Distributed driving hovercar failure and instability degree multi-dimensional evaluation method based on deep reinforcement learning

Abstract

The invention relates to the technical field of hovercar control, and discloses a distributed driven hovercar failure and instability degree multi-dimensional evaluation method based on deep reinforcement learning, and the method comprises the steps: obtaining the operation state information of a distributed driven hovercar in road-air switching and adjacent stages; an evaluation model based on a depth deterministic strategy gradient algorithm is constructed, an evaluation process is modeled as an agent and environment interaction process, and the model learns a mapping strategy of evaluation vectors from a hovercar state to a failure and instability state through reinforcement learning training; inputting the state information obtained in the flying car air-ground switching period into the trained evaluation model; and outputting a multi-dimensional failure and instability state evaluation result, quantitatively representing the current potential failure mode and severity level, and comprehensively evaluating the stability margin of each key degree of freedom. According to the invention, the comprehensive safety state of the hovercar in a complex air-ground switching process and a potential failure scene can be accurately evaluated in real time.

Description

Technical Field

[0001] This invention relates to the field of flying car control technology, specifically to a multi-dimensional evaluation method for the failure and instability of distributed-drive flying cars based on deep reinforcement learning. Background Technology

[0002] Flying cars, as a new type of transportation integrating ground driving and aerial flight capabilities, are considered an important development direction for solving urban traffic congestion, improving transportation efficiency, and expanding transportation dimensions. Distributed drive technology, which achieves independent, rapid, and precise control of the driving force of each wheel through hub / wheel-side motors, has already shown great potential in ground vehicles to improve handling stability, energy efficiency, and achieve complex functions (drive anti-slip, yaw moment compensation). Applying distributed drive technology to the ground-based chassis of flying cars, combined with their rotor propulsion system for aerial flight, constitutes a complex multi-actuator, multi-mode, and strongly coupled system—the distributed drive flying car. This highly integrated system also brings severe safety challenges:

[0003] 1. Complexity of land-air mode switching: During takeoff, landing and mode switching, aerodynamic loads and ground loads alternate rapidly. The flight control system and chassis control system need to be closely coupled and coordinated. The stability boundary changes drastically and instability is prone to occur.

[0004] 2. Limitations of Traditional Assessment Methods: Traditional model-based or rule-based fault diagnosis and stability assessment methods struggle to accurately describe the dynamic characteristics of high-dimensional, strongly nonlinear, and time-varying coupled systems like flying cars, especially in dealing with unforeseen combinations of failure modes and complex environmental disturbances. Furthermore, these methods are often computationally intensive, lack real-time performance guarantees, and typically provide relatively singular assessment results, failing to offer fine-grained, multi-dimensional risk assessment information. Deep reinforcement learning, as a cutting-edge technology in artificial intelligence, learns optimal strategies through interaction between the agent and the environment, demonstrating powerful capabilities in handling high-dimensional state spaces, nonlinear dynamics, and uncertainty problems. Applying it to the failure-instability assessment of flying cars holds promise for overcoming the shortcomings of traditional methods.

[0005] Therefore, a method is needed to evaluate the failure modes of distributed-drive flying cars under various operating conditions and their impact on stability in real time, accurately and in multiple dimensions, so as to provide support for the safe operation and intelligent fault-tolerant control of flying cars. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a multi-dimensional evaluation method for the failure and instability of a distributed-drive flying car based on deep reinforcement learning.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A multi-dimensional evaluation method for the failure and instability of a distributed-drive flying car based on deep reinforcement learning includes:

[0009] Acquire sensor data from a distributed-drive flying car, inject failure mode information and corresponding failure and instability state label data to form an operational state dataset for reinforcement learning;

[0010] A failure and instability state evaluation model based on a deep deterministic policy gradient algorithm is constructed, and the evaluation model is regarded as an intelligent agent. A multi-dimensional failure and instability state evaluation vector is defined, and a reward function is designed based on the failure and instability state label data. Finally, a deep deterministic policy gradient framework is formed and trained on the running state dataset.

[0011] The real-time operating status of the distributed-drive flying car during the land-air transition phase or the set potential risk phase is obtained and input into the trained evaluation model;

[0012] The evaluation model outputs a multi-dimensional evaluation vector to assess the vehicle's failure risk, which can be used to trigger alarms, assist in decision-making, and adjust control.

[0013] In one embodiment, the acquisition of sensor data from the distributed-drive flying car, the injection of failure mode information and corresponding failure and instability state label data, to form an operational state dataset for reinforcement learning specifically includes:

[0014] Acquire sensor data from the distributed-drive flying car, including: inertial measurement unit data, positioning data, pitot tube data, radar data, wheel speed, suspension travel, motor current, motor voltage, motor temperature, rotor speed, and rotor pitch angle;

[0015] The failure mode information simulates one or more combined failure scenarios, such as output torque mismatch, partial failure, or complete failure of the distributed drive motor, that may occur during the land-to-air transition phase or during the set potential risk phase.

[0016] The failure and instability state label data are calculated using a multiphysics coupled simulation model to simulate the dynamic behavior during the land-air handover process.

[0017] The sensor data, failure mode information, and failure and instability state label data are combined to form the operating status data.

[0018] In one embodiment, the evaluation model includes a policy network and a value network; the policy network and the value network are trained and updated using an actor-critic framework;

[0019] The training data for the evaluation model should at least cover the land-air switching process of the flying car under different loads, speeds, and environments, including takeoff and landing conditions, so that the agent can learn evaluation strategies applicable to the set key stages.

[0020] In one embodiment, the dimensions and quantification method of the multidimensional evaluation vector are determined according to a preset safety threshold and control requirements. The multidimensional evaluation vector represents the remaining safety margin of stability in each dimension and can be used to classify risk levels.

[0021] Compared with the prior art, the beneficial technical effects of the present invention are:

[0022] 1. This invention utilizes the powerful nonlinear mapping and feature learning capabilities of deep reinforcement learning to learn and identify the deep coupling relationship between failure modes and multidimensional instability states that are difficult to capture by traditional model-based or simple rule-based methods from the complex and high-dimensional operational state information of distributed-driven flying cars, especially during the land-to-air transition phase when dynamic characteristics change drastically.

[0023] 2. Unlike simple fault alarms or single-index evaluations, this invention outputs failure mode identification, severity level, and quantitative evaluation vectors for stability margins in multiple key degrees of freedom such as longitudinal, lateral, vertical, roll, pitch, and yaw, providing a more comprehensive and detailed profile of the safety status of flying cars.

[0024] 3. This invention focuses on the land-to-air transition phase, which is the riskiest and most dynamically complex during the operation of flying cars. By focusing on data collection and model training during this phase, it can provide targeted real-time safety status assessments and significantly improve the safety assurance capabilities of flying cars during this critical transition process.

[0025] 4. The trained deep reinforcement learning evaluation model has fast forward reasoning ability and can quickly output evaluation results based on real-time vehicle status information, meeting the stringent requirements of real-time safety monitoring for high-speed, dynamic systems like flying cars, and buying valuable time for timely control intervention or decision adjustment. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method in an embodiment of the present invention. Detailed Implementation

[0027] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] like Figure 1 As shown, this invention provides a multi-dimensional evaluation method for the failure and instability of a distributed-drive flying car based on deep reinforcement learning, comprising the following steps:

[0029] S1. Acquire sensor data from the distributed-drive flying car, inject failure mode information and corresponding failure and instability state label data to form an operational state dataset for reinforcement learning.

[0030] S2, construct a failure and instability state evaluation model based on a deep deterministic policy gradient algorithm, the evaluation model is regarded as an intelligent agent; define a multi-dimensional failure and instability state evaluation vector, design a reward function based on the failure and instability state label data, and finally form a deep deterministic policy gradient framework and train it on the running state dataset;

[0031] S3, obtain the real-time operating status of the distributed-drive flying car during the land-air transition phase or the set potential risk phase, and input it into the trained evaluation model;

[0032] S4, the evaluation model outputs a multi-dimensional evaluation vector to assess the vehicle's failure risk, which can be used to trigger alarms, assist in decision-making, and adjust control.

[0033] In one embodiment, the step of acquiring the operational status information of the distributed-drive flying car, pre-injecting failure mode information and corresponding failure and instability state label data, and forming a dataset for reinforcement learning specifically includes:

[0034] The sensor data of the distributed-drive flying car is acquired, including: inertial measurement unit data, positioning data, pitot tube data, radar data, wheel speed, suspension travel, motor current, motor voltage, motor temperature, rotor speed, and rotor pitch angle; these data together constitute part of the reinforcement learning state space.

[0035] Failure mode information simulates one or more combined failure scenarios, such as distributed drive motor output torque mismatch, partial failure, or complete failure, that may occur during the land-to-air transition phase or a set potential risk phase; these scenarios are used to generate labeled state data required for training.

[0036] The failure and instability state label data are calculated using a multiphysics coupled simulation model to simulate the dynamic behavior during the land-air switching process and to provide a baseline truth value for calculating the reward function in step two.

[0037] The sensor data, failure mode information, and failure and instability state label data are combined to form the operating status data.

[0038] In one embodiment, the evaluation model includes a policy network and a value network; the policy network and the value network are trained and updated using an actor-critic framework;

[0039] The training data for the evaluation model should at least cover the land-air switching process of the flying car under different loads, speeds, and environments, including takeoff and landing conditions, so that the agent can learn evaluation strategies applicable to the set key stages.

[0040] In the actual operation or simulation test environment of the flying car, the trained evaluation model is loaded; the current operating state of the flying car, which is collected in real time and processed by the same preprocessing method as in the training stage, is provided as input to the policy network of the evaluation model.

[0041] In one embodiment, the dimensions and quantification method of the multidimensional evaluation vector are determined according to preset safety thresholds and control requirements. The multidimensional evaluation vector represents the remaining safety margin of stability in each dimension and can be used to classify risk levels, such as safety, caution, warning, and danger, to provide a basis for land-to-air handover decisions.

[0042] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0043] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0044] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0045] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A multi-dimensional evaluation method for failure and instability of a distributed drive air car based on deep reinforcement learning, characterized in that, The method comprises the following steps: acquiring sensor data of the distributed drive flying car, injecting failure mode information and corresponding failure and instability state label data, and forming a running state data set for reinforcement learning; building a failure and instability state evaluation model based on a deep deterministic policy gradient algorithm, which is regarded as an intelligent agent; defining a multi-dimensional failure and instability state evaluation vector, designing a reward function based on the failure and instability state label data, and finally forming a deep deterministic policy gradient framework and training the running state data set; acquiring the real-time running state of the distributed drive flying car in the land-air switching stage or the set potential risk stage, and inputting it into the trained evaluation model; the evaluation model outputs a multi-dimensional evaluation vector for evaluating the failure risk of the vehicle, which can be used to trigger alarms, assist decision-making, and adjust control.

2. The multi-dimensional evaluation method for failure and instability degree of the distributed drive flying car based on deep reinforcement learning according to claim 1, characterized in that, The method of acquiring sensor data of the distributed drive flying car, injecting failure mode information and corresponding failure and instability state label data, and forming a running state data set for reinforcement learning comprises the following steps: acquiring sensor data of the distributed drive flying car, the sensor data including inertial measurement unit data, positioning data, pitot tube data, radar data, wheel speed, suspension travel, motor current, motor voltage, motor temperature, rotor speed, and rotor pitch angle; the failure mode information simulates one or more combined failure scenarios of distributed drive motor output torque mismatch, partial failure or complete failure that may occur in the land-air switching stage or the set potential risk stage; the failure and instability state label data are calculated by a multi-physical field coupling simulation model to simulate the dynamic behavior in the land-air switching process; the sensor data, failure mode information, and failure and instability state label data are combined to form the running state data.

3. The multi-dimensional evaluation method for failure and instability of a distributed drive flying car based on deep reinforcement learning according to claim 1, characterized in that, The evaluation model includes a policy network and a value network; the policy network and the value network are trained and updated through an actor-critic framework; The training data of the evaluation model at least covers the land-air switching process of the flying car under different loads, different speeds, and different environments, including take-off and landing conditions, so that the intelligent agent learns an evaluation strategy suitable for the set key stage.

4. The multi-dimensional evaluation method for failure and instability of a distributed drive flying car based on deep reinforcement learning according to claim 1, characterized in that, The dimensions and quantization methods of the multi-dimensional evaluation vector are determined according to the preset safety threshold and control requirements, and the multi-dimensional evaluation vector is the remaining safety margin of each dimension stability, which can be used to divide the risk level.

Images