Intelligent function recovery method and system after new energy automobile collision

Abstract

The invention discloses an intelligent function recovery method and system for a new energy automobile after collision, and belongs to the technical field of new energy automobile safety control. The method comprises the following steps: collecting collision data through a plurality of sensors, and generating a comprehensive collision situation assessment result by adopting a data fusion algorithm; diagnosing a vehicle power supply network state and actuator availability, and generating a resource state vector; splicing the comprehensive collision situation assessment result and the resource state vector into a state space, generating a plurality of function recovery sequences based on a resource perception type online Q learning algorithm, calculating a comprehensive utility value of each sequence, and selecting an optimal sequence; and executing a function recovery instruction of the selected optimal sequence, and updating the decision model of the Q learning algorithm on line according to an actual execution result. According to the method, the problem of key function loss caused by high-voltage power failure after collision is solved, the optimal function recovery sequence is selected in a resource limited environment through intelligent decision, and the survival probability of passengers is remarkably increased.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent function recovery of new energy vehicles after a collision, belonging to the field of new energy vehicle safety control technology. Background Technology

[0002] Collision safety of pure electric vehicles is a key concern in the industry. Existing technologies can reliably achieve high-voltage power-off after a collision through multi-channel collision signal fusion diagnosis, avoiding the risk of short circuits and fires. However, high-voltage power-off can cause a sudden drop or interruption in the vehicle's low-voltage network voltage, leading to the failure of critical functions such as door locks, windows, lights, and communication modules, severely hindering occupant self-rescue and external rescue.

[0003] Existing solutions are mostly rule-based controls with fixed priorities. For example, patent CN111231774A provides a reliable solution for high-voltage power cutoff and low-voltage power maintenance after a collision, but its functional recovery logic still relies on preset rules and lacks dynamic response capabilities to vehicle damage status, remaining energy, and occupant conditions. Another patent, CN112590723A, while implementing an automatic distress call function, has a linear decision-making process that does not consider complex scenarios with limited energy and partial actuator failure. In such cases, this fixed sequence may exhaust energy due to ineffective attempts, thus missing rescue opportunities.

[0004] Therefore, existing technologies struggle to intelligently select the "critical chain of survival" with the highest success rate in the chaotic and resource-constrained environment following a collision. Under conditions of extremely limited energy, a single failed functional attempt could deplete emergency power, leading to a complete loss of rescue opportunities. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes an intelligent function recovery method and system for new energy vehicles after a collision. This method can solve the problem of loss of critical functions caused by high-voltage power failure after a collision. By making intelligent decisions to select the optimal function recovery sequence in resource-constrained environments, it significantly improves the survival rate of occupants.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: In a first aspect, an embodiment of the present invention provides a method for restoring the intelligent functions of a new energy vehicle after a collision, comprising the following steps: Step S1: Collect collision data through multiple sensors and generate a comprehensive collision situation assessment result using a data fusion algorithm; Step S2: Diagnose the vehicle power network status and actuator availability, and generate a resource status vector; Step S3: The comprehensive collision situation assessment result and the resource state vector are concatenated into a state space. Multiple functional recovery sequences are generated based on the resource-aware online Q-learning algorithm. The comprehensive utility value of each sequence is calculated and the optimal sequence is selected. Step S4: Execute the function recovery instruction of the selected optimal sequence, and update the decision model of the Q-learning algorithm online based on the actual execution results.

[0007] As one possible implementation of this embodiment, step S1 includes the following steps: Step S11: After the collision event is triggered, raw data from at least four different types of sensors are collected in real time. Step S12: Use a data fusion algorithm to perform real-time fusion calculation on the raw data collected from at least four different types of sensors to generate a structured situation vector.

[0008] As one possible implementation of this embodiment, the data fusion algorithm includes a weighted Bayesian estimation algorithm or a lightweight one-dimensional convolutional neural network model, wherein the lightweight one-dimensional convolutional neural network model includes: The input layer receives a one-dimensional sensor data sequence with dimension n. The first one-dimensional convolutional layer is configured with 16 convolutional kernels of size 3, using the ReLU activation function; The first one-dimensional max pooling layer uses max pooling operation with a pooling size of 2; The second one-dimensional convolutional layer is configured with 32 convolutional kernels of size 3, using the ReLU activation function; The second one-dimensional max pooling layer uses max pooling operation with a pooling size of 2; The flattening layer converts multidimensional feature tensors into one-dimensional feature vectors. The fully connected layer contains 64 neurons and uses the ReLU activation function; The output layer has the same number of neurons as the structured situation vector and uses either the Softmax or Sigmoid activation function.

[0009] As one possible implementation of this embodiment, step S2 includes the following steps: Step S21: Quickly query the vehicle power network status. The vehicle power network status includes reading the current voltage value and estimated remaining capacity of the low-voltage battery, and reading the current voltage and remaining charge percentage of the emergency supercapacitor group that is specifically designed to power the system after a collision. Step S22: Send status query commands to each body controller via low-speed CAN bus or LIN bus to diagnose the availability of key actuators; Step S23: Summarize the current voltage and remaining charge percentage of the emergency supercapacitor bank and the diagnostic results of the key actuators into a resource status vector.

[0010] As one possible implementation of this embodiment, the resource state vector includes: Power status components: low-voltage battery voltage, remaining capacity percentage, emergency supercapacitor bank voltage, remaining charge percentage; Actuator status components: status of each door lock, status of each window, status of the lighting system, and status of the communication module; Comprehensive evaluation components: total available energy estimate, critical actuator availability score, and overall system health status.

[0011] As one possible implementation of this embodiment, step S3 includes the following steps: Step S31: Based on the preset priority objectives and combined with the comprehensive collision situation assessment results, the rule engine generates multiple feasible function recovery action sequences, which are called candidate critical chains. Step S32: A resource-aware online Q-learning algorithm is used to calculate the comprehensive utility value for each candidate key chain; Step S33: Compare the overall utility values ​​of all candidate key chains and select the key chain with the highest utility value as the final execution plan.

[0012] As one possible implementation of this embodiment, the reward function of the resource-aware online Q-learning algorithm... Designed as follows: , in, It is manually set based on the execution result of action 'a' and its contribution to the ultimate goal. This is the estimated total energy consumption of action chain a. It is the estimated total time of action chain a. It is a risk penalty item used to assess the potential risks of performing action a in state s; , , The weighting coefficients were determined through simulation and vehicle testing based on a large amount of historical collision data.

[0013] As one possible implementation of this embodiment, step S4 includes the following steps: Step S41: According to the optimal critical chain determined by the decision, the function recovery instructions are executed sequentially through the corresponding vehicle domain controller or direct drive actuator. Step S42: Monitor the execution result of each instruction in real time, calculate the actual value of the reward function R based on the actual execution result, and use the actual reward value to update the decision model of the Q-learning algorithm online.

[0014] Secondly, an embodiment of the present invention provides a new energy vehicle post-collision intelligent function recovery system for performing any of the methods described above, the system comprising: The multi-source perception and fusion module is used to realize the multi-source collision situation perception and fusion steps, collect multi-source sensor data and perform fusion processing to generate a comprehensive collision situation assessment result. The resource diagnostic module is used to implement a rapid resource availability diagnostic step, diagnose the vehicle power network status and actuator availability, and generate a resource status vector. The decision processing module includes a processor and a memory. The processor is configured to call and execute the reinforcement learning algorithm to realize dynamic key chain generation and intelligent decision-making steps. The memory is used to store the decision model of the reinforcement learning algorithm. The instruction execution module is used to implement the instruction execution and online learning steps, control each executor to complete the function recovery action, and update the decision model of the reinforcement learning algorithm based on the execution results.

[0015] Thirdly, the present invention provides a new energy electric vehicle, including the new energy vehicle collision intelligent function recovery system as described above.

[0016] The beneficial effects of the technical solutions of the embodiments of the present invention are as follows: 1. By combining multi-source sensing with intelligent decision-making, the optimal functional recovery sequence selection is achieved in resource-constrained environments after a collision; 2. Employing reinforcement learning algorithms, multi-objective dynamic optimization can be achieved under limited energy conditions; 3. The online learning mechanism enables the system to continuously improve and learn from actual collision events; 4. Compared with traditional fixed rule methods, it significantly improves the success rate of function recovery and energy utilization efficiency. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for restoring intelligent functions of a new energy vehicle after a collision, according to an exemplary embodiment. Figure 2 This is a schematic diagram illustrating the principle of an intelligent function recovery system for new energy vehicles after a collision, according to an exemplary embodiment. Figure 3 This is a logical connection diagram illustrating the intelligent function recovery process of a new energy vehicle after a collision, according to an exemplary embodiment. Figure 4 This is a schematic diagram illustrating a comparison of the utility of different candidate key chains at a specific decision point, according to an exemplary embodiment. Figure 5 This diagram illustrates the difference in remaining battery power over time between the strategy of this invention and the conventional strategy during a collision event. Detailed Implementation

[0018] To more clearly illustrate the technical features of the present invention, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following description is merely exemplary and is not intended to limit the invention.

[0019] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for restoring the intelligent functions of a new energy vehicle after a collision, comprising the following steps: Step S1: Collect collision data through multiple sensors and generate a comprehensive collision situation assessment result using a data fusion algorithm; Step S2: Diagnose the vehicle power network status and actuator availability, and generate a resource status vector; Step S3: The comprehensive collision situation assessment result and the resource state vector are concatenated into a state space. Multiple functional recovery sequences are generated based on the resource-aware online Q-learning algorithm. The comprehensive utility value of each sequence is calculated and the optimal sequence is selected. Step S4: Execute the function recovery instruction of the selected optimal sequence, and update the decision model of the Q-learning algorithm online based on the actual execution results.

[0020] As one possible implementation of this embodiment, step S1 includes the following steps: Step S11: After the collision event is triggered, raw data from at least four different types of sensors are collected in real time. Step S12: Use a data fusion algorithm to perform real-time fusion calculation on the raw data collected from at least four different types of sensors to generate a structured situation vector.

[0021] As one possible implementation of this embodiment, the raw data from the at least four different types of sensors includes: Vehicle dynamics sensor data, including triaxial acceleration, angular velocity, and yaw rate data collected by multi-axis acceleration sensors and vehicle attitude angular velocity sensors; Restraint system sensor data, including multi-stage detonation signals from the airbag controller; Battery pack sensor data, including internal voltage, temperature, insulation resistance, and short-circuit detection signals of the battery pack; Occupant and environmental sensor data, including occupant image information captured by in-vehicle cameras, seat pressure sensor signals, and impact sounds or cries for help captured by microphones.

[0022] As one possible implementation of this embodiment, the data fusion algorithm includes a weighted Bayesian estimation algorithm or a lightweight one-dimensional convolutional neural network model, wherein the lightweight one-dimensional convolutional neural network model includes: The input layer receives a one-dimensional sensor data sequence with dimension n. The first one-dimensional convolutional layer is configured with 16 convolutional kernels of size 3, using the ReLU activation function; The first one-dimensional max pooling layer uses max pooling operation with a pooling size of 2; The second one-dimensional convolutional layer is configured with 32 convolutional kernels of size 3, using the ReLU activation function; The second one-dimensional max pooling layer uses max pooling operation with a pooling size of 2; The flattening layer converts multidimensional feature tensors into one-dimensional feature vectors. The fully connected layer contains 64 neurons and uses the ReLU activation function; The output layer has the same number of neurons as the structured situation vector and uses either the Softmax or Sigmoid activation function.

[0023] As one possible implementation of this embodiment, step S12 further includes a data preprocessing sub-step: Data alignment, normalization, and missing value processing are performed on the collected multi-source heterogeneous data; The processed multi-source heterogeneous data are spliced ​​together according to the sensor channel order to form a fixed-length one-dimensional input vector.

[0024] As one possible implementation of this embodiment, the structured situation vector includes at least: The severity of a collision is classified into four levels: minor, moderate, severe, and fatal. Vehicle damage prediction includes assessment of deformed parts of the vehicle body, jammed door lock mechanisms, and battery pack intrusion risk. Occupant status assessment, including the conscious status, degree of injury, and entrapment status of each occupant; The risk level of the surrounding environment includes the vehicle's location, rollover risk, and secondary disaster risk.

[0025] As one possible implementation of this embodiment, step S2 includes the following steps: Step S21: Quickly query the vehicle power network status. The vehicle power network status includes reading the current voltage value and estimated remaining capacity of the low-voltage battery, and reading the current voltage and remaining charge percentage of the emergency supercapacitor group that is specifically designed to power the system after a collision. Step S22: Send status query commands to each body controller via low-speed CAN bus or LIN bus to diagnose the availability of key actuators; Step S23: Summarize the current voltage and remaining charge percentage of the emergency supercapacitor bank and the diagnostic results of the key actuators into a resource status vector.

[0026] As one possible implementation of this embodiment, the diagnosis of the availability of the key actuator includes: Diagnose each door lock motor to see if it can communicate and if it reports a stuck fault. Diagnose each window lift motor to see if it can communicate and if it reports mechanical blockage. A lighting system, including hazard lights and hazard warning lights, is used to diagnose whether it responds. The communication module diagnoses whether it is online and has good signal strength. The communication module includes a cellular network SOS module and a V2X module.

[0027] As one possible implementation of this embodiment, the diagnosis of the power network status further includes: Health status assessments are conducted on low-voltage batteries and emergency supercapacitor banks, and the remaining service life is predicted based on historical usage data and current measurement parameters. Calculate the total available energy and allocate the energy budget based on the current voltage, capacity, and estimated load demand.

[0028] As one possible implementation of this embodiment, the resource state vector includes: Power status components: low-voltage battery voltage, remaining capacity percentage, emergency supercapacitor bank voltage, remaining charge percentage; Actuator status components: status of each door lock, status of each window, status of the lighting system, and status of the communication module; Comprehensive evaluation components: total available energy estimate, critical actuator availability score, and overall system health status.

[0029] As one possible implementation of this embodiment, step S23 further includes resource status confidence assessment: The reliability of the resource state vector is evaluated based on the consistency of responses from various sensors and actuators. When the confidence level is lower than a preset threshold, a secondary diagnosis is triggered or a conservative resource estimation strategy is adopted.

[0030] As one possible implementation of this embodiment, step S3 includes the following steps: Step S31: Based on the preset priority objectives and combined with the comprehensive collision situation assessment results, the rule engine generates multiple feasible function recovery action sequences, which are called candidate critical chains. Step S32: A resource-aware online Q-learning algorithm is used to calculate the comprehensive utility value for each candidate key chain; Step S33: Compare the overall utility values ​​of all candidate key chains and select the key chain with the highest utility value as the final execution plan.

[0031] As one possible implementation of this embodiment, the preset priority target includes: The highest priority is for occupants to extricate themselves from the predicament. Initiating external rescue is the second highest priority; Delaying secondary disasters is the next highest priority.

[0032] As one possible implementation of this embodiment, the candidate critical chain is generated based on the following: collision severity level; vehicle damage prediction results; occupant status assessment results; and actuator availability information in the resource state vector.

[0033] As one possible implementation of this embodiment, the state space S of the resource-aware online Q-learning algorithm is defined as follows: A multidimensional feature vector formed by concatenating the comprehensive collision situation assessment results with the resource state vector; Action space A is defined as the set of all candidate key chains generated in step S31.

[0034] As one possible implementation of this embodiment, the reward function of the resource-aware online Q-learning algorithm... Designed as follows: , in, It is manually set based on the execution result of action 'a' and its contribution to the ultimate goal. This is the estimated total energy consumption of action chain a. It is the estimated total time of action chain a. It is a risk penalty item used to assess the potential risks of performing action a in state s; , , The weighting coefficients were determined through simulation and vehicle testing based on a large amount of historical collision data.

[0035] As one possible implementation of this embodiment, the Q-value update of the Q-learning algorithm uses the following formula: , Where α is the learning rate and γ is the discount factor. It is the new state after action A is performed.

[0036] As one possible implementation of this embodiment, step S4 includes the following steps: Step S41: According to the optimal critical chain determined by the decision, the function recovery instructions are executed sequentially through the corresponding vehicle domain controller or direct drive actuator. Step S42: Monitor the execution result of each instruction in real time, calculate the actual value of the reward function R based on the actual execution result, and use the actual reward value to update the decision model of the Q-learning algorithm online.

[0037] As one possible implementation of this embodiment, the execution of the function recovery instruction includes: Control commands are sent to each actuator via CAN bus or LIN bus; Monitor the execution status of commands, including success, failure, and timeout. When an instruction fails to execute, it is handled according to a preset fault tolerance strategy.

[0038] As one possible implementation of this embodiment, the monitoring of the execution result includes: Real-time monitoring of the response status of each actuator; Record the actual execution time of each instruction; Monitor system resource consumption, including energy consumption and communication load.

[0039] As one possible implementation of this embodiment, the actual value of the reward function R is calculated based on the following: the actual success rate of instruction execution; the difference between actual energy consumption and estimated energy consumption; the difference between actual execution time and estimated time; and risk events that occur during execution.

[0040] As one possible implementation of this embodiment, the decision model for the online Q-learning algorithm includes: Based on the actual reward value, update the Q value of the corresponding state-action pair in the Q table according to the Q value update formula; Adjust the policy network parameters to enhance the tendency to select successful action sequences under similar conditions.

[0041] like Figure 2 As shown in the embodiment of the present invention, an intelligent function recovery system for new energy vehicles after a collision is provided, used to execute any of the methods described above. The system includes: The multi-source perception and fusion module is used to realize the multi-source collision situation perception and fusion steps, collect multi-source sensor data and perform fusion processing to generate a comprehensive collision situation assessment result. The resource diagnostic module is used to implement a rapid resource availability diagnostic step, diagnose the vehicle power network status and actuator availability, and generate a resource status vector. The decision processing module includes a processor and a memory. The processor is configured to call and execute the reinforcement learning algorithm to realize dynamic key chain generation and intelligent decision-making steps. The memory is used to store the decision model of the reinforcement learning algorithm. The instruction execution module is used to implement the instruction execution and online learning steps, control each executor to complete the function recovery action, and update the decision model of the reinforcement learning algorithm based on the execution results.

[0042] The present invention provides a new energy electric vehicle, including the new energy vehicle collision intelligent function recovery system as described above.

[0043] This invention constitutes a tightly integrated, closed-loop feedback intelligent decision-making system, the core logic of which is: (a) Perception first: Through multi-source collision situational awareness and fusion steps, the system first forms a comprehensive understanding of the collision event, vehicle damage, occupant status and environmental risks, providing information input for subsequent decision-making.

[0044] (b) Diagnostic Foundation: The subsequent rapid resource availability diagnostic step, based on cognition, quickly clarifies the currently available energy and actuator resources of the system and defines the physical constraints of functional recovery.

[0045] (c) Intelligent decision-making: The dynamic critical chain generation and intelligent decision-making steps are the core of the system. It receives the output of the first two steps as its state input and uses reinforcement learning algorithms to weigh success rate, energy consumption, time and risk among many possible action sequences, dynamically generating and selecting the recovery plan with the highest comprehensive utility.

[0046] (d) Execution optimization: The final instruction execution and online learning steps are responsible for executing the decision and feeding the execution results back to the reinforcement learning model for online policy updates, so that the system can learn from each collision event and continuously optimize its decision-making ability in similar situations in the future.

[0047] The above four steps form a complete closed loop from "environmental perception" to "resource assessment", then to "intelligent decision-making" and finally "execution learning", ensuring the system's adaptability, efficiency and reliability under extreme conditions.

[0048] like Figure 3 As shown, the specific process of restoring the intelligent functions of a new energy vehicle after a collision according to the present invention includes the following five logically connected steps, and the steps are linked by a clear information flow.

[0049] Step S100, Environmental Perception and Situational Awareness: The system first collects collision data from multiple sources, including vehicle dynamics sensors, constraint system sensors, battery pack sensors, and occupant and environmental sensors. A weighted Bayesian estimation algorithm or a lightweight one-dimensional convolutional neural network model is then used to fuse the multi-source heterogeneous data in real time, generating a structured comprehensive collision situation assessment result. This step provides environmental state input for subsequent decision-making.

[0050] Specifically, step S100 includes the following steps: S101. Data Acquisition Sub-step: After a collision event is triggered, raw data from at least four different types of sensors is acquired in real time: a) Vehicle dynamics sensors: including multi-axis acceleration sensors and vehicle attitude angular velocity sensors, used to detect collision acceleration, force and rollover state, and collect triaxial acceleration (±50g range), angular velocity (±2000rad / s) and yaw rate data; b) Restraint system sensors: receive multi-stage detonation signals from the airbag controller (including pretensioners, primary airbags, and secondary airbags); c) Battery pack sensors: monitor the internal voltage, temperature, insulation resistance, and short-circuit detection signals of the battery pack; d) Occupant and environmental sensors: Occupant image information collected by in-vehicle cameras, seat pressure sensor signals, and impact sounds or cries for help collected by microphones.

[0051] S102. Data Fusion and Situation Assessment Sub-step: A lightweight one-dimensional convolutional neural network model is used to perform real-time fusion calculations on the multi-source heterogeneous data collected in step S101; The lightweight one-dimensional convolutional neural network model is a pre-trained compressed model with fewer than 100k parameters. Its input is a one-dimensional sequence of multi-sensor data, and its output is a structured situation vector. It is specifically designed for the real-time processing requirements of automotive embedded platforms. Its specific implementation process is as follows: (1) Data preprocessing and input construction: After a collision event is triggered, raw data collected in real time from at least four of the following sources: vehicle dynamics sensors, restraint system sensors, battery pack sensors, and occupant and environmental sensors. First, data alignment, normalization, and missing value processing are performed. Then, the processed multi-source heterogeneous data is concatenated according to the sensor channel order to form a fixed-length one-dimensional input vector, represented as: , Where n is the total number of sensor data channels, The standardized data represents the i-th sensor channel.

[0052] (2) Model forward propagation and feature extraction: The preprocessed one-dimensional vector is input into the lightweight one-dimensional convolutional neural network model. The model performs forward propagation calculations through multiple network layers connected in sequence, automatically extracting and fusing deep features from the data.

[0053] The structure and function of each layer of the lightweight one-dimensional convolutional neural network model are as follows: Input layer: Receives one-dimensional sensor data sequences with dimension (n); The first one-dimensional convolutional layer: configured with 16 convolutional kernels of size 3, using the ReLU activation function, extracts local spatiotemporal feature patterns from the sensor sequence through one-dimensional convolution operations; The first one-dimensional max pooling layer: uses max pooling operation with a pooling size of 2 to downsample the feature map, retaining significant features and reducing data dimensionality; The second one-dimensional convolutional layer is configured with 32 convolutional kernels of size 3, using the ReLU activation function to perform deeper feature combination and abstraction based on the primary features; The second one-dimensional max pooling layer: Max pooling with a pooling size of 2 is used again to further compress the feature dimension and enhance the translation invariance of the model; Flattening layer: Converts the multidimensional feature tensor after convolution and pooling into a one-dimensional feature vector so that it can be input into the fully connected layer; Fully connected layer: Contains 64 neurons, uses the ReLU activation function, and performs nonlinear combination and information integration on all the high-level features extracted above; Output layer: The number of neurons is consistent with the dimension of the structured situation vector. The softmax activation function (used for mutually exclusive classification, such as collision level) or the sigmoid activation function (used for multi-label classification, such as simultaneously predicting injury site and occupant status) is used to output the comprehensive collision situation assessment result.

[0054] (3) Model output and situation analysis: The model output is the structured situation vector, which directly corresponds to the comprehensive collision situation assessment result, including at least four dimensions of quantitative assessment: collision severity level, vehicle damage prediction, occupant status assessment, and surrounding environment risk level.

[0055] The lightweight one-dimensional convolutional neural network model, through end-to-end training, can directly learn effective feature representations for situation assessment from raw sensor data, avoiding the complex manual feature design and probabilistic model construction that rely on prior knowledge in traditional methods, and has stronger adaptability and higher fusion accuracy.

[0056] The model outputs a comprehensive collision situation assessment, which includes at least: a) Collision severity levels: minor, moderate, severe, fatal; b) Vehicle damage prediction, such as: deformation of the left side of the vehicle body, possible jamming of the right front door locking mechanism, and risk of battery pack intrusion; c) Occupant status assessment, such as: driver occupant is conscious, front passenger occupant is unconscious, and there are no occupants in the rear seats; d) Risk level of the surrounding environment, such as: the vehicle is located inside the tunnel and there is a risk of rollover.

[0057] Step S200, Resource Boundary Diagnosis: Based on the environmental awareness gained in step S100, the system rapidly diagnoses the power network status and actuator availability. Power status diagnosis includes reading the low-voltage battery voltage and the emergency supercapacitor bank charge; actuator diagnosis queries the status of each key actuator via bus communication. The diagnostic results are summarized into a resource status vector, clearly defining the physical constraints for functional recovery.

[0058] Specifically, step S200 includes the following steps: S201. Power Status Diagnosis Sub-step: Quickly check the vehicle's power network status, including: Read the current voltage value of the 12V low-voltage battery and estimate the remaining capacity; Read the current voltage and remaining charge percentage of the emergency supercapacitor bank specifically designed to power the system after a collision, such as... Figure 5 As shown.

[0059] S202. Actuator and Load Diagnostic Sub-step: Status query commands are sent to each body controller via low-speed CAN bus or LIN bus to diagnose the availability of critical actuators, including: Diagnose whether each door lock motor can communicate and whether it reports a stuck fault. Diagnose whether each window lift motor can communicate and whether it reports mechanical blockage; Diagnose the lighting system, including hazard lights and flashers, to see if they are responding. Diagnose whether the communication modules, cellular network SOS module, V2X module, etc., are online and have good signal strength.

[0060] S203. Resource aggregation sub-step: The diagnostic results of S201 and S202 are summarized into a resource status vector.

[0061] Step S300, Intelligent Planning and Optimization Decision: The situation assessment result output in step S100 is concatenated with the resource state vector output in step S200 to form a complete state space. Based on a resource-aware online Q-learning algorithm, multiple candidate function recovery sequences are generated and the comprehensive utility value of each sequence is calculated. The algorithm employs an ε-greedy strategy to balance exploration and utilization, weighing success rate, energy consumption, time, and risk through a reward function, and selecting the sequence with the highest utility value as the execution plan.

[0062] Specifically, step S300 includes the following steps: S301. Target Definition and Candidate Action Chain Generation Sub-Steps: Based on the preset priority objectives: occupants autonomously escape > initiate external rescue > delay secondary disasters, and combined with the situation assessment results output in step S100, the rule engine generates multiple feasible functional recovery action sequences, which are called candidate critical chains. Example: If the situation assessment is "driver conscious, left front door deformed, battery low risk", then the generated candidate chain may include: Candidate chain 1: Attempt to unlock all doors -> lower the driver's side window -> activate the headlight alarm; Candidate Chain 2: Immediately send a rescue message via SOS -> activate the light alarm -> attempt to lower the car window.

[0063] S302. Utility calculation sub-steps based on reinforcement learning: A resource-aware online Q-learning algorithm embedded in the vehicle domain controller is used to calculate the comprehensive utility value of each candidate critical chain for controlling the vehicle actuators, so as to achieve precise scheduling of limited power, avoid ineffective energy consumption, and extend the survival window of critical functions. The state space (S) of the algorithm is defined as a multi-dimensional feature vector formed by concatenating the comprehensive collision situation assessment result output in step S100 with the resource state vector output in step S200. The action space (A) of the algorithm is defined as: the set of all candidate key chains generated in step S301; The reward function of the algorithm Designed as follows: , in, It is a basic reward item, which is manually set according to the execution result of action 'a' and its contribution to the ultimate goal. Successfully completing key actions, such as sending SOS, will be given a high positive reward, while failure to execute or triggering system protection will be given a high negative reward. It is the estimated total energy consumption of action chain a, calculated based on the nominal power consumption and estimated running time of each actuator; It is the estimated total time of action chain a. It is a risk penalty item used to assess the potential risks of performing action a in state s, such as the risk of short circuit caused by attempting to raise or lower a car window when the battery is hot. , , This is a weighting coefficient used to balance the relative importance of rewards and various costs. Its specific value is determined through simulation and whole-vehicle testing optimization based on a large amount of historical collision data.

[0064] The Q-value update of the algorithm uses the following formula: , Where α is the learning rate (value 0.1) and γ is the discount factor (value 0.9). It is the new state after action A is performed.

[0065] S303. Decision Sub-step: Compare the overall utility value (Q value) of all candidate critical chains, and select the critical chain with the highest utility value as the final execution plan.

[0066] Step S400: Execution Verification and Experience Accumulation Following the optimal sequence determined in step S300, function recovery commands are executed sequentially via the vehicle domain controller. The execution results are monitored in real time, and reward values ​​are calculated based on the actual performance, updating the Q-learning algorithm's decision model online. This step forms a closed-loop feedback loop from theoretical decision-making to practical verification.

[0067] Specifically, step S400 includes the following steps: S401. Instruction execution sub-step: The optimal critical chain determined in step S303 is used to sequentially execute function recovery commands through the corresponding body domain controller (BDCU) or direct drive actuator.

[0068] S402. Results Monitoring and Learning Sub-steps: Monitor the execution result of each instruction in real time: success, failure, timeout; Based on the actual execution results, the actual value of the reward function R is calculated; Using this actual reward value, the Q value of the corresponding state-action pair in the Q table is updated online according to the Q value update formula in S302, thereby achieving real-time optimization of the algorithm strategy.

[0069] Step S500, Process Traceability and System Optimization: Throughout the process, key data for each step is recorded in parallel, including timestamps, system status, selected actions, reward values, Q-values, and remaining battery power. A data analysis report is output after the event concludes, providing data support for system verification and continuous optimization.

[0070] Specifically, step S500 includes the following steps: S501. Operation Data Recording Sub-step: Throughout the entire system operation, continuously record the following data to non-volatile memory in time sequence: Timestamp, system state vector, selected action, Q-value of all candidate actions, instant reward obtained, and remaining battery power.

[0071] S502. Post-event analysis output sub-step: After the event processing is completed, provide an interface to output the recorded data, which is used to generate utility comparison bar charts and resource consumption curves for each decision point, for system verification, debugging and performance optimization.

[0072] Steps S100 to S500 constitute a complete closed-loop intelligent decision-making system, wherein: The output of step S100 serves as the input condition for step S200, guiding the targeted execution of resource diagnosis. The outputs of steps S100 and S200 are used together as the input to step S300 to construct a complete decision state space; The output of step S300 serves as the basis for the execution of step S400, driving the specific function recovery action; The execution result of step S400 is fed back to step S300 for online optimization of the decision model; Step S500 records the running data of each step in parallel, providing data support for system optimization.

[0073] S100 execution: Data collected: X-axis acceleration -35g, airbag deployment signal is true, battery temperature 32℃ with no short circuit, camera image identified by CNN shows the driver and passenger are conscious but have a painful expression, microphone analysis shows no violent impact sound. Fusion model output situation vector: [Collision level = moderate, predicted damage = left side vehicle deformation, battery risk = low, driver status = conscious and trapped, environmental risk = ordinary road].

[0074] S200 diagnostics: The diagnostics show a 12V voltage of 11.8V and an emergency capacitor charge of 70%. The left front door lock motor is unresponsive, indicating damage or a circuit break. The left front window motor is unresponsive, while the right rear door motor responds normally. The SOS module signal strength is good.

[0075] Execute S301: Based on the above states, generate 3 candidate chains: A: Try unlocking the left front door -> lowering the left front window (Note: the success rate is extremely low, possibly due to actuator failure). B: Unlock all unlocked doors -> Lower all available windows -> Activate hazard lights; C: Immediately send location and occupant status via SOS -> activate hazard lights -> attempt to unlock the right rear door.

[0076] Execute S302: The algorithm calculates the utility of each chain. For example... Figure 4 As shown, chain A has the lowest utility value due to its low success rate and high energy consumption. Chains B and C need to be compared: chain B provides more escape routes but has slightly higher energy consumption; chain C can call for rescue the fastest and has an extremely high success rate. Calculations show that chain C has the highest utility value, or Q value.

[0077] Execute S303: Select chain C as the execution plan.

[0078] Execute S401: Execute the following in sequence: SOS message sent successfully (reward +100), dual flash activation successful (reward +20), right rear door unlocked successfully (reward +80). Total reward R = 200. Energy consumption is approximately 15mAh.

[0079] Execute S402: Update the Q table based on the actual result (success, power consumption 15mAh) to reinforce the preference for chain C under similar conditions.

[0080] Execution of S501: Record status, actions, rewards, and battery level changes throughout the process. Post-event analysis showed that choosing chain C at the decision point was the optimal solution, avoiding energy waste on the damaged left front door.

[0081] This invention solves the problem of critical function loss caused by high-voltage power failure after a collision. By making intelligent decisions to select the optimal function recovery sequence in resource-constrained environments, it significantly improves the chances of occupant survival.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A new energy vehicle post-collision intelligent function recovery method, characterized in that, The method comprises the following steps: Step S1, collecting collision data through multiple sensors, and generating a comprehensive collision situation assessment result by using a data fusion algorithm; Step S2, diagnosing the state of the vehicle power supply network and the availability of actuators, and generating a resource state vector; Step S3, concatenating the comprehensive collision situation assessment result and the resource state vector into a state space, generating multiple functional recovery sequences based on a resource-aware online Q-learning algorithm, calculating the comprehensive utility value of each sequence and selecting the optimal sequence; Step S4, executing the functional recovery instructions of the selected optimal sequence, and updating the decision model of the Q-learning algorithm online according to the actual execution result. 2.The new energy vehicle post-collision intelligent function recovery method according to claim 1, characterized in that, The step S1 comprises the following steps: Step S11, after the collision event is triggered, real-time collection of raw data from at least four different types of sensors; Step S12, real-time fusion calculation of the collected raw data of at least four different types of sensors by using a data fusion algorithm to generate a structured situation vector. 3.The new energy vehicle post-collision intelligent function recovery method according to claim 2, characterized in that, The data fusion algorithm comprises a weighted Bayesian estimation algorithm or a lightweight one-dimensional convolutional neural network model, and the lightweight one-dimensional convolutional neural network model comprises: An input layer for receiving a one-dimensional sensor data sequence with a dimension of n; A first one-dimensional convolutional layer configured with 16 convolutional kernels with a size of 3 and using a ReLU activation function; A first one-dimensional maximum pooling layer using a maximum pooling operation with a pooling size of 2; A second one-dimensional convolutional layer configured with 32 convolutional kernels with a size of 3 and using a ReLU activation function; A second one-dimensional maximum pooling layer using a maximum pooling operation with a pooling size of 2; A flattening layer for converting a multi-dimensional feature tensor into a one-dimensional feature vector; A fully connected layer containing 64 neurons and using a ReLU activation function; An output layer with a number of neurons consistent with the dimension of the structured situation vector and using a Softmax or Sigmoid activation function.

4. The method for recovering intelligent functions after a collision of a new energy vehicle according to claim 1, characterized in that, The step S2 comprises the following steps: Step S21, quickly querying the state of the vehicle power supply network, which comprises reading the current voltage value and estimating the remaining capacity of the low-voltage storage battery, and reading the current voltage and the remaining capacity percentage of the emergency supercapacitor group specially used for powering the post-crash system; Step S22, sending a state query instruction to each body controller through a low-speed CAN bus or a LIN bus to diagnose the availability of key actuators; Step S23, collecting the current voltage and the remaining capacity percentage of the emergency supercapacitor group and the diagnosis results of the key actuators into a resource state vector. 5.The method of claim 1, wherein, The resource state vector comprises: Power supply state components: low-voltage storage battery voltage, remaining capacity percentage, emergency supercapacitor group voltage, remaining capacity percentage; Actuator state components: the state of each door lock, the state of each window, the state of the lighting system, and the state of the communication module; Comprehensive evaluation components: total available energy estimation, key actuator availability score, and overall system health status. 6.The method of claim 1, wherein, The step S3 comprises the following steps: Step S31, according to the preset priority target, combining the comprehensive collision situation assessment result, and generating multiple feasible functional recovery action sequences by a rule engine, referred to as candidate key chains; Step S32, using resource-aware online Q-learning algorithm to calculate the comprehensive utility value for each candidate critical chain; Step S33, comparing the comprehensive utility values of all candidate critical chains, selecting the critical chain with the highest utility value as the final execution scheme. 7.The method of claim 6, wherein, The resource-aware online Q-learning algorithm's reward function is designed to: ， wherein, is manually set according to the execution result of the action a and its contribution to the ultimate goal, is the estimated total energy consumption of the action chain a, is the estimated total time consumption of the action chain a, is a risk penalty term for evaluating the risk that may be caused by executing the action a in the state s; , , is a weight coefficient, which is determined by a large amount of historical collision data simulation and vehicle test optimization.

8. The method of claim 1-7, wherein, The step S4 includes the following steps: Step S41, according to the optimal critical chain determined, executing the function recovery instructions in sequence through the corresponding body domain controller or directly driving the actuators; Step S42, monitoring the execution results of each instruction in real time, calculating the actual value of the reward function R according to the actual execution results, and updating the decision model of the Q-learning algorithm online using the actual reward value.

9. A new energy vehicle post-collision intelligent function recovery system, configured to perform the method according to any one of claims 1-8, characterized in that, Comprise: A multi-source perception and fusion module for realizing the multi-source collision situation perception and fusion step, collecting multi-source sensor data and performing fusion processing to generate a comprehensive collision situation assessment result; A resource diagnosis module for realizing the resource availability rapid diagnosis step, diagnosing the vehicle power supply network state and actuator availability, and generating a resource state vector; A decision processing module including a processor and a memory, the processor being configured to call and execute the reinforcement learning algorithm, realize the dynamic critical chain generation and intelligent decision step, and the memory being used to store the decision model of the reinforcement learning algorithm; An instruction execution module for realizing the instruction execution and online learning step, controlling the actuators to complete the function recovery action, and updating the decision model of the reinforcement learning algorithm according to the execution results.

10. A new energy electric vehicle, characterized in that, The intelligent function recovery system for new energy vehicles after collision comprises the system according to claim 9.

Images