Electric vehicle energy efficiency optimization method and system based on fault prediction

Abstract

The invention provides an electric vehicle energy efficiency optimization method and system based on fault prediction, and the method comprises the steps: collecting the operation data of an electric vehicle in real time, and predicting the fault probability of the electric vehicle according to the operation data, constructing an optimization objective function including minimization of energy consumption cost, minimization of battery life attenuation and minimization of travel time, and setting constraint conditions of the optimization objective function according to the fault probability; solving the optimization objective function by using a local model of the electric vehicle to obtain a charging strategy and a driving strategy of the electric vehicle, and pushing the charging strategy and the driving strategy to a user side; taking the adoption rate and the adjustment record of the push strategy by the user as a log file, removing the position information and the user information in the log file to obtain desensitization data, uploading the desensitization data to a cloud, calculating the difference weight of each user, aggregating the local models of all the users according to the difference weight, and obtaining a desensitization result; and updating local models of all the electric vehicles according to the cross-vehicle knowledge in the aggregation model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy vehicles, and particularly relates to an energy efficiency optimization method and system for electric vehicles based on fault prediction. BACKGROUND

[0002] The global new energy vehicle industry is expanding rapidly, and the demand for intelligent operation and maintenance of electric vehicles is increasing. As the most core component of the vehicle, the fault warning precision and energy efficiency management level of the power battery system directly determine the safety and economy of the vehicle. However, how to balance between efficient and accurate fault warning and comprehensive intelligent energy efficiency optimization, and make them real-time collaborative, is still a problem that needs to be solved in the industry.

[0003] The traditional battery management system (BMS) has long relied on a threshold alarm mechanism of a single parameter such as voltage and temperature, which cannot capture abnormal signals in the early stage of major faults such as thermal runaway. When the alarm is triggered, the battery is often irreversibly damaged. Industry statistics show that the false alarm rate is higher than 15% all year round due to sensor noise and environmental interference, which weakens the trust of users. These isolated warning logic also lack effective linkage with the vehicle energy efficiency strategy. When the battery is running with a disease, the system cannot adjust the charging power or driving recommendations in time, but may accelerate battery degradation or exacerbate faults due to extreme energy-saving strategies. Energy efficiency management is also one-dimensional: charging plans and driving tips only refer to static SOC, ignoring dynamic price fluctuations, and lacking quantitative assessment of driving behaviors such as sudden acceleration and frequent braking; single-target optimization algorithms are difficult to balance energy consumption costs, travel time, and battery life, leading to strategies deviating from real-world scenarios. In addition, traditional solutions generally do not include real-time health status and potential fault risk in energy efficiency decisions, highlighting the contradiction between energy saving and safety. In order to break through the above limitations, the industry has begun to adopt cloud AI models on a large scale as a new technical path in recent years. By aggregating massive amounts of vehicle data, cloud platforms can train fault diagnosis and energy efficiency optimization models in parallel, and then use OTA to re-inject models or strategies to the vehicle end, in order to achieve higher-dimensional joint analysis.

[0004] Although the cloud intelligent solution can handle massive amounts of data, it still has some problems: the network delay caused by the return of full data makes the fault diagnosis response time exceed 200ms, which cannot meet the real-time requirements of vehicle-level; if the high-performance computing unit supporting the original model is deployed locally, the hardware cost of BMS will increase dramatically, hindering the implementation of the solution. In addition, the cloud overload architecture separates the high-frequency collaboration between fault warning and energy efficiency optimization - the vehicle end can only periodically pull strategies, and cannot dynamically adjust energy-saving goals according to the instantaneous health status. SUMMARY

[0005] The application provides an energy efficiency optimization method and system for electric vehicles based on fault prediction, which solves the problem of poor real-time performance of existing cloud intelligent solutions.

[0006] To solve the above technical problems, the application provides an energy efficiency optimization method for an electric vehicle based on fault prediction, comprising the following steps: Step S1: Real-time collection of operation data of the electric vehicle, prediction of a fault probability of the electric vehicle according to the operation data, construction of an optimization objective function including minimization of energy consumption cost, minimization of battery life attenuation and minimization of travel time, and setting of a constraint condition of the optimization objective function according to the fault probability; Step S2: Solving of the optimization objective function by using a local model of the electric vehicle to obtain a charging strategy and a driving strategy of the electric vehicle, and pushing of the charging strategy and the driving strategy to a user end; Step S3: Recording of a user's adoption rate and adjustment of the pushed strategy as a log file, removal of location information and user information in the log file to obtain desensitized data, uploading of the desensitized data to a cloud end, calculation of a difference weight of each user according to the desensitized data, aggregation of local models of all users according to the difference weight, and updating of the local models of all electric vehicles according to knowledge of the aggregated models across vehicles.

[0007] Preferably, the real-time collection of operation data of the electric vehicle in step S1 comprises the following steps: acquisition of original data of a battery voltage, temperature distribution and acceleration of the electric vehicle, normalization processing of the original data by using a sliding window, and construction of a time series data matrix.

[0008] Preferably, in step S1, electrochemical characteristics and thermodynamic characteristics in the operation data are extracted by using the local model of the electric vehicle, the fault probability of the electric vehicle is predicted according to the electrochemical characteristics and the thermodynamic characteristics, and an expression for predicting the fault probability is as follows: ; In the formula, is the fault probability; is an activation function; is a weight matrix of a full connection layer of the local model; is an electrochemical characteristic; is a thermodynamic characteristic vector; represents splicing of the electrochemical characteristic and the thermodynamic characteristic vector; is a bias term of the full connection layer.

[0009] Preferably, in step S2, an expression of the optimization objective function is as follows: ; ; In the formula, is the optimization objective function; , , , , , , denotes the minimization of energy cost; denotes the minimization of battery life degradation; denotes the minimization of travel time; is the total discrete time steps within the planning horizon; is the charging power at time step ; is the real-time electricity price at time step ; is the average growth rate of battery internal resistance; is the average state of charge during charging or discharging cycles; is the travel time of the electric vehicle; is the weight coefficient; is the total charging time within the planning horizon.

[0010] Preferably, a fuzzy controller is employed in step S2 to dynamically adjust the weights in the optimization objective function according to the battery health status and the user-set driving mode, and the fuzzy rules employed by the fuzzy controller are as follows: (1) if the battery health status is healthy and the user driving mode is economy priority, then increase the weight of , and decrease the weights of and ; (2) if the battery health status is healthy and the user driving mode is balance mode, then maintain the weights of , and ; (3) if the battery health status is healthy and the user driving mode is performance priority, then increase the weight of , and decrease the weights of and ; (4) if the battery health status is mild degradation and the user driving mode is economy priority, then increase the weight of , maintain the weight of , and decrease the weight of ; (5) if the battery health status is mild degradation and the user driving mode is balance mode, then increase the weight of , and maintain the weights of and ; (6) if the battery health status is mild degradation and the user driving mode is performance mode, then increase the weights of and the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased; the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased; (7) if the battery health state is serious degradation and the user driving mode is economy priority, the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased; the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased; the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased; the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased; (8) if the battery health state is serious degradation and the user driving mode is balance mode, the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased; the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased; the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased; the weight of the battery life target is reduced, the weight of the trip efficiency target is maintained, and the weight of the energy consumption target is increased.

[0011] Preferably, after uploading the user's adoption rate of the push strategy and the adjustment record in step S3 to the cloud as a log file, reinforcement learning is adopted to update the fuzzy rules according to the log file, and the expression of the fuzzy rules is updated as follows: ; wherein, is the value table of reinforcement learning; is the state, , is the state space, including the battery health state, the user driving mode preference, the current working condition of the vehicle, and the adoption rate of the last optimization suggestion; is the action, , is the action space, including the weight of the energy consumption optimization target, the weight of the battery life target, and the weight of the trip efficiency target; is the learning rate; is the reward; is the discount factor; represents the next time state obtained by the environment transition after performing the current action ; represents the next time action selected under the state .

[0012] Preferably, the constraint conditions in step S1 include: (1) when the failure probability , the discharge current upper limit of the branch where the faulty cell is located is lowered; (2) when the failure probability , the energy recovery strategy is suspended, and the upper limit of the accelerator pedal response is limited; (3) when the failure probability ​​​​When the high-voltage relay is cut off, the liquid cooling system is controlled to operate at full power.

[0013] Preferably, the expression for calculating the difference weight of each user according to the desensitization data in step S3 is: ; wherein, is the difference weight of the user i ; , , are weighting coefficients of , , respectively; is the relative data volume factor of the vehicle of the user i in the current training cycle; is the quality credibility factor of the uploaded data of the vehicle of the user i ; is the verification loss improvement factor of the local model of the vehicle of the user i after updating; is the number of valid samples uploaded by the user i ; is the weighting coefficient; is the proportion of outliers in the uploaded data of the user i ; is the signal noise fluctuation of the uploaded data of the user i ; is the sampling interval missing rate of the user i ; is the verification loss of the aggregated model before downloading; is the verification loss of the local model of the vehicle of the user i after updating.

[0014] Preferably, in step S3, the update package of the local model of all electric vehicles is generated through a differential compression algorithm by extracting the knowledge across vehicles in the aggregated model, and is downloaded to the vehicle of the user through over-the-air technology (OTA), so as to realize the updating of the local model of the electric vehicle.

[0015] The application also provides an energy efficiency optimization system for electric vehicles based on fault prediction, which is realized based on the above-mentioned energy efficiency optimization method for electric vehicles based on fault prediction and comprises a data acquisition module, a fault prediction module, an optimization strategy generation module, a cloud collaborative updating module and a man-machine interaction module. The data acquisition module: acquires the battery pack voltage, temperature distribution and acceleration data of the vehicle in real time; The fault prediction module: extracts the electrochemical characteristics and thermodynamic characteristics of the vehicle from the data collected by the data acquisition module, and calculates the fault probability of the vehicle; The optimization strategy generation module: constructs a multi-objective optimization function, realizes multi-objective dynamic optimization under fault constraints, and outputs charging time period suggestions, vehicle speed control intervals, power response parameters, etc. of the vehicle; The cloud collaborative update module: calculates the difference weight of the user according to the user's strategy preference, aggregates the local models of all vehicles, updates the global model in the cloud, generates an incremental update package and distributes it to the vehicle end, and realizes the update of the local model of the vehicle; The human-computer interaction module: visually presents the fault level heat map, energy optimization scheme and battery health trend of the vehicle to the user, allows the user to adjust the vehicle speed interval offset and manually modify the charging time period, records the strategy adoption rate and behavior deviation of the user, and updates the driver portrait of the user.

[0016] The beneficial effects of the present application at least include: 1. The optimization calculation is completed by the local model on the vehicle end, avoiding network delay caused by uploading full data, meeting the real-time requirement of vehicle level, and without the need of stacking high-performance computing units on the vehicle end, significantly reducing the hardware cost; 2. The adoption rate and adjustment record are collected through the push-feedback closed loop, the real user preference is quantified into difference weight, so that the local model continuously fits the individual driving habits, and the strategy landing gap is reduced; 3. The cloud only processes desensitized weight and gradient information, which protects privacy and generalizes the driving habits of each user into cross-vehicle knowledge through federated aggregation, and feeds back to the local model of each vehicle, realizing the continuous upgrade of the local model of the vehicle without returning the original working condition big data. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The method flowchart of the embodiment of the present application; Figure 2 The decision logic diagram of the multi-objective optimization algorithm of the embodiment of the present application; Figure 3 The system architecture diagram of the embodiment of the present application; Figure 4 The collaborative interaction diagram of the fault prediction module and the optimization strategy generation module of the embodiment of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0019] AsFigure 1 As shown, the embodiment of the present application provides a method for optimizing energy efficiency of an electric vehicle based on fault prediction, comprising the following steps: Step S1: Real-time acquisition of operation data of the electric vehicle, prediction of fault probability of the electric vehicle according to the operation data, construction of an optimization objective function including minimization of energy consumption cost, minimization of battery life attenuation and minimization of travel time, and setting of constraint conditions of the optimization objective function according to the fault probability.

[0020] Specifically, the power system bus of the electric vehicle is accessed through the on-board automatic diagnostic system OBD-II interface, and data such as battery pack voltage, temperature distribution including single temperature, and acceleration of the vehicle are collected in real time. The original data collected are subjected to sliding window normalization processing to eliminate the dimensional differences of sensors and construct a time series feature matrix. The system continuously monitors and records the above parameters to provide data support for subsequent average temperature calculation and historical baseline construction. The data acquisition process follows the SAE J1939 protocol standard, and the acceleration data of the inertial measurement unit IMU are used to assist in compensating non-standard parameters.

[0021] The electrochemical characteristics and thermodynamic characteristics in the operation data are extracted by using the local model of the electric vehicle, and the preprocessed voltage / current time series data are input into a one-dimensional convolution layer (Conv1D, kernel_size=5) to capture local fluctuation characteristics through a sliding window: ; In the formula, is a trainable convolution kernel weight; is t the voltage sampling value at the moment.

[0022] After maximum pooling (Pool_size=3) dimensionality reduction is performed on the output feature map, an internal resistance change trend spectrum is generated.

[0023] The two-dimensional spatial data of the temperature detection array are rearranged into a space-time sequence: ; In the formula, represents a complete temperature time series data set composed of observation values of all temperature detection units at consecutive time steps; represents the temperature measurement value of the th temperature detection unit at time step , reflecting the dynamic change of the unit over time; represents the total number of sampling time steps, i.e. the length of the time series data.

[0024] The above spatio-temporal sequence is input into a bidirectional LSTM network for spatio-temporal feature fusion. Through the forward and backward network outputs of the bidirectional LSTM, a prediction matrix of the heat propagation rate is finally obtained , which reflects the spatio-temporal evolution law in the thermal runaway process . Key thermodynamic characteristics such as the maximum heat propagation rate, average heat propagation rate, heat diffusion center position and thermal runaway trend are extracted from the matrix, and are integrated into a low-dimensional vector through dimensionality reduction processing such as principal component analysis or autoencoder, which comprehensively represents the heat diffusion state and trend of the battery. The thermodynamic feature vector is obtained by feature extraction .

[0025] At the same time, based on the historical stable temperature data of the battery pack continuously collected under different operating conditions, a dynamically changing temperature baseline is constructed, which represents the expected average temperature under non-fault state based on historical operating data. The temperature warning threshold is dynamically adjusted according to the current state of charge SOC of the battery and : ; In the formula, , are experimental calibration coefficients, which are set to , in the embodiment of the present application. is the deviation of the average temperature of the current battery pack from its historical normal operating temperature baseline.

[0026] - ; In the formula, is the average temperature of the current battery pack obtained in real time.

[0027] The electrochemical and thermodynamic feature vectors are fused, and the failure probability of the electric vehicle is predicted through the full connection layer of the local model: ; In the formula, is the failure probability; is an activation function Sigmoid, used to map the output value to (0, 1); is a weight matrix of the full connection layer; is an electrochemical feature; is a thermodynamic feature vector; is a bias term of the full connection layer.

[0028] According to the size of the failure probability, different response actions of the system are triggered to ensure timely response to potential failure threats. Specifically, when the failure probability is in the interval of 0.4 to 0.6, the system sends a diagnostic fault code (DTC) through the CAN bus, requiring a response time of less than 100 milliseconds; when the failure probability is in the interval of 0.6 to 0.8, the system triggers a voice warning and flashes a red indicator light on the dashboard, requiring a response time of less than 50 milliseconds; when the failure probability is greater than or equal to 0.8, the system immediately cuts off the high-voltage relay and starts the liquid cooling system, requiring a response time of less than 10 milliseconds.

[0029] After the above process completes the multi-modal feature fusion and risk probability evaluation, the diagnosis result is not only used to trigger the hierarchical response mechanism, but also directly as one of the constraint conditions for energy efficiency optimization. Specifically, when the failure probability exceeds 0.4, the system will temporarily reduce the discharge current upper limit of the branch where the fault-related battery is located, and adjust the vehicle power output curve to limit the occurrence of high-load working conditions; when the battery heat diffusion trend exceeds the preset threshold, the system will suspend the energy recovery strategy and adjust the cooling cycle interval to reduce the system thermal stress. The above mechanism ensures that the energy efficiency optimization strategy always operates within the controllable risk range, avoiding the optimization behavior exacerbating the failure risk, thereby building a diagnosis-constraint-optimization linkage control path to realize the coupling and synergy of vehicle running safety and economy.

[0030] The optimization module is deployed on the vehicle-mounted embedded AI processor, and under the safety state constraint, a multi-objective collaborative decision process is performed with battery health, electricity price fluctuation, driving behavior characteristics and navigation road condition data as input, realizing comprehensive optimization of energy consumption cost, battery life and travel efficiency.

[0031] As Figure 2 The decision logic diagram of the multi-objective optimization algorithm of the embodiment of the application is shown, four types of dynamic parameters are fused in the multi-objective optimization function, thereby constructing an input vector space, including: a) Battery health parameters: including capacity attenuation rate, internal resistance growth coefficient and polarization voltage offset of the battery; b) Economic parameters: obtain time-of-use electricity price data from the power grid API, and construct an electricity price-location association matrix combined with the distribution of charging piles; c) Behavior characteristic parameters: through the user's historical driving data, the frequency of sudden acceleration, average braking intensity and cruise speed preference are counted; d) Road condition parameters: integrate real-time traffic flow, slope change and path length information provided by the navigation system.

[0032] The battery health parameter in the above parameters is directly quantified as the degree of battery attenuation and is taken as a safety constraint; the economy parameter directly participates in the calculation and optimization of energy consumption cost; the behavior characteristic parameter and the road condition parameter provide key basis for dynamic weight adjustment of the optimization model, generation of driving strategy and charging path planning.

[0033] The constructed input vector space is the underlying data basis for the generation of the entire dynamic energy efficiency optimization strategy. Through direct variable of the objective function, influence on the weight of the objective function, application of optimization constraints or as the basis for strategy generation, etc., it jointly acts on the collaborative optimization process, thereby achieving comprehensive and dynamic optimization of energy consumption, battery life and travel efficiency.

[0034] The expression of the constructed optimization objective function is: ; ; In the formula, is the optimization objective function; , , are the weights of , , ; represents minimizing energy consumption cost; represents minimizing battery life attenuation; represents minimizing travel time; is the total discrete time step in the planning period; is the charging power of time step ; is the real-time electricity price of time step ; is the average growth rate of the internal resistance of the battery, which is a key indicator of battery aging, and is directly provided by the internal resistance growth coefficient in the battery health parameter; is the average state of charge during charging or discharging cycles; is the driving time of the electric vehicle, which is affected by road condition parameters such as real-time traffic flow and path length; is a weight coefficient used to balance the relative importance of charging time and driving time in total travel time, and its value can be dynamically adjusted according to user driving mode preference; is the total charging time in the planning period.

[0035] Step S2: Use the local model of the electric vehicle to solve the optimization objective function to obtain the charging strategy and driving strategy of the electric vehicle, and push the charging strategy and driving strategy to the user end.

[0036] Specifically, to further improve the accuracy of optimization decision, the three optimization objectives are dynamically adjusted by the fuzzy logic controller. The fuzzy input of target weight includes battery health state and user driving mode preference. The fuzzy logic controller outputs the real-time weight coefficient of the three objective functions according to the rule base, and adjusts the relative importance of each objective according to the user mode and battery health state. The battery health state includes: healthy, mild attenuation, severe attenuation; the user driving mode preference includes: economy priority, balanced mode, performance priority. The combination of the two forms 9 fuzzy rule items as shown in Table 1.

[0037] Table 1 Fuzzy rule base In actual operation, the system continuously collects vehicle operation data, user behavior feedback and optimization strategy execution effect data. Using machine learning method, the initial rule base is iteratively optimized and adaptively adjusted. By analyzing the influence of weight distribution on comprehensive benefit or user satisfaction under different battery states and driving behaviors in historical data, the system automatically learns and refines the rules. Specifically, by analyzing a large amount of actual driving data, the significant influence of battery on life under specific temperature interval or load mode is identified, and the related rules are dynamically adjusted to realize the refinement of weight distribution.

[0038] For fuzzy input and fuzzy output, the corresponding fuzzy sets are defined, and the membership functions of each fuzzy set are set. Specifically, the health state can be divided into good, medium and poor; the weight adjustment can be divided into significant increase, increase, keep, decrease and significant decrease. The membership function maps the exact input value to the membership degree of the fuzzy set, so that the rule base can handle continuous changing input.

[0039] Through the above method, the rule base is constructed and continuously optimized, so that the fuzzy logic controller can dynamically and intelligently adjust the weight of each objective function in multi-objective optimization according to the real-time working condition of the vehicle, thereby generating more adaptive current scene energy efficiency optimization strategy. Based on this, the weight of the above optimization objectives is fused, and a multi-objective Pareto solution set is constructed based on the improved fast elitist multi-objective genetic algorithm NSGA-II. Through the human-computer interaction terminal, the driver is recommended customized driving and charging strategy including recommended speed interval, charging plan period and power response adjustment parameter.

[0040] Meanwhile, based on real-time data, the optimization calculation is performed every 5 minutes. The optimization process generates driving speed interval and charging period recommendations according to the latest battery status, environmental conditions and user preferences. These recommendations are displayed through the vehicle-mounted human-machine interface (HMI), and users can manually adjust parameters such as vehicle speed interval or charging period offset as needed. User adjustment deviation (δ) will be recorded and updated through reinforcement learning. Reinforcement learning algorithm is used to adjust the weight parameters or membership function shape of fuzzy rules, so that the weight distribution of fuzzy controller in the next round of operation is more in line with user preferences and comprehensive performance goals. The expression for updating the rule base through reinforcement learning is: ; wherein, is the value table of reinforcement learning ; is the state, , is the state space, including battery health status, user driving mode preference, current vehicle operating condition and adoption rate of the last optimization recommendation; is the action, , is the action space, including the weight of energy consumption optimization target, the weight of battery life target, and the weight of trip efficiency target; is the learning rate; is the reward; is the discount factor; represents the next time state obtained by the environment transition after executing the current action ; represents the next time action selected by the agent in state according to the policy.

[0041] To ensure that the optimization behavior is executed within the controlled risk range, the risk factor is loaded synchronously in the optimization process. Specifically, the risk factor is a semantic mapping of the failure probability , which is used as a risk constraint indicator in the energy efficiency optimization module. The system calculates the failure probability value, combined with the dynamic early warning threshold , to calculate the risk factor by the following function relationship: ; wherein, the function represents the risk response strategy set in the actual system, mainly manifested as segmented control or conditional judgment, including: (1) When the failure probability When the faulty battery cell is located, the upper limit of the discharge current of the branch is lowered. Specifically, the upper limit of the branch discharge current is reduced by 20% to 35% linearly, the regenerative braking power is limited to less than 20% of the original value, the upper limit torque of the drive pedal is reduced by 10% to 15%, the cooling pump duty cycle is increased to 70% to 80%, the fan duty cycle is increased to 60% to 70%, and the charging rate is limited to below 0.5C.

[0042] (2) When the failure probability At this time, the energy recovery strategy is suspended, the accelerator pedal response limit is restricted, and the battery SOC is maintained between 20% and 90%, with the cell temperature controlled between 5°C and 45°C. Specifically, this is manifested as follows: regenerative braking power is directly shut off; the upper limit of drive torque is limited to 60% to 75% of the rated value, and the upper limit of effective pedal travel is limited to 60% to 70%; the branch discharge current is further reduced to 40% to 60%; the cooling pump speed is increased to 90% duty cycle, the fan to 85% duty cycle, and a uniform temperature cycle is implemented; the charging rate is limited to below 0.3C, and the DC fast charging power is limited to below 30kW.

[0043] (3) When the failure probability When the high-voltage relay is activated, the liquid cooling system will operate at full power. Immediately disconnect the main positive / main negative relays and keep the pre-charge relay disconnected; the liquid cooling pump and fan will operate at full power, and if compressor cooling is available, the frequency will be increased to the maximum until the cell temperature drops below 40°C; all drive and charging requests will be turned off, leaving only necessary low-load functions such as defrosting / dehumidification in the cabin; at the same time, a voice alarm will be triggered and the red light on the instrument panel will flash continuously, and at least 30 seconds of voltage, current, temperature, and DTC diagnostic data will be saved as a fault record.

[0044] Step S3: Collect the user's adoption rate and adjustment records of the push strategy as a log file, remove the location information and user information from the log file to obtain de-identified data, upload the de-identified data to the cloud, calculate the difference weight of each user based on the de-identified data, aggregate the local models of all users based on the difference weight, and update the local models of all electric vehicles based on the cross-vehicle knowledge in the aggregated model.

[0045] Existing battery life prediction methods are limited to static analysis based on a single data source and have not yet established a cross-vehicle data collaboration and continuous model iteration mechanism. The prediction results deviate significantly from the actual degradation patterns, which not only weakens the forward-looking judgment of energy efficiency optimization on battery health but also restricts the continuous evolution of fault diagnosis models.

[0046] To improve the generalization ability and dynamic adaptability of the fault diagnosis model and the optimization model, an embodiment of the present application designs a cloud model training and issuing mechanism based on the fusion of federated learning and transfer learning. The mechanism completes the aggregation and knowledge transfer of multi-vehicle model parameters without uploading original user sensitive data, and synchronizes the model update to the vehicle-end embedded AI processor through the over-the-air technology (OTA), so as to realize the self-evolution of the edge intelligent module.

[0047] The cloud model updating mechanism includes the following steps: Step S31: Daily scheduled upload of data summaries including abnormal segments, energy efficiency strategy execution records, user preference adjustment logs, etc. The data structure does not contain location information or user identity, ensuring privacy compliance.

[0048] Step S32: The cloud platform fuses the local model gradients from multiple vehicles based on the federated average (FedAvg) algorithm to train a unified fault diagnosis model and energy efficiency optimization model. A difference weight coefficient is introduced in the model aggregation process to reflect the differences in training data volume, data quality and model effectiveness of each vehicle. The difference weight coefficient is calculated as follows: ; In the formula, is the difference weight of user i ; , , are weighting coefficients for balancing the contributions of different indicators in the calculation of the difference weight, for balancing the influences of data volume, data quality and model effectiveness, and the sum of the three is 1 and remains stable within the training period, only adjusting when there are significant data fluctuations or business target changes. Specifically, represents the weight allocated to the relative data volume factor, reflecting the importance of the data scale uploaded by user in aggregation; represents the weight allocated to the data quality reliability factor, highlighting the contribution of high-quality, low-noise data to the model aggregation result; represents the weight allocated to the validation loss improvement factor, emphasizing the effectiveness of the local model of user in improving the overall validation performance after updating.

[0049] is the relative data volume factor of the vehicle of user i in the current training period; is the quality reliability factor of the data uploaded by the vehicle of user i ; is the validation loss improvement factor of the vehicle of user ia verification loss improvement factor of the local model of the vehicle after the update; an effective sample number uploaded by the user i an effective sample number uploaded by the user a weighting coefficient an effective sample number uploaded by the user i an effective sample number uploaded by the user an effective sample number uploaded by the user i an effective sample number uploaded by the user an effective sample number uploaded by the user i an effective sample number uploaded by the user a verification loss of the aggregated model before downloading a verification loss of the local model of the vehicle after the update i a verification loss of the local model of the vehicle after the update

[0050] Step S33: On the basis of the global model, the aging law of the battery, the typical driving behavior model of the user and the charging mode preference are extracted to realize knowledge migration across vehicle platforms.

[0051] Step S34: The model update result after training is compressed into an update package by a differential compression algorithm, the update package only contains the model weight change amount, and is pushed to the corresponding vehicle terminal through the Internet of Vehicles communication link and automatically loaded by the OTA upgrade subsystem.

[0052] Step S35: After receiving and loading the update at the vehicle end, a round of quick verification (Quick Test) needs to be performed to ensure model compatibility and stability, and the original model version is automatically replaced after verification, and the update process is transparent to the user.

[0053] Through the above cloud collaborative mechanism, the system can continuously iterate and optimize the model on the basis of maintaining efficient edge response, ensuring that the fault identification logic and energy efficiency strategy always fit the latest vehicle working conditions and user behavior characteristics, forming a cloud-edge integrated continuous learning closed loop.

[0054] In order to improve the explainability and user acceptance of the system, a human-machine interface (HMI) with feedback capability is set in the system, which visualizes the diagnosis results and energy efficiency optimization suggestions to the user, and records the user's response behavior, so that the user's response participates in the adaptive learning process of the system model. The HMI module supports hierarchical information display technology, mainly including the following functional components: Risk state display interface: display early warning information such as current fault level, abnormal position of cell temperature, battery health trend chart, etc. in a graphical way.

[0055] Energy efficiency suggestion interface: dynamically presents the current recommended speed interval, feasible charging window, SOC prediction curve and target optimization weight.

[0056] User preference adjustment module: support user manual adjustment of recommended parameters, such as vehicle speed upper and lower limits, charging time period offset, etc., all user modification operations are recorded in real time, and are returned to the optimization module as strategy offset.

[0057] Behavior adoption record module: the system constructs a driver behavior portrait based on user adoption rate, and is linked with the system strategy adoption reward mechanism, dynamically updates the state-action mapping relationship of the optimization model, and improves the individual adaptation ability.

[0058] When the system detects that the user behavior deviates from the strategy suggestion, such as ignoring energy saving prompts for 3 times in a row, frequently adjusting the acceleration upper limit, etc., the adaptive learning mechanism will be triggered, the behavior portrait will be reconstructed, and the fuzzy controller rule library will be adjusted, so that the strategy recommendation evolves from global generalization to individual customization.

[0059] As shown in Figure 3 , the embodiment of the present application also provides an energy efficiency optimization system for electric vehicles based on fault prediction, which is realized based on the above-mentioned energy efficiency optimization method for electric vehicles based on fault prediction, and includes a data acquisition module, a fault prediction module, an optimization strategy generation module, a cloud collaborative update module and a human-computer interaction module.

[0060] Data acquisition module: real-time acquisition of battery voltage, temperature distribution and acceleration data of the vehicle.

[0061] Fault prediction module: extract the electrochemical characteristics and thermodynamic characteristics of the vehicle from the data collected by the data acquisition module, and calculate the fault probability of the vehicle.

[0062] Optimization strategy generation module: construct a multi-objective optimization function, realize multi-objective dynamic optimization under fault constraints, and output the charging time period suggestion, vehicle speed control interval, power response parameters, etc.

[0063] As shown in Figure 4 , the embodiment of the present application establishes an optimization strategy generation module that integrates battery health (SOH), real-time electricity price, driving behavior characteristics and navigation road conditions under the safety state constraints provided by the fault prediction module. An improved NSGA-II algorithm is used to generate a Pareto optimal solution set, and the optimization weights of energy cost, battery life and travel time are dynamically adjusted by combining a fuzzy logic controller. During vehicle driving, the edge computing unit analyzes driving behavior data in real time, combines cloud electricity price prediction information and road condition updates, and iteratively generates driving speed recommendation interval and charging time period suggestion every 5 minutes. This optimization strategy fully integrates the battery health information from the fault prediction module, ensuring that the optimization process prioritizes battery safety and life. Effectively balance the composite demand of economy, safety and user experience, realize the technical leap of global energy efficiency management from static planning to dynamic adaptation and fault risk avoidance.

[0064] Cloud collaborative update module: calculate the difference weight of the user according to the user's strategy preference, aggregate all local models of the vehicles, update the global model in the cloud, generate an incremental update package and distribute it to the vehicle end, and realize the update of the local model of the vehicle.

[0065] Human-computer interaction module: visually presents the vehicle's fault level heat map, energy consumption optimization scheme, and battery health trend to the user, allows the user to adjust the vehicle speed interval offset and manually modify the charging period, records the user's strategy adoption rate and behavior deviation, and updates the user's driver portrait.

[0066] The technical features of the above embodiments can be combined in any way. To make the description simple, not all possible combinations of the technical features in the above embodiments are described, only the preferred embodiments of the present application are expressed, and the description is more specific and detailed, but it cannot be understood as limiting the scope of the present application. As long as the combination of these technical features does not exist, it should be considered as the scope of the present application.

[0067] It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for optimizing energy efficiency of an electric vehicle based on failure prediction, characterized in that, The method comprises the following steps: Step S1: Real-time acquisition of operation data of the electric vehicle, prediction of the failure probability of the electric vehicle according to the operation data, construction of an optimization objective function including minimization of energy consumption cost, minimization of battery life attenuation and minimization of travel time, and setting of a constraint condition of the optimization objective function according to the failure probability; Step S2: Solving of the optimization objective function by using a local model of the electric vehicle to obtain a charging strategy and a driving strategy of the electric vehicle, and pushing of the charging strategy and the driving strategy to a user end; Step S3: Recording of a user's adoption rate and adjustment of the pushed strategy as a log file, removal of location information and user information in the log file to obtain desensitized data, uploading of the desensitized data to a cloud end, calculation of a difference weight of each user according to the desensitized data, aggregation of local models of all users according to the difference weight, and updating of the local models of all electric vehicles according to knowledge of cross-vehicles in the aggregated model.

2. The method for optimizing energy efficiency of an electric vehicle based on failure prediction according to claim 1, characterized in that: In step S1, the real-time acquisition of the operation data of the electric vehicle comprises the following steps: obtaining raw data of a battery voltage, a temperature distribution and an acceleration of the electric vehicle, performing normalization processing on the raw data by using a sliding window, and constructing a time series data matrix.

3. The method for optimizing energy efficiency of an electric vehicle based on failure prediction according to claim 1, characterized in that: In step S1, the local model of the electric vehicle is used to extract electrochemical characteristics and thermodynamic characteristics in the operation data, and the failure probability of the electric vehicle is predicted according to the electrochemical characteristics and the thermodynamic characteristics. The expression for predicting the failure probability is: ； wherein, is the failure probability; is the activation function; is the weight matrix of the fully connected layer of the local model; is the electrochemical feature; is the thermodynamic feature vector; denotes concatenation of the electrochemical feature and the thermodynamic feature vector; is the bias term of the fully connected layer.

4. The method for optimizing energy efficiency of an electric vehicle based on failure prediction according to claim 1, characterized in that: In step S2, the expression of the optimization objective function is: ； ； wherein is the objective function to be optimized; , , are the weights for , , respectively; denotes the minimization of energy cost; denotes the minimization of battery life degradation; denotes the minimization of travel time; is the total discrete time steps within the planning horizon; is the charging power at time step ; is the real-time electricity price at time step ; is the average growth rate of the battery internal resistance; is the average state-of-charge during charging or discharging cycles; is the travel time of the electric vehicle; is the weight coefficient; is the total charging time within the planning horizon.

5. The method of claim 4, wherein: In step S2, a fuzzy controller is used to dynamically adjust the weight in the optimization objective function according to a battery health state and a driving mode set by a user. The fuzzy rules used by the fuzzy controller are: If the battery health state is healthy and the user driving mode is economy priority, increase the weight of and decrease the weight of and ; If the battery health state is healthy and the user driving mode is balance mode, the weights of , , and are maintained. , and ​ if the battery health state is healthy and the user driving mode is performance priority, increase the weight of , decrease the weight of and ; If the battery health status is slightly degraded and the user's driving mode is set to economy priority, then increase... The weight, maintain The weight is reduced. The weights; If the battery health state is slight attenuation, and the user driving mode is balance mode, the weight of is increased, and the weights of and are maintained; If the battery health status is slightly degraded and the user's driving mode is performance mode, then improve... and The weight is reduced. The weights; If the battery health is severely degraded and the user's driving mode is set to economy priority, then reduce... Weighting, increasing The weight, maintain The weights; If the battery health state is serious degradation, and the user driving mode is balance mode, increase the weight of , and decrease the weight of and ; If the battery health state is severely degraded and the user driving mode is performance priority, increase the weight of and decrease the weight of and .

6. The method of claim 5, wherein the method further comprises: In step S3, after the user's adoption rate and adjustment record of the pushed strategy are uploaded to the cloud end as the log file, reinforcement learning is used to regularly update the fuzzy rules. The expression for updating the fuzzy rules is: ； wherein, is a value table for reinforcement learning . is a state, , is a state space including battery health state, user driving mode preference, vehicle current operating condition and adoption rate of last optimization recommendation; is an action, , is an action space including weight of energy consumption optimization target, weight of battery life target, weight of trip efficiency target; is a learning rate; is a reward; is a discount factor; denotes the next time state obtained by the environment transition after performing the current action . denotes the next time action selected in state .

7. The method for optimizing energy efficiency of an electric vehicle based on failure prediction according to claim 1, characterized in that: In step S1, the constraint condition comprises: When the failure probability is greater than the threshold, the discharge current upper limit of the branch where the failed cell is located is lowered. When the failure probability is greater than the threshold value, the energy recovery strategy is suspended, and the upper limit of the accelerator pedal response is restricted. When the failure probability is greater than the threshold, the high-voltage relay is cut off, and the liquid cooling system is controlled to operate at full power.

8. The method for optimizing energy efficiency of an electric vehicle based on failure prediction according to claim 1, characterized in that: In step S3, the expression for calculating the difference weight of each user according to the desensitized data is: ； In the formula, is the difference weight of the user i ; , , are the weighting coefficients of , , respectively; is the relative data volume factor of the vehicle of the user i in the current training period; is the quality credibility factor of the uploaded data of the vehicle of the user i ; is the verification loss improvement factor of the local model of the vehicle of the user i after the update; is the number of valid samples uploaded by the user i ; is the weighting coefficient; is the proportion of outliers in the uploaded data of the user i ; is the signal noise fluctuation of the uploaded data of the user i ; is the sampling interval missing rate of the user i ; is the verification loss of the aggregated model before downloading; is the verification loss of the local model of the vehicle of the user i after the update.

9. The method for optimizing energy efficiency of an electric vehicle based on failure prediction according to claim 1, characterized in that: In step S3, by extracting the knowledge of cross-vehicles in the aggregated model, an update package of the local models of all electric vehicles is generated by using a differential compression algorithm, and is downloaded to user vehicles by using over-the-air (OTA) technology, so as to realize the updating of the local models of the electric vehicles.

10. A failure prediction based energy efficiency optimization system for electric vehicles, implemented based on a failure prediction based energy efficiency optimization method for electric vehicles according to any one of claims 1 to 9, characterized in that, It comprises: a data acquisition module, a failure prediction module, an optimization strategy generation module, a cloud end collaborative updating module and a man-machine interaction module; The data acquisition module: real-time acquisition of battery pack voltage, temperature distribution and acceleration data of a vehicle; The failure prediction module: extraction of electrochemical characteristics and thermodynamic characteristics of a vehicle from data collected by the data acquisition module, and calculation of a failure probability of the vehicle; The optimization strategy generation module: construction of a multi-objective optimization function, multi-objective dynamic optimization under a failure constraint, and output of a charging time period suggestion, a vehicle speed control interval, a power response parameter and the like of the vehicle; The cloud collaborative updating module: calculates the difference weight of the user according to the strategy preference of the user, aggregates the local models of all vehicles, updates the global model in the cloud, generates an incremental update package and delivers it to the vehicle end, and realizes updating of the local model of the vehicle; The human-computer interaction module: visually presents the fault level heat map of the vehicle, the energy consumption optimization scheme, and the battery health trend to the user, allows the user to adjust the vehicle speed interval offset and manually modify the charging period, records the strategy adoption rate and behavior deviation of the user, and updates the driver portrait of the user.

Images