A physical information driven federated thermal state estimation model parameter collaborative training and end-side cloud layered deployment method and system

Abstract

The application discloses a physical information driven federal thermal state estimation model parameter collaborative training and end side cloud layered deployment method and system, belongs to the technical field of lithium ion battery thermal management and state estimation, and comprises the following steps: each edge node collects local battery operation data, takes a data fitting term and a non-steady-state heat conduction equation discretization residual as a physical constraint term to construct a loss function, and trains a local thermal state estimation model; after model parameters are uploaded to a server through privacy protection processing, a federal average algorithm is adopted to aggregate and update a global model and distribute the global model to each edge node; after the global model is incrementally fine-tuned by the edge node, the global model is deployed on a battery management system embedded controller through light compression, and temperature prediction and thermal runaway early warning are output in real time; and the model is updated online and continuously through the circulation of global model distribution, incremental fine-tuning, parameter uploading and federal aggregation. The application considers data privacy protection, physical consistency, end side real-time performance and full life cycle adaptive capacity.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery thermal management and state estimation technology, and in particular to a physical information-driven method for collaborative training of parameters of a federated thermal state estimation model and hierarchical deployment of edge-cloud. Background Technology

[0002] Lithium-ion batteries generate a significant amount of Joule heat and reaction heat during charging and discharging. If this heat cannot be dissipated in time, the battery temperature will rise, potentially leading to thermal runaway. Accurately and in real-time estimating the internal temperature distribution and heat diffusion trends of the battery module is one of the core tasks of a battery management system.

[0003] Existing methods for estimating thermal state can be mainly classified into the following categories: (1) Estimation method based on physical model: The temperature distribution is obtained by solving partial differential equations by establishing an electrochemical-thermal coupling model or an equivalent thermal network model. This type of method has high accuracy, but the model is complex and the computation is large, making it difficult to deploy in an embedded BMS; at the same time, the model parameters drift with battery aging, requiring frequent calibration and having poor adaptability.

[0004] (2) Pure data-driven estimation method: This method uses machine learning models such as neural networks to learn the mapping relationship between inputs such as current and voltage and output temperature from historical data. This type of method performs well under a single operating condition, but has poor generalization ability across operating conditions; moreover, training requires the centralized collection of a large amount of battery data (the operating data of different car companies and users involves trade secrets and personal privacy), which poses a serious risk of data privacy leakage.

[0005] (3) Application of Federated Learning in the Battery Field: Federated learning, as a privacy-preserving distributed machine learning framework, has been initially applied in the field of battery state of health (SOH) estimation (e.g., CN121302313A, CN120180101A). However, existing federated learning technologies all focus on SOH / SOC estimation. Their data modes (electrochemical impedance spectroscopy, capacity decay curve) are fundamentally different from thermal state estimation (temperature spatiotemporal sequence, thermal diffusion characteristics). Furthermore, the slow-changing characteristics of SOH (time scale of months or years) and the real-time requirements (second level) of thermal state estimation are different in terms of task objectives and model architecture, and cannot be simply transferred.

[0006] (4) Cloud-edge collaborative thermal management method: Existing patents (such as CN120511803A and CN121011768A) have proposed a cloud-edge collaborative thermal management architecture. However, these solutions adopt a centralized training mode, and the data of each vehicle or energy storage site needs to be uploaded to the cloud, which cannot guarantee data privacy. At the same time, knowledge between different sites cannot be shared, and each site trains the model independently, which limits the generalization ability of the model.

[0007] (5) Application of Physical Information Neural Network (PINN): CN121848931B has disclosed a method for identifying the thermal runaway state of lithium batteries based on physical information neural network. However, this method focuses on the thermal runaway "state identification" of a single cell battery. The core is to extract heat source features from electrical signals through PINN to detect the thermal runaway critical point, which belongs to the single cell level thermal runaway detection. In contrast, this invention is aimed at the continuous estimation of thermal state at the module level (temperature distribution prediction, thermal diffusion trend analysis), and adopts federated learning to achieve distributed collaborative training. The two are fundamentally different in terms of technical goals, architecture level and implementation method.

[0008] In summary, existing technologies cannot simultaneously meet multiple requirements, including data privacy protection, cross-operating condition knowledge sharing, physical consistency, edge real-time performance, and full lifecycle adaptability. Summary of the Invention

[0009] To address the shortcomings of the existing technologies, the present invention aims to provide a method for collaborative training of parameters and hierarchical deployment of edge-cloud-based federated thermal state estimation models driven by physical information. This method ensures that the original operating data of each edge node remains local, guarantees that the model conforms to thermodynamic laws through a loss function based on physical constraints, leverages federated learning to aggregate knowledge from multiple nodes to enhance cross-condition generalization capabilities, reduces communication overhead and meets data localization compliance requirements through a hierarchical aggregation architecture, achieves millisecond-level real-time inference for the BMS embedded controller through lightweight compression, and updates the model online following battery aging using a federated continuous learning mechanism. This achieves a synergistic unity of data privacy protection, physical consistency, cross-condition knowledge sharing, edge-side real-time performance, and full lifecycle adaptation.

[0010] The technical solution adopted by this invention to solve its technical problem is as follows: A method for collaborative training of parameters and hierarchical deployment of edge-cloud federated thermal state estimation models driven by physical information is provided, including the following steps: S1: At multiple edge nodes, a local thermal state estimation model is trained using locally collected battery operating data. The loss function during training includes a data fitting term and a physical constraint term based on the battery thermodynamic equation. S2: The trained local model parameters are transmitted to the server after privacy protection processing, and the raw data of battery operation data is not included in the transmission process. S3: The server executes a federated aggregation algorithm based on the model parameters received from multiple edge nodes to update the global thermal state estimation model; S4: Distribute the updated global thermal state estimation model to multiple edge nodes; S5: Edge nodes perform local adaptation processing on the received global model based on local battery operation data; S6: Perform lightweight compression on the locally adapted model to generate an edge inference model, and deploy it in the embedded controller of the battery management system. S7: The deployed end-side inference model receives real-time measurement data from the battery module and outputs thermal state estimation results.

[0011] Furthermore, in step S1, the physical constraint term is constructed based on the discretized residual of the unsteady-state thermal conduction equation of the lithium-ion battery, specifically in the form of: ; ; in, For the first Spatial nodes at Temperature at any moment Where is the thermal diffusivity, For the Laplace operator, For heat generation rate, For density, Specific heat capacity; The total loss function is: ; in, For data fitting terms, This is the balance coefficient, with a value ranging from 0.01 to 0.1.

[0012] Furthermore, in step S2, the privacy protection method is one or more of differential privacy noise, homomorphic encryption, or secure multi-party computation.

[0013] Furthermore, in step S3, the federated aggregation employs a federated averaging algorithm or an improved variant thereof, the formula of which is: ; in, The number of participating nodes, For nodes The number of local samples, , For nodes In the The model parameters uploaded in rounds; Each time the server performs aggregation, it may randomly select some nodes from all edge nodes to participate, or select all nodes to participate.

[0014] Furthermore, in step S3, the federated aggregation adopts hierarchical federated aggregation: First-level aggregation: Multiple edge nodes within the same region upload model parameters to the regional edge server for the first aggregation; Second-layer aggregation: The edge servers in each region upload the aggregated model parameters to the global server in the cloud for a second aggregation.

[0015] Furthermore, in step S5, the local adaptation process includes: incrementally fine-tuning the global model using local battery operating data to obtain a personalized thermal state estimation model adapted to local operating conditions.

[0016] Furthermore, the incremental fine-tuning only updates some layer parameters of the global model, while freezing the parameters of the remaining layers.

[0017] Furthermore, in step S6, the lightweight compression includes at least one of the following: knowledge distillation, network pruning, and parameter quantization.

[0018] Furthermore, it also includes federated continuous learning: repeatedly executing the loop of steps S4, S5, S2, and S3 to continuously update the global model online.

[0019] A system for estimating the thermal state of lithium-ion batteries employs a physically-informed, federated thermal state estimation model parameter co-training and edge-cloud hierarchical deployment method. The system includes edge nodes and servers. The edge nodes are deployed on the battery system side to train models locally based on collected battery operation data, and receive global models from the server for local adaptation processing to generate end-side inference models for the battery management system to call. The server communicates with multiple edge nodes to collect model parameters uploaded by each edge node for federated aggregation, and distributes the aggregated and updated global model to each edge node. Data transmission between the edge node and the server is based on model parameters, while battery operation data is stored locally.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Each edge node only uploads model parameters instead of raw battery running data, which solves the problem that raw data cannot be separated from local data. It further enhances privacy protection through mechanisms such as differential privacy and homomorphic encryption, fundamentally eliminating the risk of data leakage and solving the privacy compliance problem faced by centralized training. 2. Federated learning aggregates thermal state knowledge of battery systems from different regions, vehicle models, and charging / discharging strategies. The global model absorbs the thermal diffusion characteristics under various operating conditions, and each edge node can benefit from the "experience" of other nodes, significantly improving the model's generalization ability under unknown operating conditions. Experiments show that the prediction error of the global model after federated aggregation is reduced by about 60% compared with the single-node model. 3. The physical constraint term enables the model to output temperature predictions that conform to thermodynamic laws even in the early stages when data is sparse or the response to measurable temperature points is delayed. This mechanism provides a more accurate temperature rise trend in the early stages of thermal runaway, improving the timeliness of early warning by more than 30% and reducing the false alarm rate by about 62%. 4. Through lightweight compression techniques such as knowledge distillation, pruning, and quantization, the generated edge-side inference model can achieve millisecond-level inference (typical value 10~15ms) on the BMS embedded MCU, without relying on network connection, ensuring the reliability and real-time performance of thermal runaway early warning; 5. The federated continuous learning mechanism enables the model to continuously absorb changes in thermal characteristics after battery aging (such as changes in thermal behavior caused by increased internal resistance and capacity decay), solving the problem of decreased accuracy of traditional static models with aging. 6. The hierarchical federated aggregation architecture is suitable for ultra-large-scale deployment scenarios, such as fleets of tens of thousands of vehicles and energy storage networks of hundreds of stations, reducing communication volume by more than 90% while meeting regional compliance requirements that data does not leave the domain. Attached Figure Description

[0021] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a schematic diagram of the three-layer system architecture of the edge-cloud of the present invention; Figure 3 This is a diagram of the hierarchical federated aggregation architecture of the present invention (applicable to large-scale deployment scenarios). Figure 4 This is a schematic diagram of the local thermal state estimation model structure of the present invention; Figure 5 This is a timing diagram of the federated continuous learning mechanism of this invention. Detailed Implementation

[0022] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] like Figure 1 As shown, this embodiment provides a method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model, including the following steps:

[0025] At multiple edge nodes (each corresponding to a battery system, such as an electric vehicle, an energy storage container, or a battery swapping station), local thermal state estimation models are trained using locally collected battery operation data, such as... Figure 4 The diagram shown is a schematic diagram of the local thermal state estimation model of the present invention.

[0026] Input data: Time-series temperature data, total voltage, total current, and state of charge (SOC) at multiple temperature measurement points within the battery module. Optional data includes ambient temperature.

[0027] Model structure: Temporal neural network, such as Temporal Convolutional Network (TCN), Long Short-Term Memory Network (LSTM), or Transformer. The input consists of features from the current and multiple historical moments (e.g., the past 10 seconds), and the output consists of predicted temperature values ​​and temperature field distribution features for all temperature measurement points (including virtual internal points not directly measured) at multiple future moments.

[0028] Loss function design: Data fitting term : Mean square error (MSE) between predicted and measured temperatures; Physical constraints Discretized residual construction based on the one-dimensional unsteady-state heat conduction equation of lithium-ion battery.

[0029] The physical constraint term is constructed based on the discretized residual of the unsteady-state thermal conduction equation of a lithium-ion battery, and its specific form is as follows: ; ; in, For the first Spatial nodes at Temperature at any moment Where is the thermal diffusivity, For the Laplace operator, For heat generation rate, For density, Specific heat capacity; The total loss function is: ; in, For data fitting terms, This is the balance coefficient, with a value ranging from 0.01 to 0.1.

[0030] Training process: Optimization algorithms such as Adam are used to perform multiple rounds of iterative training on local historical data until the loss converges.

[0031] In the local training of step S1, a multi-task learning framework is used to optimize both branches simultaneously: Data-driven branch: Learn temperature change patterns related to operating conditions from local runtime data.

[0032] Physics-driven branch: Utilizes the residuals of the heat conduction equation to provide prior regularization, without relying on label data.

[0033] The two branches are optimized collaboratively in the loss function, which enables the model to maintain thermodynamic rationality under data sparse conditions (such as newly added battery packs) and improves the physical consistency of the global model after federated aggregation.

[0034] S2: Upload Model Parameters Each edge node transmits the trained local model parameters (weights and biases) to the server after privacy protection processing.

[0035] Privacy protection methods include one or more of the following: Differential privacy: Add Gaussian or Laplace noise that satisfies (ϵ,δ)-differential privacy to the model gradient or weights.

[0036] Homomorphic encryption: Parameters are transmitted in encrypted form, and the server aggregates them in encrypted form.

[0037] Secure multi-party computation: Multiple nodes collaborate to aggregate parameters without leaking the parameters of a single node.

[0038] Key features: The transmitted content consists only of model parameters and does not contain any raw battery operating data, such as raw temperature, voltage, and current sequences.

[0039] S3: Federated aggregation updates the global model The server (global server in the cloud or regional edge server) executes a federated aggregation algorithm based on the model parameters received from multiple edge nodes to update the global thermal state estimation model.

[0040] Aggregation algorithm: FedAvg or its improved variants (such as FedProx, which adds near-end terms during aggregation to deal with non-independent and identically distributed data) can be used.

[0041] The formula for the federal average algorithm is: ; in, The number of participating nodes, For nodes The number of local samples, , For nodes In the The model parameters uploaded in rounds; Each time the server performs aggregation, it may randomly select some nodes from all edge nodes to participate, or select all nodes to participate.

[0042] like Figure 2 As shown, this invention adopts a three-layer system architecture: the edge-cloud layer includes multiple battery modules and their embedded inference modules; the edge layer includes multiple edge nodes (each node contains a local training module, a federated communication module, and a model compression module); and the cloud layer includes a global federated server (containing a parameter aggregation module, a global model repository, a model distribution module, and a node management module). Based on this, as... Figure 3 As shown, when the number of edge nodes is large, the federated aggregation adopts a hierarchical federated aggregation: First-level (regional) aggregation: Multiple edge nodes within the same geographical region or the same management domain upload local model parameters to the regional edge server (such as a battery swapping station server or a park server). The regional edge server performs the first federated aggregation to obtain the regional model parameters. Second-layer (cloud-level) aggregation: The edge servers in each region upload the aggregated model parameters to the global server in the cloud for a second aggregation, resulting in a global thermal state estimation model.

[0043] The advantages of this solution are: it significantly reduces cloud communication pressure (communication volume can be reduced by more than 90%), and data within the region does not leave the regional boundary, thus meeting the data localization compliance requirements.

[0044] S4: Global Model Distribution The server distributes the updated global thermal state estimation model to each edge node.

[0045] Edge nodes perform local adaptation processing on the received global model based on local battery operating data. Specifically, the local adaptation processing includes: incrementally fine-tuning the global model using local battery operating data to obtain a personalized thermal state estimation model adapted to local operating conditions.

[0046] Fine-tuning strategy: Typically, the shallow network parameters of the global model (these layers learn general features) are frozen, and only the parameters of the deep fully connected layers are updated; gradient descent is performed in a small number of epochs (e.g., 3-5 epochs) using a small amount of new local data (e.g., 1000 samples).

[0047] Function: To enable the global model to quickly adapt to the aging state and recent operating conditions of the local battery by absorbing the common knowledge of other nodes.

[0048] S6: Lightweight Compression and End-Side Deployment The fine-tuned personalized model or the global model is directly compressed to generate an edge inference model, which is then deployed in the embedded controller (MCU or NPU) of the battery management system.

[0049] The lightweight compression includes at least one of the following: knowledge distillation, network pruning, and parameter quantization.

[0050] Knowledge distillation: Using the finely tuned model as the teacher model, train a simpler student model (e.g., with more than 80% fewer parameters) so that the student model output approximates the teacher model output.

[0051] Network pruning: Remove redundant neuron connections or convolutional channels, prune according to importance score, and then fine-tune to restore accuracy.

[0052] Parameter quantization: Quantize floating-point parameters into INT8 or INT4, significantly reducing storage and computational overhead.

[0053] Adaptive Compression: Based on the computing power level of the target hardware (such as different BMS models), the compression ratio and quantization bit depth are automatically selected to generate the optimal lightweight model. Due to the significant differences in BMS hardware across different edge nodes, an adaptive compression strategy is designed: Before deployment, the edge node automatically detects the computing power of the target hardware (MCU clock speed, memory size, and whether it supports NPU acceleration), selects the appropriate combination of compression techniques and quantization bit depth according to a preset level mapping table, and generates a lightweight model adapted to that hardware. For nodes supporting NPUs, the model is further converted to an NPU-specific format (such as ONNX, TFLite, TensorRT) to utilize hardware acceleration.

[0054] The deployed end-side inference model receives real-time measurement data from the battery module, outputs thermal state estimation results, and issues a thermal runaway warning signal based on a preset threshold.

[0055] The output includes the current and short-term forecast of the highest temperature, average temperature distribution, rate of temperature rise, remaining safe operating time, and thermal runaway risk level.

[0056] Warning logic: Multi-level threshold settings (e.g., 60℃ Level 1 warning, 70℃ Level 2 warning, 80℃ Level 3 warning; warning triggered when the temperature rise rate exceeds 3℃ / s or 5℃ / s). The warning signal is output to the vehicle controller or energy storage management system via CAN bus or hard-wired interface to trigger active cooling, power limiting, or emergency power cut-off.

[0057] like Figure 5The diagram shown illustrates the timing of the federated continuous learning mechanism of this invention. To adapt to the aging changes throughout the battery's lifecycle, a lifelong learning closed loop is designed: repeatedly executing the cycle of "global model distribution → edge node incremental fine-tuning → model parameter upload → server federated aggregation". The cycle period is configurable (e.g., once daily, once weekly, or event-triggered). This mechanism enables the global model to continuously absorb newly added battery aging data and continuously distribute the updated model to each edge node, achieving online model updates in line with battery aging.

[0058] Example 1: Cooperative Estimation of Thermal State of Electric Vehicle Fleets Scene description: A ride-sharing company owns 1,000 electric vehicles, each equipped with a Battery Management System (BMS). This system collects data on the battery module's voltage, current, state of charge (SOC), and six NTC temperature measurement points via a CAN bus, with a sampling frequency of 1Hz. The vehicles communicate with a cloud server via 4G / 5G networks.

[0059] Step 1: Local model initialization and physical constraint construction Each vehicle's BMS controller pre-installs a Temporal Convolutional Network (TCN) as the foundation for its local thermal state estimation model. The input layer receives feature data from the past 10 seconds (a total of 9 features: current, voltage, SOC, ambient temperature, and 6 temperature measurement points), with an input tensor size of (10, 9). The model contains 3 dilated convolutional layers (dilation rates of 1, 2, and 4), each followed by batch normalization and ReLU activation. Finally, a fully connected layer outputs the predicted temperature values ​​for the next 5 seconds from 8 measurement points (6 actual measurement points plus 2 virtual interior points).

[0060] A physical constraint term is added to the loss function: based on the one-dimensional heat conduction equation along the battery thickness direction, it is discretized using the finite difference method, and the mean square value of the residual between the predicted temperature field and the heat conduction equation is calculated. The total loss function is: ; in, The value is 0.1.

[0061] Step 2: Local Training Each vehicle's TCN model was trained using operational data accumulated over the past 7 days (approximately 600,000 samples) with the Adam optimizer, an initial learning rate of 0.001, a batch size of 64, and a training duration of 10 epochs. Model parameters were saved after training.

[0062] Step 3: Federated Aggregation The cloud-based global server initiates a federated aggregation process daily at 2:00 AM. 200 vehicles are randomly selected from 1000 vehicles to participate in this round of aggregation (to reduce communication load). Each vehicle uploads its model parameters (weights and biases) to the cloud via TLS encryption. The server executes a federated averaging algorithm (FedAvg) to generate new global model parameters by weighting the parameters according to the number of samples reported by each vehicle. After aggregation is complete, the global model is distributed to all 1000 vehicles.

[0063] Step 4: Incremental Fine-tuning and Lightweight Deployment After each vehicle receives the global model, it performs incremental fine-tuning using a small amount of newly generated data (approximately 1000 samples) from that day. During fine-tuning, the parameters of the first two dilated convolutional layers are frozen, and only the last dilated convolutional layer and the fully connected layer are updated. The learning rate is set to 0.0001, and training is conducted for 3 epochs.

[0064] After fine-tuning, knowledge distillation is performed: using the fine-tuned model as the teacher model, a simpler student model (3 convolutional layers plus 2 fully connected layers, reducing the number of parameters from 2.3M to 0.4M) is trained. Further INT8 quantization is then performed, compressing the model size to 0.1MB. The quantized model is then deployed on the STM32F4 series MCU of the BMS.

[0065] Step 5: Real-time Reasoning and Early Warning The deployed edge model receives the latest sliding window data every 10ms (updated in real time), completes inference within 15ms, and outputs the highest temperature, average temperature, and temperature rise rate for the next 5 seconds. The warning logic is set as follows: a Level 1 warning (yellow) is issued when the predicted highest temperature exceeds 60℃, a Level 2 warning (orange) when it exceeds 70℃, and a Level 3 warning (red) when it exceeds 80℃; an emergency warning is triggered directly when the temperature rise rate exceeds 5℃ / s. The warning signal is sent to the vehicle controller via the CAN bus, triggering the fan or liquid cooling pump to accelerate and limiting drive power.

[0066] Operational Results: After three months of federated continuous learning (aggregation once a day), the mean absolute error (MAE) of temperature prediction for the fleet decreased from an initial 2.1℃ to 0.7℃. The average early warning time for precursors of thermal runaway (abnormal temperature fluctuations) was 45 seconds, and the false alarm rate was reduced by 62% compared to the traditional single-node model.

[0067] Example 2: Hierarchical Federated Learning for Distributed Energy Storage Power Stations Scene description: A power grid operator owns 50 distributed energy storage power stations, each containing 10-20 battery clusters. Each power station is equipped with a local edge server (industrial computer). Communication bandwidth between power stations is limited (10Mbps shared), and it is required that battery operation data within a power station must not leave the station's boundaries (localization compliance requirement).

[0068] Step 1: First layer of polymerization (inside the power plant) Within each energy storage power station, the BMS of each battery cluster acts as an end-side device. Following steps 1-2 of Example 1, it performs local training and uploads the model parameters to the power station's edge server. The power station's edge server performs internal federated aggregation (FedAvg) to obtain the power station-level model parameters.

[0069] Step 2: Second-layer aggregation (cloud) The edge servers of 50 power plants upload their respective power plant-level model parameters to the global server in the cloud, and perform a second federated aggregation to obtain a global thermal state estimation model.

[0070] Step 3: Global model distribution and local adaptation The cloud distributes the global model to the edge servers of each power station. Each power station's edge server performs minor adjustments to the global model based on the data distribution characteristics of its own power station before distributing it to the BMS of each battery cluster within the power station for lightweight deployment and real-time inference.

[0071] Beneficial effects: The cloud communicates directly with 50 edge servers instead of hundreds of BMSs, reducing the number of communication connections by more than 80%; data within the power plant never leaves the plant, fully meeting localization compliance requirements. Experimental data shows that the temperature prediction accuracy of the global model after hierarchical aggregation is comparable to that of direct aggregation across all nodes, but cloud communication volume is reduced by approximately 95%.

[0072] Example 3: Cross-battery pack knowledge transfer in a battery swapping station scenario Scene description: In a certain battery swapping station network, battery packs are frequently transferred between different vehicles, and newly deployed battery packs lack historical operating data (cold start problem).

[0073] Step 1: Building a Global Knowledge Base The cloud-based global server has accumulated thermal state knowledge covering various battery models and aging stages (SOH from 90% to 70%) through federated learning (from various battery swapping stations and operating vehicles), forming a global model.

[0074] Step 2: Cold start of the new battery pack When a new battery pack is first put into a battery swapping station, the cloud directly sends the global model to its BMS as the initial thermal state estimation model. This model already includes "general thermal behavior knowledge" of other battery packs and can provide an initial estimate (MAE of about 1.2℃, while the cold start error of traditional pure data-driven methods is as high as 3.5℃) without local historical data.

[0075] Step 3: Online Adaptation After a few charge-discharge cycles (about 1 day), the new battery pack uses locally accumulated data to incrementally fine-tune the model (only updating the deep fully connected layers), so that the model transitions from "general knowledge" to "personalized knowledge", and the thermal state estimation accuracy is gradually improved to below 0.8℃ MAE.

[0076] Beneficial effects: The cold start phase does not require intensive data collection, which significantly shortens the adaptation period for new battery packs and improves the operational efficiency of battery swapping stations.

[0077] Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-mentioned technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-mentioned technical features or their equivalent features without departing from the inventive concept. For example, technical solutions formed by substituting the above-mentioned features with technical features disclosed in this application (but not limited to) that have similar functions.

Claims

1. A method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model, characterized in that, Includes the following steps: S1: At multiple edge nodes, a local thermal state estimation model is trained using locally collected battery operating data. The loss function during training includes a data fitting term and a physical constraint term based on the battery thermodynamic equation. S2: The trained local model parameters are transmitted to the server after privacy protection processing, and the raw data of battery operation data is not included in the transmission process. S3: The server executes a federated aggregation algorithm based on the model parameters received from multiple edge nodes to update the global thermal state estimation model; S4: Distribute the updated global thermal state estimation model to multiple edge nodes; S5: Edge nodes perform local adaptation processing on the received global model based on local battery operation data; S6: Perform lightweight compression on the locally adapted model to generate an edge inference model, and deploy it in the embedded controller of the battery management system. S7: The deployed end-side inference model receives real-time measurement data from the battery module and outputs thermal state estimation results.

2. The method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model according to claim 1, characterized in that, In step S1, the physical constraint term is constructed based on the discretized residual of the unsteady-state thermal conduction equation of the lithium-ion battery, specifically in the form of: ； ； in, For the first Spatial nodes at Temperature at any moment Where is the thermal diffusivity, For the Laplace operator, For heat generation rate, For density, Specific heat capacity; The total loss function is: ； in, For data fitting terms, This is the balance coefficient, with a value ranging from 0.01 to 0.

1.

3. The method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model according to claim 1, characterized in that, In step S2, the privacy protection method is one or more of differential privacy noise, homomorphic encryption, or secure multi-party computation.

4. The method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model according to claim 1, characterized in that, In step S3, the federated aggregation employs a federated averaging algorithm or an improved variant thereof, the formula of which is: ； in, The number of participating nodes, For nodes The number of local samples, , For nodes In the The model parameters uploaded in rounds; Each time the server performs aggregation, it may randomly select some nodes from all edge nodes to participate, or select all nodes to participate.

5. The method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model according to claim 1, characterized in that, In step S3, the federated aggregation adopts hierarchical federated aggregation: First-level aggregation: Multiple edge nodes within the same region upload model parameters to the regional edge server for the first aggregation; Second-layer aggregation: The edge servers in each region upload the aggregated model parameters to the global server in the cloud for a second aggregation.

6. The method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model according to claim 1, characterized in that, In step S5, the local adaptation process includes: incrementally fine-tuning the global model using local battery operating data to obtain a personalized thermal state estimation model adapted to local operating conditions.

7. The method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model according to claim 6, characterized in that, The incremental fine-tuning only updates some layer parameters of the global model, while freezing the parameters of the remaining layers.

8. The method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model according to claim 1, characterized in that, In step S6, the lightweight compression includes at least one of the following: knowledge distillation, network pruning, and parameter quantization.

9. The method for collaborative training of parameters and hierarchical deployment of edge-cloud in a physically-informed federated thermal state estimation model according to claim 1, characterized in that, It also includes federated continuous learning: repeatedly executing the loop of steps S4, S5, S2, and S3 to continuously update the global model online.

10. A system for estimating the thermal state of a lithium-ion battery, employing the method described in any one of claims 1-9, characterized in that, The system includes edge nodes and servers; The edge nodes are deployed on the battery system side to train models locally based on collected battery operation data, and receive global models from the server for local adaptation processing to generate end-side inference models for the battery management system to call. The server communicates with multiple edge nodes to collect model parameters uploaded by each edge node for federated aggregation, and distributes the aggregated and updated global model to each edge node. Data transmission between the edge node and the server is based on model parameters, while battery operation data is stored locally.

Images