physical consistency constraints, These are constraint coefficients to prevent non-physical abrupt changes in physical parameters. This represents the true value of the battery's health status. This is a predicted value for battery health. This represents the actual remaining battery life. This is a predicted value for the remaining battery life. It is a Transformer model.

[0046] Optionally, the method further includes: using big data analysis in the cloud to determine the actual performance deviations of various battery physical models under different operating conditions, and updating the weight parameters of the gated neural network. The updated gated neural network parameters are then sent back to the edge computing terminal; the weight parameters of the gated neural network... The update is as follows:

[0047]

[0048] in, For learning rate, The weight parameters of the gated neural network before the update. For the updated gated neural network weight parameters, To update the gradient.

[0049] The energy storage lithium battery status privacy protection sensing system includes an edge computing terminal and a cloud server. The edge computing terminal is equipped with:

[0050] The data acquisition module is used to collect battery operating data in real time.

[0051] A secure storage module is used to store the battery's operating data in a privacy-secure domain;

[0052] A collaborative identification and twin engine is used to extract operating condition feature vectors from local battery operating data, perform parallel parameter identification based on several battery physical models, and then construct a hybrid physical proxy model based on parameter observability; a digital twin of the battery's physical information is generated based on the hybrid physical proxy model.

[0053] The virtual mapping module is used to generate standardized virtual data using physical information digital twins and upload it to a cloud server.

[0054] The cloud server is equipped with:

[0055] A data aggregation module is used to aggregate the standardized virtual data from multiple edge computing terminals;

[0056] The training module is used to train a global deep learning model for battery state awareness based on the standardized virtual data.

[0057] The strategy optimization module is used to send the trained global deep learning model and / or updated gated neural network parameters to the edge computing terminal.

[0058] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0059] 1. This invention constructs a hybrid physical proxy model on an edge computing terminal, which intelligently schedules multiple battery physical models and data-driven models. Then, it transforms the original sensitive data into a standardized, information-rich digital twin of physical information, generating virtual data based on standard operating conditions, which is then used for model training. This invention achieves thorough privacy protection while integrating physical mechanisms to improve the model's training efficiency, accuracy, and interpretability.

[0060] 2. This invention employs a counterfactual virtual mapping technique, which generates virtual data based on public / standard operating conditions. This not only avoids the risk of privacy leakage of sensitive battery data, but also ensures high consistency since the virtual data of each edge computing terminal is generated based on the same standard virtual current. This facilitates the aggregation of cloud data and the training of global models.

[0061] 3. When constructing the hybrid physical proxy model, several battery physical models are combined to cover different perspectives of describing battery characteristics, and dynamic weights are set based on observable rows for weighting. This enables the hybrid physical proxy model to accurately characterize the complex electrochemical mechanism and life cycle evolution of lithium batteries, and it can be adapted to various scenarios, especially those where the physical meaning needs to be clearly defined. Attached Figure Description

[0062] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention;

[0063] Figure 2 This is a schematic diagram of the basic computing framework based on multi-model fusion in Embodiment 2 of the present invention;

[0064] Figure 3 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0065] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0066] A method for privacy protection sensing of the state of energy storage lithium batteries includes:

[0067] Battery operation data is collected in real time from each edge computing terminal and saved to a local privacy and security domain;

[0068] The operating condition feature vector is extracted from the local battery operating data, and parallel parameter identification is performed based on several battery physical models. Then, a hybrid physical proxy model is constructed based on parameter observability.

[0069] A digital twin of the battery's physical information is generated based on the hybrid physical proxy model;

[0070] Standardized virtual data is generated using physical information digital twins and uploaded to a cloud server;

[0071] The cloud server aggregates the standardized virtual data from multiple edge computing terminals and trains a global deep learning model for battery status awareness based on this data.

[0072] The trained global deep learning model is distributed to edge computing terminals, and each edge computing terminal uses local data to perform battery status perception and prediction.

[0073] Example 1

[0074] The parallel parameter identification is based on a multi-model fusion computing framework pre-built on an edge computing terminal. Therefore, as Figure 1 As shown, in one embodiment, the energy storage lithium battery state privacy protection sensing method includes the following specific steps.

[0075] Step S10: Pre-install a physical agent model library and gating network on the edge computing terminal to build a basic computing framework for multi-model fusion.

[0076] The physical proxy model library contains several battery physical models used to describe battery characteristics from different perspectives; the gated neural network is used to dynamically assign weights to each battery physical model based on the battery's operating conditions, thus constructing a basic computational framework for multi-model fusion.

[0077] In one implementation, the physical proxy model library needs to contain 3-4 different types or complexities of battery physical models. The 3-4 types of battery physical models must cover equivalent circuit models, electrochemical mechanism models and multi-physics coupling models, so that the physical proxy model library can describe battery characteristics from different perspectives of electricity, chemistry and multi-physics coupling.

[0078] Step S20: Collect and store the battery's raw operating data locally on the edge computing terminal to establish a local privacy and security domain.

[0079] In one embodiment, the edge computing terminal is connected to a high-precision sensor that collects battery operating data in real time at a frequency of 10Hz-100Hz. The vector definition of the battery operating data is D(t)=[V(t), I(t), T(t)]. T , where V(t), I(t), and T(t) are the terminal voltage, current, and battery surface temperature in time series form, respectively.

[0080] The collected battery operation data will be written directly to the hardware-encrypted partition of the edge terminal without any compression or encoding, and a local privacy database will be established to form a local privacy and security domain. The raw operation data is strictly confined to this storage area and will not be transmitted to the external network in plaintext; it is only available for use by local algorithms.

[0081] Step S30: Perform collaborative identification and dynamic weighting of multiple models on the edge computing terminal to obtain a hybrid physical proxy model.

[0082] In one embodiment, step S30 specifically includes:

[0083] Step S301: Extract a current sequence of length L from the battery operating data during the constant voltage charging process, and calculate the multi-scale current excitation entropy H. exc This is used to characterize the degree of activation of the internal electrochemical reactions of the battery under current operating conditions.

[0084]

[0085] in, This represents the normalized energy percentage of the m-th frequency component after frequency domain transformation of the current sequence. This represents the total number of frequency components. This represents the maximum step amplitude of the current within the window. Rated current; This is the adjustment coefficient.

[0086] Construct the working condition feature vector as follows ,in, This represents the average battery temperature. This is an estimate of the battery's state of charge. This represents the voltage recovery rate during the battery relaxation phase.

[0087] Since the current value at a single point in time is a scalar and lacks frequency characteristics, while infinitely long historical data does have frequency, the computational burden is too great and it contains outdated information. Therefore, only by selecting a waveform segment of a certain length L can a Fast Fourier Transform or Wavelet Transform be performed to extract the frequency distribution information reflecting the dynamic characteristics of the operating condition. Therefore, as a preferred approach, the value of L is in the range of 50-100.

[0088] Preferred, The value range is 0.5-2.0. This range is chosen based on the following physical and signal processing principles:

[0089] Considering the inherent measurement noise of on-board current sensors, when If the value is less than 0.5, the algorithm is too sensitive to high-frequency small-amplitude noise, which can easily lead to misjudgments by the gating network under steady-state conditions; when When the value is greater than 2, the distinguishability of the entropy value decreases, making it difficult to effectively capture transient spikes in the current signal (such as rapid acceleration conditions), resulting in a decrease in the ability to characterize high-frequency excitations.

[0090] choose A value between 0.5 and 2.0 allows the excitation entropy to map the characteristic timescale of the establishment and elimination of polarization impedance within the battery to the greatest extent possible, thus providing the most physically relevant decision-making basis for the gating network.

[0091] Step S302: Perform parallel parameter identification based on multiple models.

[0092] For example, the parameters that need to be identified for the first-order Thevenin model are R0, R1, and C1. R0 is the ohmic internal resistance, used to track the instantaneous voltage response; R1 and C1 are the polarization resistance and polarization capacitance, respectively, used to track the diffusion polarization process inside the battery. For the single-particle model, the parameters that need to be identified are D, K, and R. D is the effective solid-phase diffusion coefficient, reflecting the transport capability of lithium ions within the particle; K is the electrochemical reaction rate constant, reflecting the interfacial charge transfer rate; and R is the electrode surface film resistance, reflecting the impedance change caused by the thickening of the SEI film.

[0093] The online identification algorithm is run in parallel for each physical model in the model library. In this embodiment, a recursive least squares method with a forgetting factor is used to solve for the optimal parameter set of the i-th model at time k using local raw data. The specific process is as follows:

[0094] For the i-th battery model at time k, the output equation for the terminal voltage observation can be linearized as follows: The parameter iteration update process is as follows:

[0095] (1) First, based on the covariance matrix of the previous time step and the data vector at the current time Calculate the gain matrix This gain matrix determines the weight of the current measurement data on the parameter correction. The calculation formula is as follows:

[0096]

[0097] in, It is the gain factor.

[0098] (2) Next, calculate the prior error. This error reflects the measured voltage at time k. Compared with predictions based on parameters from the previous time step The deviation between the voltages. Prior error. The calculation formula is as follows:

[0099]

[0100] (3) Subsequently, by correcting the physical model parameter set, the latest parameter estimates at the current time are obtained. This enables adaptive tracking of parameters in response to changes in battery operating state.

[0101]

[0102] (4) Finally, the covariance matrix is ​​updated. Among them, a forgetting factor is introduced. Its value ranges from 0.95 to 0.99. The weights of older data are exponentially decayed to prevent data saturation and ensure the algorithm maintains sensitivity to time-varying parameter characteristics throughout the battery's entire lifespan. Covariance matrix. The updated formula is as follows:

[0103]

[0104] Recursive least squares with forgetting factor is used to evaluate the parameters. By continuously iterating and updating until convergence, the optimal parameter set of each model at time k can be obtained. .

[0105] Step S303: Generate dynamic weights based on observability. Input the working condition feature vector x into the gating network, and the network sequentially calculates the comprehensive score S of the i-th model. i (k) represents the contribution of each physical model under the current operating conditions.

[0106]

[0107] in, This represents the degree of matching and reliability of the i-th physical model in the model library relative to the current working condition at time k. These are the weight parameters currently learned by the gated neural network; It is the feature vector of the working condition at time k; These are the weighting coefficients for the observability of model parameters; The Jacobian matrix of the output voltage of the i-th model with respect to the parameter set to be identified reflects the degree of influence of changes in model parameters on the output voltage. is the Frobenius norm; cond(⋅) is the condition number of the matrix. The larger this ratio is, the easier it is to accurately identify the parameters of the physical model under the current operating conditions; This is a penalty coefficient for calculating computational complexity; Let be the computational complexity index for the i-th model.

[0108] Finally, the final dynamic weights are generated using the Softmax function. :

[0109]

[0110] Step S304: Construct a hybrid physical proxy model. Based on the weights obtained in step S303 and the optimal parameter set obtained in step S302, construct the optimal physical proxy model for the current moment:

[0111]

[0112] Step S40: Based on the hybrid physical agent model constructed in step S30, instantiate or update the digital twin of the battery's physical information on the edge computing terminal.

[0113] The physical information digital twin object encapsulates the set of physical parameters calibrated in real time at time k. and weight vector .

[0114] In a preferred embodiment, the encrypted hash value of the core data of the physical information digital twin object is uploaded to the blockchain for evidence storage, forming an immutable battery health record.

[0115] Step S50: Use the local physical information digital twin to generate standardized virtual data and upload it to the cloud server.

[0116] Because users have different usage habits, charging conditions, and charging environments for energy storage batteries, the statistical characteristics (such as mean, variance, and covariance) of local real charging current and voltage data change over time, becoming non-stationary data. The edge terminal controls the physical information digital twin to perform simulations, mapping these non-stationary local real operating data to virtual data under standard operating conditions, and only uploading the virtual data and model parameters.

[0117] In one embodiment, step S50 specifically includes:

[0118] Step S501: Load the standard procedure. The edge terminal has a pre-stored set of unified standard DST operating condition test procedures, and the current used in this test procedure is marked as the standard virtual current sequence I. std (t).

[0119] Step S502: Perform counterfactual virtual mapping. Utilize locally identified real physical parameters. The physical models are constructed sequentially, and a standard virtual current I is applied to the input terminals of each model. std (t) drives the digital twin to perform simulation and generate a virtual response voltage "if the battery is operating under standard DST conditions". :

[0120]

[0121] in, and These are the set of physical parameters and the weight vector for real-time calibration at time k, respectively, encapsulated within the object of the physical information digital twin. To use local battery operating data to solve for the optimal parameters of the i-th battery physical model at time k; Let N be the i-th battery physical model, and N be the number of battery physical models.

[0122] Step S503, Data Upload. Only the generated virtual data packet is uploaded. Uploaded to the cloud server via wireless network. Due to I std (t) is the standard virtual current, which is publicly available information shared between the cloud and the edge. It is a virtual voltage generated based on publicly available operating conditions. The original user's real current and voltage trajectory never leaves the local area, thus completely cutting off the path of privacy leakage.

[0123] Step S60: The cloud server aggregates virtual data from multiple edge computing terminals and trains a global deep learning model based on this data for battery status awareness.

[0124] In one implementation, step S60 specifically includes:

[0125] Step S601, Cloud Data Aggregation. The cloud server receives virtual data packets uploaded by a massive number of terminals. Since the data is based on the same I... std (t) is generated and has a high degree of consistency.

[0126] Step S602, Global Model Training. Using a deep neural network, with virtual voltage sequences and physical parameter evolution characteristics as input, a global battery health status and remaining lifespan prediction model is trained to achieve battery status awareness.

[0127] In this embodiment, a cloud-based global deep learning model based on Transformer is constructed, and the input of the model is a virtual voltage sequence. and physical parameter sequence The goal is to predict the battery's health status and remaining lifespan. The cloud-based global deep learning model and its loss function are as follows:

[0128]

[0129]

[0130] in, For physical consistency constraints, These are constraint coefficients to prevent non-physical abrupt changes in physical parameters. This represents the true value of the battery's health status. This is a predicted value for battery health. This represents the actual remaining battery life. This is a predicted value for the remaining battery life. It is a Transformer model.

[0131] Step S70: The cloud server performs model distribution and closed-loop optimization.

[0132] In one embodiment, step S70 specifically includes:

[0133] Step S701, Model Deployment. The trained global deep learning model is deployed to the edge terminal via OTA technology. The edge terminal uses this model to make high-precision predictions of the future battery state locally.

[0134] Step S702, Gated Network Evolution. The cloud utilizes big data analysis to examine the actual performance deviations of various physical models under different operating conditions, and updates the weight parameters of the gated neural network accordingly. The updated gating network parameters are then sent back to the edge terminal to achieve closed-loop optimization of "perception-learning-evolution". The updated weight parameters of the gating neural network are:

[0135]

[0136] in, For learning rate, The weight parameters of the gated neural network before the update. For the updated gated neural network weight parameters, To update the gradient.

[0137] Example 2

[0138] like Figure 2 As shown, the construction of the basic computing framework for multi-model fusion includes:

[0139] Step S101: Construct a physical proxy model library: Pre-configure a physical proxy model library containing N different types and complexities of battery physical models on the edge terminal. The physical proxy model library preferably includes macroscopic equivalent circuit models and simplified microscopic electrochemical mechanism models. The macroscopic equivalent circuit models include, but are not limited to: first-order and second-order Thevenin models, and PNGV models; the simplified microscopic electrochemical mechanism models include, but are not limited to: single-particle models and their reduced-order models, reduced-order P2D models, and thermal coupling models.

[0140] In this embodiment, the physical proxy model library includes:

[0141] (1) Second-order Thevenin model: focuses on describing the polarization effect of the battery, has low computational complexity, and is suitable for steady-state conditions;

[0142] (2) PNGV model: Based on the basic circuit model, a description of capacitance change is added, which is suitable for working conditions with drastic load changes;

[0143] (3) Single-particle model: simplifies the microscopic mechanism and focuses on describing the solid-phase diffusion process;

[0144] (4) Reduced-order thermal coupling P2D model: focuses on describing the coupling between electrochemical and thermal properties.

[0145] Step S102: Construct a gated neural network: This network is a lightweight fully connected neural network, with the input layer dimension consistent with the dimension of the working condition feature vector. The output layer uses the Softmax activation function to ensure the output weight vector... = [ , ,..., ], The weights are dynamic, and N is the number of battery physics models to ensure the non-negativity and normalization of the output weights.

[0146] Step S103: Deployment at the edge terminal: In edge terminals such as the electric vehicle onboard T-BOX and energy storage BMS controller, a physical agent model library and gating network are pre-installed, and the corresponding initial parameters and configuration files are stored; this provides a basic computing framework for the collaborative identification and dynamic weighting of the subsequent hybrid physical agent model.

[0147] like Figure 3 As shown, the energy storage lithium battery status privacy protection sensing system of the present invention includes: an edge computing terminal 10 and a cloud server 20.

[0148] The edge computing terminal 10 is used to execute steps S20 to S50 of the above method, uploading virtual data packets to the cloud server.

[0149] In one embodiment, the edge computing terminal 10 internally includes:

[0150] Data acquisition module 101: includes a high-precision ADC and signal conditioning circuit, performs high-frequency acquisition of sensor data in step S20 to obtain the battery's raw operating data;

[0151] Secure storage module 102: Uses eMMC or UFS storage chips and divides the area into TrustZone secure zones to store raw sensor sensitive data.

[0152] Collaborative Identification and Twin Engine 103: Step S30 involves parallel identification of multiple models, calculation of operational condition excitation entropy and observability score, generation of dynamic weights, and finally, obtaining a hybrid physical agent model. Based on this model, step S40 integrates parameters and weights to complete the instantiation of the digital twin.

[0153] Virtual mapping module 104: Executes the virtual data generation in step S50 to generate desensitized virtual data.

[0154] The cloud server 20 is used to execute steps S60 to S70 of the above method, and to send the gradient and model parameters to the edge computing terminal.

[0155] In one embodiment, the internal functional modules of the cloud server 20 include:

[0156] Data aggregation module: Executes step S60 to receive and parse virtual data packets via MQTT or HTTPS protocols;

[0157] Global model training cluster: Equipped with a GPU acceleration card, it executes the training of a global deep learning model in step S60 to achieve battery status awareness;

[0158] Strategy optimization module: Executes step S70 to calculate the update gradient of the gating network and manage OTA push.

Claims

1. A method for sensing the state privacy of energy storage lithium batteries, characterized in that, The method includes: Battery operation data is collected in real time from each edge computing terminal and saved to a local privacy and security domain; The operating condition feature vector is extracted from the local battery operating data, and parallel parameter identification is performed based on several battery physical models. Then, a hybrid physical proxy model is constructed based on parameter observability. A digital twin of the battery's physical information is generated based on the hybrid physical proxy model; Standardized virtual data is generated using physical information digital twins, including: based on locally identified real physical parameters. A battery physical model is constructed sequentially, and a standard virtual current I is applied to the input terminal of each model. std (t) drives the digital twin of physical information to perform simulation and generate a virtual response voltage generated by the battery under public operating conditions: ;in, and These are the set of physical parameters and the weight vector for real-time calibration at time k, respectively, encapsulated within the object of the physical information digital twin. To use local battery operating data to solve for the optimal parameters of the i-th battery physical model at time k; Let N be the i-th battery physical model, and N be the number of battery physical models. virtual data packets Uploaded to the cloud server; The cloud server aggregates the standardized virtual data from multiple edge computing terminals and trains a global deep learning model for battery status awareness based on this data. The trained global deep learning model is distributed to edge computing terminals, and each edge computing terminal uses local data to perform battery status perception and prediction.

2. The energy storage lithium battery state privacy protection sensing method according to claim 1, characterized in that, The method further includes: pre-installing a physical proxy model library and a gated neural network on the edge computing terminal to construct a basic computing framework for multi-model fusion; wherein, the physical proxy model library contains several battery physical models used to describe battery characteristics from different perspectives; the gated neural network is used to dynamically assign weights to each battery physical model based on the battery's operating conditions to construct a basic computing framework for multi-model fusion.

3. The energy storage lithium battery state privacy protection sensing method according to claim 1, characterized in that, The extraction of the operating condition feature vector from the local battery operating data includes: Extract a current sequence of length L from the local battery operation data during the constant voltage charging process, and calculate the multi-scale current excitation entropy H. exc This is used to characterize the degree of activation of the internal electrochemical reactions of the battery under current operating conditions. in, This represents the normalized energy percentage of the m-th frequency component after frequency domain transformation of the current sequence. This represents the total number of frequency components. This represents the maximum step amplitude of the current within the window. Rated current; This is the adjustment coefficient; Construct the working condition feature vector as follows ,in, This represents the average battery temperature. This is an estimate of the battery's state of charge. This represents the voltage recovery rate during the battery relaxation phase.

4. The energy storage lithium battery state privacy protection sensing method according to claim 1, characterized in that, The hybrid physical proxy model is the optimal physical proxy model synthesized based on the dynamic weights of observability and the optimal parameter set obtained by parallel parameter identification, as shown in the following formula: in, For dynamic weights, To utilize local battery operating data to solve for the optimal parameter set of each battery physical model at time k, where N is the number of battery physical models, Let i be the physical model of the i-th battery.

5. The energy storage lithium battery state privacy protection sensing method according to claim 4, characterized in that, The dynamic weights are obtained according to the following steps: the operating condition feature vector x is input into the gated network, and the network calculates the comprehensive score of each battery physical model at time k. , Let be the total number of battery physics models, and then use the Softmax function to generate the dynamic weights of the i-th battery physics model. : 。 6. The energy storage lithium battery state privacy protection sensing method according to claim 5, characterized in that, The comprehensive score represents the contribution of the i-th battery physics model under the current operating conditions: in, This represents the degree of matching and reliability of the i-th battery physical model relative to the current operating condition at time k; These are the weight parameters currently learned by the gated neural network; Let k be the feature vector of the working condition at time k; These are the weighting coefficients for the observability of model parameters; is the Jacobian matrix of the output voltage of the i-th battery physical model with respect to the set of parameters to be identified, which reflects the degree of influence of changes in model parameters on the output voltage; is the Frobenius norm; cond(⋅) is the condition number of the matrix. The larger this ratio is, the easier it is to accurately identify the parameters of the physical model under the current operating conditions; This is a penalty coefficient for calculating computational complexity; Let be the computational complexity index for the i-th model.

7. The energy storage lithium battery state privacy protection sensing method according to claim 1, characterized in that, The cloud server constructed a global deep learning model based on Transformer, with the model's input being a virtual voltage sequence. and physical parameter sequence The goal is to predict the health status and remaining lifespan of batteries. The cloud-based global deep learning model and its loss function are as follows: in, For physical consistency constraints, These are constraint coefficients to prevent non-physical abrupt changes in physical parameters. This represents the true value of the battery's health status. This is a predicted value for battery health. This represents the actual remaining battery life. This is a predicted value for the remaining battery life. It is a Transformer model.

8. The energy storage lithium battery state privacy protection sensing method according to claim 1, characterized in that, The method also includes: using big data analysis in the cloud to identify the actual performance deviations of various battery physical models under different operating conditions, and updating the weight parameters of the gated neural network. The updated gated neural network parameters are then sent back to the edge computing terminal; the weight parameters of the gated neural network... The update is as follows: in, For learning rate, The weight parameters of the gated neural network before the update. For the updated gated neural network weight parameters, To update the gradient.

9. A privacy protection sensing system for the status of energy storage lithium batteries, characterized in that, The system includes edge computing terminals and cloud servers; The edge computing terminal is equipped with: The data acquisition module is used to collect battery operating data in real time. A secure storage module is used to store the battery's operating data in a privacy-secure domain; A collaborative identification and twin engine is used to extract operating condition feature vectors from local battery operating data, perform parallel parameter identification based on several battery physical models, and then construct a hybrid physical proxy model based on parameter observability; a digital twin of the battery's physical information is generated based on the hybrid physical proxy model. The virtual mapping module is used to generate standardized virtual data using physical information digital twins, including: based on locally identified real physical parameters. A battery physical model is constructed sequentially, and a standard virtual current I is applied to the input terminal of each model. std (t) drives the physical information digital twin to perform simulation and generate a virtual response voltage generated by the battery under public operating conditions. : in, and These are the set of physical parameters and the weight vector for real-time calibration at time k, respectively, encapsulated within the object of the physical information digital twin. To use local battery operating data to solve for the optimal parameters of the i-th battery physical model at time k; Let N be the number of battery physical models, and let N be the number of battery physical models; The virtual data packet... Upload to the cloud server; The cloud server is equipped with: A data aggregation module is used to aggregate the standardized virtual data from multiple edge computing terminals; The training module is used to train a global deep learning model for battery state awareness based on the standardized virtual data. The strategy optimization module is used to send the trained global deep learning model and / or updated gated neural network parameters to the edge computing terminal.

Images