Control method of triphibian search and rescue device based on multi-sensor fusion and intelligent algorithm

Abstract

The invention discloses a control method of a triphibian search and rescue device based on multi-sensor fusion and an intelligent algorithm. A variety of sensors are deployed on the triphibian search and rescue device, data are collected by using a time synchronization technology, and accurate environmental perception information is obtained through fusion of Kalman filtering and particle filtering algorithms. A deep reinforcement learning intelligent algorithm is used for decision making, a control instruction is generated, and parameters are dynamically adjusted in cooperation with self-adaptive fuzzy control. The system can intelligently plan paths, switch modes, monitor states in real time and process emergently according to environment changes and task requirements. According to the method, the adaptability, reliability and rescue efficiency of the triphibian search and rescue device in a complex environment are remarkably improved, and the multi-sensor fusion system further has good expandability.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicles (UAVs), and in particular relates to a control method for an amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithms. Background Art

[0002] In modern rescue operations, amphibious search and rescue vehicles are playing an increasingly important role. For example, Chinese invention patent publication number CN106976367A discloses an amphibious drone capable of performing amphibious operations with high precision. It is portable and easy to operate, and equipped with various sensor modules, it can achieve multifunctional detection of disaster rescue sites and perform certain rescue missions. However, existing amphibious search and rescue vehicle control methods have many limitations: ① Insufficient sensor fusion: Although existing amphibious drones are equipped with multiple sensors, such as visual and range sensors, the data from these sensors is not fully integrated. Each sensor operates independently, failing to form an integrated whole. This results in an inability to fully and accurately perceive the complex rescue environment. For example, at a smoke-filled fire scene, visual sensors may be interfered with. While range sensors can provide some obstacle information, lacking integration with other sensors makes it impossible to accurately determine the nature of obstacles and the overall surrounding environment, compromising search and rescue efficiency and safety.

[0003] ② Lack of intelligent control methods: Existing technologies lack publicly available effective control methods, and their control strategies are likely to be traditional and fixed, lacking flexibility and adaptability. Control parameters cannot automatically adjust to real-time environmental changes and mission requirements in different environmental modes (flight, land travel, and surface navigation). For example, if a sudden change in current occurs while sailing on the surface, the propulsion force and heading cannot be adjusted in time, resulting in deviation from the planned route and potentially dangerous situations.

[0004] ③ Limited path planning capabilities: When planning paths in complex environments, existing amphibious drones may rely solely on simple preset routes or basic obstacle avoidance algorithms. When encountering dynamic obstacles or complex terrain, they are unable to quickly and intelligently plan the optimal path. For example, in the rubble of an earthquake, with numerous collapsed buildings and falling objects, existing drones may be unable to avoid dangerous areas in a timely manner, resulting in mission failure or equipment damage.

[0005] ④ Poor environmental adaptability: Due to the lack of multi-sensor fusion and intelligent control, the performance of existing amphibious drones is severely affected when faced with extreme environmental conditions such as strong winds, heavy rain, and low temperatures. When flying in strong winds, they may be unable to maintain a stable attitude and may even be blown off course. In low-temperature environments, battery performance degrades, significantly reducing the device's endurance and operating efficiency.

[0006] ⑤ Insufficient decision-making capabilities: Existing amphibious drones lack the ability to make autonomous decisions when performing search and rescue missions. They cannot quickly determine rescue priorities and the best course of action based on the information they collect. For example, when multiple trapped people are discovered, they cannot rationally arrange the rescue sequence based on factors such as their vital signs and location, which can lead to a waste of rescue resources and prolonged rescue time. Summary of the Invention

[0007] In view of the above-mentioned shortcomings of existing amphibious drones, in order to improve the operational capability and rescue efficiency of amphibious search and rescue vehicles in complex environments, the present invention proposes a control method for amphibious search and rescue vehicles based on multi-sensor fusion and intelligent algorithms. This method deeply integrates the data of multiple sensors and uses intelligent algorithms to achieve flexible control, intelligent path planning and autonomous decision-making of the search and rescue vehicle, thereby significantly improving the adaptability and reliability of the amphibious search and rescue vehicle in various complex environments and providing more powerful support for rescue work. The specific technical solutions adopted are: A control method for an amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm includes the following steps: A) Deploy multiple types of sensors on the amphibious search and rescue vehicle and use time synchronization technology to ensure that each sensor collects data under the same time reference; B) Use data fusion algorithms to fuse data collected by different sensors to obtain comprehensive and accurate environmental perception information; C) Based on the fused environmental perception information, intelligent algorithms are used to make decisions and generate control instructions for the amphibious search and rescue vehicle; D) According to the generated control instructions, the operating status of the amphibious search and rescue vehicle in flight, land driving and surface navigation modes is controlled.

[0008] Furthermore, the multiple types of sensors include at least three of laser radar, visual camera, ultrasonic sensor, infrared sensor, multispectral sensor and harmful gas detection sensor.

[0009] Furthermore, the data fusion algorithm uses a Kalman filter algorithm combined with a particle filter algorithm to fuse data from different types of sensors, eliminating noise and errors in the data and improving the accuracy and reliability of environmental perception information. The specific steps include: B1) Initialization: For each sensor, the state vector and covariance matrix are initialized based on its characteristics and historical data. The state vector covers the physical quantities measured by the sensor, such as position and velocity. The covariance matrix reflects the uncertainty of the state estimate. The particle filter particle set is initialized, including the state and weight of each particle. The particle state is also related to the physical quantity measured by the sensor, and the weight reflects the degree of match between the particle and the actual state. B2) Kalman filter prediction: The system model is used to predict the state vector and covariance matrix at the next moment. During the prediction process, the covariance matrix is updated to account for the influence of system noise to reflect the increased uncertainty of the prediction. B3) Sensor data collection: At a specific time point, data is collected from various sensors, which may contain noise and errors; B4) Kalman filter update: Based on the collected sensor data, the Kalman gain is calculated. The Kalman gain is used to balance the weights of the predicted and measured values. It depends on the prediction error covariance matrix and the measurement noise covariance matrix. The Kalman gain is used to update the state vector and covariance matrix to make the state estimate closer to the true value while reducing the uncertainty of the estimate. B5) Particle filter resampling: Based on the updated state estimate from the Kalman filter, the particle weights in the particle filter are adjusted. The weight adjustment depends on the degree of match between the particle state and the Kalman filter estimate state. The resampling operation is performed to eliminate particles with smaller weights and replicate particles with larger weights to ensure that the particle set can better represent the actual state distribution. B6) Data fusion output: This combines the updated state estimate from the Kalman filter with the resampled particle set from the particle filter to obtain fused environmental perception information. This fused information is then smoothed and error corrected to further improve its accuracy and reliability. B7) Iteration loop: Repeat the above steps of Kalman filter prediction, sensor data acquisition, Kalman filter update, particle filter resampling and data fusion output to achieve real-time update and optimization of environmental perception information.

[0010] Furthermore, the intelligent algorithm is a deep reinforcement learning algorithm, which takes the state information of the amphibious search and rescue vehicle as input. The state information includes position, speed, posture and fused sensor data. Through learning and training of the neural network, it outputs the optimal action instructions. The action instructions include flight direction adjustment, speed control, and mode switching. The specific steps include: C1) Initialize the environment and agent: Build a simulation environment for the amphibious rescue vehicle or use an actual search and rescue scenario as a training environment. This environment must accurately simulate various conditions in the three modes of flight, land travel, and surface navigation, including different terrain, weather conditions, and obstacle distribution. Initialize the deep reinforcement learning agent, define the neural network structure, and determine parameters such as the number of layers and neurons. Also, initialize the weights of the agent's policy network and value network. C2) State Information Acquisition and Preprocessing: At each time step, state information is acquired from the amphibious search and rescue vehicle's sensor system, including the current location's latitude and longitude, altitude, linear and angular velocity, attitude's pitch, roll, and yaw angles, as well as fused sensor data such as obstacle distance, ambient temperature, and light intensity. The acquired state information is normalized and preprocessed to be within an appropriate numerical range to improve the training efficiency and stability of the neural network. C3) Action Selection: The agent inputs preprocessed state information into the policy network, which outputs a probability distribution for each possible action based on the current policy. Based on the action probability distribution, an ε-greedy strategy is used to select an action. In the early stages of training, actions are randomly selected with a certain probability ε to fully explore the environment. As training progresses, the ε value is gradually reduced to select more actions with the highest probability. C4) Action Execution and Environmental Feedback: The amphibious rescuer executes the selected action, including adjusting flight direction, changing speed, or switching modes. The environment changes accordingly based on the rescuer's actions and returns new status information and reward values. The reward value should be designed to be relevant to the objectives of the search and rescue mission, with positive rewards for successfully finding the target and negative rewards for colliding with obstacles or failing to complete the mission on time. C5) Experience Storage: The current state, selected action, reward, and next state are stored as an experience sample in the experience replay buffer. The experience replay buffer is used to store the experience accumulated by the agent during training to break the correlation between data and improve the stability of training. C6) Neural Network Training: Randomly sample a batch of experience samples from the experience replay buffer. For the value network, calculate the target value, that is, calculate the target value of the current state-action pair based on the reward value and the estimated value of the next state. Then, use a loss function such as mean squared error to update the weights of the value network through the backpropagation algorithm to bring the estimated value close to the target value. For the policy network, use the policy gradient part of the actor-critic algorithm to update the weights of the policy network so that the actions output by the policy network can obtain higher cumulative rewards. C7) Termination condition determination: Determine whether the termination condition of training has been met, including whether the preset number of training steps has been reached, the cumulative reward has reached a certain threshold, or the convergence standard has been reached. If the termination condition has not been met, return to step C2) to continue training. If the termination condition has been met, terminate the training and obtain a trained deep reinforcement learning model. Furthermore, the method includes an adaptive control step, which uses an adaptive fuzzy control algorithm to dynamically adjust control parameters based on real-time data collected by sensors and environmental changes, fuzzify environmental factors and control targets, establish a fuzzy rule base, and derive corresponding control quantities based on fuzzy reasoning. The specific steps of the adaptive fuzzy control algorithm are as follows: E1) Determine input and output variables: The environmental factors that affect the control of the amphibious search and rescue vehicle and the physical quantities related to the control objectives are used as input variables, and the control parameters that need to be adjusted are used as output variables; E2) Define fuzzy sets and membership functions: Define several fuzzy sets for each input variable and output variable, and determine an appropriate membership function for each fuzzy set to describe the degree to which the input variable or output variable belongs to a fuzzy set; E3) Establish a fuzzy rule base: Based on the control experience and expert knowledge of amphibious search and rescue vehicles, a series of fuzzy rules are developed to ensure that the fuzzy rule base covers all possible input conditions and avoids control blind spots. At the same time, check whether there are any conflicts between the rules. If there are any conflicts, adjust and optimize them. E4) Fuzzy processing: Obtain real-time physical quantity data related to environmental factors and control targets from sensors; substitute the collected input variable data into the corresponding membership function and calculate the membership degree of the input variable to each fuzzy set; E5) Fuzzy reasoning: Based on the membership of the input variables obtained through fuzzification, each rule in the fuzzy rule base is matched and the activation strength of each rule is calculated. The activation strength of a rule is the minimum of the membership of each input variable in the rule's premise. Based on the activation strength of the rule and the conclusion of the rule, the fuzzy set and its membership corresponding to each output variable are determined. E6) Defuzzification: Select the center of gravity method, maximum membership method or weighted average method to convert the fuzzy set of output variables into precise control variables; E7) Parameter update: Apply the precise control quantity obtained by defuzzification to the control system of the amphibious search and rescue vehicle and adjust the corresponding control parameters; E8) Iterative loop: Continuously collect real-time data from sensors and repeat the above steps of fuzzification, fuzzy inference, defuzzification, and control parameter adjustment. Dynamically adjust the control parameters in real time according to environmental changes and the operating status of the amphibious search and rescue vehicle to achieve adaptive control.

[0011] Furthermore, the environmental factors include wind speed, water flow speed, temperature, humidity, and light intensity; the control objectives include maintaining the stable operation of the amphibious search and rescue vehicle and improving the accuracy and reliability of the operation; and the control quantities include the output power of the motor, the speed of the propeller, and the steering angle of the servo.

[0012] Furthermore, during the mode switching process of the amphibious search and rescue vehicle, it is judged whether the mode switching conditions are met based on the fused environmental perception information and the decision results of the intelligent algorithm. If so, the amphibious search and rescue vehicle is controlled to switch between flight, land driving, and surface navigation modes. At the same time, the control parameters are adjusted according to the characteristics of different modes during the switching process.

[0013] Furthermore, the method also includes a path planning step, which uses intelligent algorithms and fused environmental perception information to plan the travel path of the amphibious search and rescue vehicle. During the planning process, factors such as obstacle distribution, target location, energy consumption, etc. are considered to generate the optimal path.

[0014] Furthermore, the method also includes real-time monitoring and evaluation of the operating status of the amphibious search and rescue vehicle. When an abnormal situation is detected in the amphibious search and rescue vehicle, an intelligent algorithm is used to generate and execute a corresponding emergency response strategy based on the type and severity of the abnormal situation.

[0015] Furthermore, the multi-sensor fusion system is scalable and can easily add new types of sensors to adapt to different rescue scenario requirements without changing the core algorithm architecture, and the data of the newly added sensors can be seamlessly connected to the data fusion algorithm for processing.

[0016] The control method of the amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm of the present invention has the following beneficial technical effects: 1. Comprehensive and accurate environmental perception: By deploying multiple types of sensors on the amphibious search and rescue vehicle and utilizing time synchronization technology and advanced data fusion algorithms (Kalman filter algorithm combined with particle filter algorithm), deep fusion of sensor data is achieved, effectively eliminating noise and errors. This allows for comprehensive and accurate perception of complex rescue environments, making up for the shortcomings of insufficient sensor fusion in existing technologies and improving the efficiency and safety of search and rescue.

[0017] 2. Intelligent and flexible control strategy: Using a deep reinforcement learning algorithm, the system takes the status information of the amphibious search and rescue vehicle as input, and outputs the optimal action instructions through neural network learning and training. It can automatically adjust the control parameters according to real-time environmental changes and mission requirements, so that the amphibious search and rescue vehicle can respond flexibly in flight, land driving, and surface navigation modes, solving the problem that existing control methods are not intelligent enough, lack flexibility and adaptability.

[0018] 3. Efficient path planning capability: Path planning is performed using intelligent algorithms and integrated environmental perception information, fully considering factors such as obstacle distribution, target location, and energy consumption. It can quickly and intelligently plan the optimal path, enabling the amphibious search and rescue vehicle to avoid dangerous areas in complex environments such as earthquake ruins, thereby improving the mission success rate and overcoming the limited existing path planning capabilities.

[0019] 4. Good environmental adaptability: Combined with the adaptive fuzzy control algorithm, the control parameters (such as motor output power, propeller speed, steering angle of the servo, etc.) are dynamically adjusted according to the real-time data collected by the sensor and environmental changes (such as wind speed, water flow speed, temperature, humidity, light intensity, etc.), ensuring that the amphibious search and rescue vehicle can still maintain a stable posture and good endurance and work efficiency under extreme environmental conditions (strong wind, heavy rain, low temperature, etc.), improving the poor environmental adaptability of existing amphibious drones.

[0020] 5. Improved autonomous decision-making capabilities: Deep reinforcement learning algorithms enable amphibious search and rescue vehicles to quickly determine the priority of rescue and the best course of action based on the collected information (such as the vital signs and location of trapped people), reasonably arrange the rescue sequence, avoid wasting rescue resources and extending rescue time, and solve the problem of insufficient decision-making capabilities of existing amphibious drones.

[0021] 6. Intelligent and stable mode switching: During the mode switching process of the amphibious search and rescue vehicle, the switching conditions are judged based on the fused environmental perception information and the decision results of the intelligent algorithm, and the control parameters are adjusted according to the characteristics of different modes, ensuring the rationality and stability of the mode switching.

[0022] 7. Real-time monitoring and emergency response: The operating status of the amphibious search and rescue vehicle is monitored and evaluated in real time. When an abnormal situation occurs, an intelligent algorithm can be used to generate and execute emergency response strategies based on the type and severity of the abnormality to ensure the safe operation of the equipment.

[0023] 8. Strong system scalability: The multi-sensor fusion system is scalable and can easily add new types of sensors without changing the core algorithm architecture. The new sensor data can be seamlessly integrated into the data fusion algorithm processing to adapt to the needs of different rescue scenarios, thereby improving the versatility and flexibility of the system. DETAILED DESCRIPTION

[0024] In order to explain the technical solution of the present invention more clearly and in detail, the control method of the amphibious search and rescue device based on multi-sensor fusion and intelligent algorithm of the present invention is described in detail below in combination with specific embodiments.

[0025] Sensor deployment A variety of sensor types should be deployed on amphibious search and rescue vehicles. Depending on the specific rescue scenario and needs, at least three of the following should be selected: lidar, visual cameras, ultrasonic sensors, infrared sensors, multispectral sensors, and hazardous gas detection sensors. For example, in a fire rescue scenario, lidar can be used to measure the distance to obstacles, visual cameras can capture images of the surrounding environment, and hazardous gas detection sensors can monitor the concentration of hazardous gases generated by the fire. A high-precision time synchronization module should be used to ensure that all sensors collect data based on the same time reference, avoiding data fusion errors caused by time asynchrony.

[0026] Data fusion algorithm implementation 1. Initialization phase: For the lidar, the state vector is initialized, such as position, velocity and other information, based on its ranging accuracy and scanning frequency, as well as past measurement data in similar environments. At the same time, the covariance matrix is initialized. This matrix reflects the uncertainty of the state estimation, and the value is obtained based on the historical measurement error statistics of the lidar. For the visual camera, the relevant state vectors are initialized based on its resolution, field of view and other characteristics, such as shooting direction, image feature position, etc., and the covariance matrix is determined. For the particle set of the initialized particle filter, the state of each particle is set to a value related to the physical quantity measured by the sensor. For example, the position state of the particle can be randomly distributed according to the measurement range of the lidar, and the weights are initialized to equal values, indicating that the degree of match between each particle and the true state is the same at the beginning.

[0027] Kalman filter prediction: Utilizing the amphibious search and rescue vehicle's motion model, such as its dynamics model in flight mode, combined with system noise parameters, the system predicts the state vector and covariance matrix of sensors such as lidar and visual cameras at the next moment. For example, based on the current flight speed, acceleration, and the statistical characteristics of system noise, the system predicts the likely position and velocity values measured by the lidar at the next moment, and updates the covariance matrix to reflect the increased uncertainty in the prediction process.

[0028] Sensor data acquisition: At specific time intervals, such as every 0.1 second, distance data is acquired from the LiDAR, image data from the camera, and near-field obstacle information from the ultrasonic sensor. Due to various interferences in the real environment, these data may contain noise and errors. For example, LiDAR data may produce measurement deviations due to varying surface characteristics of reflective objects, and camera images may be affected by lighting variations and produce noise.

[0029] Kalman filter update: Based on the collected sensor data, the Kalman gain is calculated. For example, for lidar data, the Kalman gain is calculated based on the prediction error covariance matrix and the lidar measurement noise covariance matrix. This gain is used to balance the weights of the predicted and measured values. The Kalman gain is used to update the state vector and covariance matrix of sensors such as lidar and visual cameras, bringing the state estimate closer to the true value while reducing the uncertainty of the estimate. For example, the lidar's position state estimate is updated based on the actual distance measured by the lidar and the Kalman gain.

[0030] Particle filter resampling: Based on the Kalman filter's updated state estimate, the particle weights in the particle filter are adjusted. For example, if a particle's state is closer to the Kalman filter's estimated lidar position, its weight is increased; otherwise, its weight is decreased. Resampling eliminates particles with smaller weights and replicates particles with larger weights to ensure that the particle set better represents the true state distribution. For example, the 10% of particles with the smallest weights are eliminated, and the 10% of particles with the largest weights are replicated several times to maintain the total number of particles.

[0031] Data fusion output: The updated state estimate from the Kalman filter and the resampled particle set from the particle filter are combined to generate fused environmental perception information. This fused information is smoothed, for example, using a moving average filter to remove glitches and outliers. Error correction is also performed, for example, by comparing the fused obstacle position information with a pre-established environmental model to further improve its accuracy and reliability.

[0032] Iterative loops repeat the aforementioned steps of Kalman filter prediction, sensor data acquisition, Kalman filter update, particle filter resampling, and data fusion output to achieve real-time updating and optimization of environmental perception information. For example, during the entire rescue mission, a complete iteration is performed every second to ensure that the amphibious search and rescue vehicle always obtains the latest and most accurate environmental information.

[0033] Intelligent algorithm application 1. Initializing the Environment and Agent: A high-precision simulation environment for the amphibious rescuer was constructed, capable of accurately simulating various scenarios in its three modes: flight, land travel, and surface navigation. In the flight simulation, different meteorological conditions were considered, such as the impact of strong winds on flight attitude, and the complex airflow changes in different terrains, such as mountainous flight. In the land travel simulation, the effects of different road conditions, such as rugged mountain roads and muddy roads, on driving stability were considered. In the surface navigation simulation, the interference of varying water velocities and wave heights on the navigation trajectory was simulated. The deep reinforcement learning agent was initialized, and the neural network was configured to adopt a multilayer perceptron structure. The number of neurons in the input layer was determined based on the dimensionality of the amphibious rescuer's state information. For example, if position, velocity, attitude, and fused sensor data have N dimensions, then the number of neurons in the input layer would be N. Three hidden layers were set, with 128, 64, and 32 neurons in each layer, respectively. The number of neurons in the output layer was determined based on the number of action commands. For example, if there are M commands, such as flight direction adjustment, speed control, and mode switching, then the number of neurons in the output layer would be M. At the same time, the weights of the agent's strategy network and value network are initialized, and a random initialization method is used to make the initial weights randomly distributed within a certain range.

[0034] State Information Acquisition and Preprocessing: At each time step, state information is acquired from the amphibious rescue vehicle's sensor system. For example, the GPS module obtains the current location's latitude and longitude, the barometric altimeter obtains altitude, the inertial measurement unit obtains linear velocity and angular velocity, and the attitude sensor obtains attitude information such as pitch, roll, and yaw angles. This information is also fused with sensor data, such as obstacle distances measured by lidar, ambient temperature measured by infrared sensors, and light intensity obtained by visual cameras. The acquired state information is preprocessed by normalization, for example, normalizing position information to the [0, 1] range and normalizing velocity information based on the amphibious rescue vehicle's maximum speed to an appropriate numerical range to improve the training efficiency and stability of the neural network.

[0035] Action Selection: The agent inputs preprocessed state information into the policy network, which outputs a probability distribution for each possible action based on the current policy. For example, for a flight direction adjustment action, the policy network outputs the probabilities of turning left, turning right, and maintaining the current direction. Based on this action probability distribution, an ε-greedy strategy is used to select an action. Initially, ε is set to 0.8, resulting in an 80% probability of randomly selecting an action to fully explore the environment. As training progresses, ε is gradually reduced by 0.01 every 1000 steps, gradually decreasing the ε value to more frequently select the action with the highest probability.

[0036] Action Execution and Environmental Feedback: The amphibious rescuer executes its selected action. For example, in flight mode, it adjusts its flight direction and speed by adjusting motor speed according to action commands. In land driving mode, it controls wheel steering and power output according to action commands. In surface navigation mode, it adjusts propeller speed and servo steering angle according to action commands. The environment changes accordingly based on the rescuer's actions and returns new status information and reward values. For example, successfully avoiding an obstacle gives a reward of +5; colliding with an obstacle gives a reward of -10; successfully finding the target gives a reward of +20; and failing to complete the mission on time gives a reward of -15.

[0037] Experience Storage: The current state, selected action, reward, and next state are stored as an experience sample in the experience replay buffer. The experience replay buffer uses a first-in-first-out queue structure with a buffer size of 10,000 samples. When the buffer is full, the oldest sample is automatically deleted to ensure that the stored experience samples are always up-to-date and valid, breaking the correlation between data and improving training stability.

[0038] Neural network training: Randomly sample a batch of experience samples from the experience replay buffer, such as 64 samples each time. For the value network, calculate the target value based on the reward value and the estimated value of the next state, using the formula Calculate the target value of the current state-action pair, where r is the reward value, is the discount factor, set to 0.95, is the estimated value of the next state. Then, using the mean square error loss function, the weights of the value network are updated through the back propagation algorithm to make the estimated value close to the target value. For the policy network, the policy gradient part of the Actor-Critic algorithm is used to update the weights of the policy network so that the actions output by the policy network can obtain higher cumulative rewards. For example, according to the policy gradient formula Calculate the gradient of the policy network, where are the parameters of the policy network, is the probability of taking action a in state s, A(s,a) is the advantage function, and the policy network weights are updated through continuous iteration.

[0039] 7. Termination condition judgment: judge whether the termination condition of training is met, set the preset number of training steps to 100,000 steps, or the cumulative reward reaches a threshold of 5,000, or when the parameter changes of the strategy network and the value network are less than a certain minimum value (such as 10 -5 If the termination condition is not met, the training is continued in step 2. If the termination condition is met, the training is terminated and a trained deep reinforcement learning model is obtained.

[0040] Adaptive control implementation 1. Determine input and output variables: Consider environmental factors such as wind speed, water flow, temperature, humidity, and light intensity, as well as physical quantities related to control objectives such as maintaining stable operation of the amphibious rescuer and improving operational accuracy and reliability. For example, consider measured wind speed and water flow as inputs. Consider control parameters that require adjustment, such as motor output power, propeller speed, and steering angle, as output variables.

[0041] Define fuzzy sets and membership functions: For the wind speed input variable, define fuzzy sets such as "low wind speed," "medium wind speed," and "high wind speed," and determine appropriate membership functions for each fuzzy set. For example, the "low wind speed" fuzzy set uses a Gaussian membership function with a center value of 5 m / s and a standard deviation of 1 m / s, indicating that wind speeds within the 4–6 m / s range are more likely to belong to the "low wind speed" fuzzy set. For the motor output power output variable, define fuzzy sets such as "low power," "medium power," and "high power," and determine corresponding membership functions. For example, the "low power" fuzzy set uses a triangular membership function with a value range determined by the motor's rated power and actual operating requirements.

[0042] Establish a fuzzy rule base: Based on control experience and expert knowledge of amphibious rescue vehicles, a series of fuzzy rules are developed. For example, when the wind speed is "high" and the amphibious rescue vehicle is in flight mode, a rule is established to increase motor output power to maintain stable operation. In other words, the fuzzy rule is "If the wind speed is high and the mode is flight mode, then the motor output power is high." Ensure that the fuzzy rule base covers all possible input conditions. Establish a comprehensive rule base by analyzing different combinations of environmental factors and different operating modes. At the same time, carefully check for conflicts between rules. If so, optimize by adjusting the membership functions or rule conclusions.

[0043] Fuzzy processing: Real-time environmental factor data is obtained from wind speed sensors, water flow velocity sensors, etc., and physical quantity data related to the control target is obtained from the status monitoring system of the amphibious search and rescue vehicle. The collected wind speed input variable data is substituted into the membership function corresponding to the fuzzy sets such as "low wind speed", "medium wind speed", and "high wind speed", and the membership degree of the wind speed to each fuzzy set is calculated. For example, if the measured wind speed is 7m / s, the membership degree calculated by substituting it into the "high wind speed" membership function is 0.8, and the membership degree calculated by substituting it into the "medium wind speed" membership function is 0.2.

[0044] Fuzzy reasoning: Each rule in the fuzzy rule base is matched based on the membership of the input variables obtained through fuzzification. For example, in the case of a wind speed of 7 m / s, the rule "If the wind speed is high and the mode is flight mode, then the motor output power is high power" is matched. Since the wind speed belongs to the "high wind speed" fuzzy set with a membership of 0.8, if the amphibious rescuer is in flight mode at this time, the activation strength of this rule is 0.8. Based on the activation strength of the rule and the conclusion of the rule, the fuzzy set corresponding to the motor output power and its membership are determined. For example, the conclusion of the rule is that the motor output power is "high power", and its membership is 0.8.

[0045] Defuzzification: Select the centroid method to convert the fuzzy set of motor output power into an accurate control quantity. For example, according to the membership distribution of the fuzzy set of motor output power "low power", "medium power" and "high power", the centroid method is used to calculate the formula Calculate the accurate motor output power value, where is a discrete value in the fuzzy set, is the corresponding membership degree.

[0046] Parameter Update: The precise control variables obtained through defuzzification are applied to the amphibious rescue vehicle's control system to adjust the motor output power. For example, if the calculated motor output power is 80% of the rated power, the control system adjusts the motor output power to that value.

[0047] Iterative loop: Continuously collect real-time data from sensors and repeat the aforementioned steps of fuzzification, fuzzy inference, defuzzification, and control parameter adjustment. For example, one iteration per second is performed to dynamically adjust control parameters in real time based on environmental changes and the operating status of the amphibious rescuer, achieving adaptive control.

[0048] Mode switch control During the operation of the amphibious rescuer, the conditions for mode switching are determined based on the fused environmental perception information and the decision-making results of the intelligent algorithm. For example, when the visual camera and lidar detect a large area of water ahead and the water depth exceeds a certain threshold, and the intelligent algorithm determines that a surface search is required based on the mission requirements, the conditions for switching from flight mode to surface navigation mode are met. If the conditions are met, the amphibious rescuer is controlled to switch between flight, land driving, and surface navigation modes. During the switching process, the control parameters are adjusted according to the characteristics of different modes. For example, when switching from flight mode to surface navigation mode, the control method of the motor is converted from controlling the flight attitude to controlling the propeller thrust, and the propeller speed and the steering angle control parameters of the servo are adjusted to meet the needs of surface navigation.

[0049] Path planning implementation The amphibious rescuer's path is planned using intelligent algorithms and fused environmental perception information. During the planning process, the obstacle distribution is fully considered, and the obstacle location and shape information is obtained by fusing data from lidar and visual cameras; the target location is considered based on the target location information set in the rescue mission; and energy consumption is considered based on factors such as the battery power and fuel consumption model of the amphibious rescuer. For example, the A* algorithm is combined with the Dijkstra algorithm, with the target location as the end point and the current location of the amphibious rescuer as the starting point, to search for the optimal path while taking obstacles and energy consumption into consideration. When encountering dynamic obstacles, such as moving vehicles or personnel, the environmental perception information is updated in real time and the path is replanned to ensure that the amphibious rescuer can reach the target location safely and efficiently.

[0050] Operation status monitoring and emergency response The operating status of the amphibious rescuer is monitored in real time, with sensors monitoring parameters such as the motor's operating temperature, battery charge, and the device's attitude stability. An evaluation model is established to assess whether the amphibious rescuer's operating status is normal based on the monitoring data. When an abnormality is detected in the amphibious rescuer, such as an excessively high motor operating temperature exceeding a safety threshold, an intelligent algorithm is used to generate and execute a corresponding emergency response strategy based on the type and severity of the abnormality. For example, if the motor temperature is too high, the intelligent algorithm generates an emergency response strategy to reduce the motor's output power and increase the cooling fan speed to reduce the motor temperature and ensure safe operation of the equipment.

[0051] System scalability implementation The multi-sensor fusion system is scalable. When a new type of sensor needs to be added, such as a biosensor used to detect biological signals to locate trapped people, a standard data interface is reserved in the hardware design to facilitate the integration of the new sensor. In terms of software algorithms, the core algorithm architecture does not change. It is only necessary to initialize the state vector and covariance matrix for the new sensor according to the characteristics of the new sensor during the initialization phase of the data fusion algorithm, as well as initialize the relevant parameters of the particle set in the particle filter. The data from the newly added sensor can be seamlessly integrated into the data fusion algorithm for processing. By adapting the data fusion algorithm process, the new sensor data is combined with the original sensor data for Kalman filter prediction, update, and particle filter resampling operations to meet the needs of different rescue scenarios.

[0052] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made by those skilled in the art without departing from the spirit and principles of the present invention shall be included within the scope of protection of the present invention. The scope of protection of the present invention shall be based on the scope defined by the claims.

Claims

1. A control method for an amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm, characterized in that: The following steps are involved: A) Deploy multiple types of sensors on the amphibious search and rescue vehicle and use time synchronization technology to ensure that each sensor collects data under the same time reference; B) Use data fusion algorithms to fuse data collected by different sensors to obtain comprehensive and accurate environmental perception information; C) Based on the fused environmental perception information, intelligent algorithms are used to make decisions and generate control instructions for the amphibious search and rescue vehicle; D) According to the generated control instructions, the operating status of the amphibious search and rescue vehicle in flight, land driving and surface navigation modes is controlled.

2. The control method of the amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm according to claim 1 is characterized in that: The multiple types of sensors include at least three of laser radar, visual camera, ultrasonic sensor, infrared sensor, multi-spectral sensor and harmful gas detection sensor.

3. The control method of the amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm according to claim 1 is characterized in that: The data fusion algorithm uses the Kalman filter algorithm combined with the particle filter algorithm to fuse data from different types of sensors, eliminate noise and errors in the data, and improve the accuracy and reliability of environmental perception information. The specific steps include: B1) Initialization: For each sensor, the state vector and covariance matrix are initialized based on its characteristics and historical data. The state vector covers the physical quantities measured by the sensor, such as position and velocity. The covariance matrix reflects the uncertainty of the state estimate. The particle filter particle set is initialized, including the state and weight of each particle. The particle state is also related to the physical quantity measured by the sensor, and the weight reflects the degree of match between the particle and the actual state. B2) Kalman filter prediction: The system model is used to predict the state vector and covariance matrix at the next moment. During the prediction process, the covariance matrix is updated to account for the influence of system noise to reflect the increased uncertainty of the prediction. B3) Sensor data collection: At a specific time point, data is collected from various sensors, which may contain noise and errors; B4) Kalman filter update: Based on the collected sensor data, the Kalman gain is calculated. The Kalman gain is used to balance the weights of the predicted and measured values. It depends on the prediction error covariance matrix and the measurement noise covariance matrix. The Kalman gain is used to update the state vector and covariance matrix to make the state estimate closer to the true value while reducing the uncertainty of the estimate. B5) Particle filter resampling: Based on the updated state estimate from the Kalman filter, the particle weights in the particle filter are adjusted. The weight adjustment depends on the degree of match between the particle state and the Kalman filter estimate state. The resampling operation is performed to eliminate particles with smaller weights and replicate particles with larger weights to ensure that the particle set can better represent the actual state distribution. B6) Data fusion output: This combines the updated state estimate from the Kalman filter with the resampled particle set from the particle filter to obtain fused environmental perception information. This fused information is then smoothed and error corrected to further improve its accuracy and reliability. B7) Iteration loop: Repeat the above steps of Kalman filter prediction, sensor data acquisition, Kalman filter update, particle filter resampling and data fusion output to achieve real-time update and optimization of environmental perception information.

4. The control method of the amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm according to claim 1 is characterized in that: The intelligent algorithm is a deep reinforcement learning algorithm that takes the state information of the amphibious search and rescue vehicle as input. The state information includes position, speed, posture and fused sensor data. Through learning and training of the neural network, it outputs the optimal action instructions. The action instructions include flight direction adjustment, speed control, and mode switching. The specific steps include: C1) Initialize the environment and agent: Build a simulation environment for the amphibious rescue vehicle or use an actual search and rescue scenario as a training environment. This environment must accurately simulate various conditions in the three modes of flight, land travel, and surface navigation, including different terrain, weather conditions, and obstacle distribution. Initialize the deep reinforcement learning agent, define the neural network structure, and determine parameters such as the number of layers and neurons. Also, initialize the weights of the agent's policy network and value network. C2) State Information Acquisition and Preprocessing: At each time step, state information is acquired from the amphibious search and rescue vehicle's sensor system, including the current location's latitude and longitude, altitude, linear and angular velocity, attitude's pitch, roll, and yaw angles, as well as fused sensor data such as obstacle distance, ambient temperature, and light intensity. The acquired state information is normalized and preprocessed to be within an appropriate numerical range to improve the training efficiency and stability of the neural network. C3) Action Selection: The agent inputs preprocessed state information into the policy network, which outputs a probability distribution for each possible action based on the current policy. Based on the action probability distribution, an ε-greedy strategy is used to select an action. In the early stages of training, actions are randomly selected with a certain probability ε to fully explore the environment. As training progresses, the ε value is gradually reduced to select more actions with the highest probability. C4) Action Execution and Environmental Feedback: The amphibious rescuer executes the selected action, including adjusting flight direction, changing speed, or switching modes. The environment changes accordingly based on the rescuer's actions and returns new status information and reward values. The reward value should be designed to be relevant to the objectives of the search and rescue mission, with positive rewards for successfully finding the target and negative rewards for colliding with obstacles or failing to complete the mission on time. C5) Experience Storage: The current state, selected action, reward, and next state are stored as an experience sample in the experience replay buffer. The experience replay buffer is used to store the experience accumulated by the agent during training to break the correlation between data and improve the stability of training. C6) Neural Network Training: Randomly sample a batch of experience samples from the experience replay buffer. For the value network, calculate the target value, that is, calculate the target value of the current state-action pair based on the reward value and the estimated value of the next state. Then, use a loss function such as mean squared error to update the weights of the value network through the backpropagation algorithm to bring the estimated value close to the target value. For the policy network, use the policy gradient part of the actor-critic algorithm to update the weights of the policy network so that the actions output by the policy network can obtain higher cumulative rewards. C7) Termination condition determination: Determine whether the termination conditions for training have been met, including whether the preset number of training steps has been reached, the cumulative reward has reached a certain threshold, or the convergence standard has been reached. If the termination conditions have not been met, return to step C2) to continue training. If the termination conditions have been met, the training ends and a trained deep reinforcement learning model is obtained.

5. The control method of the amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm according to claim 1 is characterized in that: The system also includes an adaptive control step, which uses an adaptive fuzzy control algorithm to dynamically adjust control parameters based on real-time data collected by sensors and environmental changes, fuzzify environmental factors and control targets, establish a fuzzy rule base, and derive corresponding control quantities based on fuzzy reasoning. The specific steps of the adaptive fuzzy control algorithm are as follows: E1) Determine input and output variables: The environmental factors that affect the control of the amphibious search and rescue vehicle and the physical quantities related to the control objectives are used as input variables, and the control parameters that need to be adjusted are used as output variables; E2) Define fuzzy sets and membership functions: Define several fuzzy sets for each input variable and output variable, and determine an appropriate membership function for each fuzzy set to describe the degree to which the input variable or output variable belongs to a fuzzy set; E3) Establish a fuzzy rule base: Based on the control experience and expert knowledge of amphibious search and rescue vehicles, a series of fuzzy rules are developed to ensure that the fuzzy rule base covers all possible input conditions and avoids control blind spots. At the same time, check whether there are any conflicts between the rules. If there are any conflicts, adjust and optimize them. E4) Fuzzy processing: Obtain real-time physical quantity data related to environmental factors and control targets from sensors; substitute the collected input variable data into the corresponding membership function and calculate the membership degree of the input variable to each fuzzy set; E5) Fuzzy reasoning: Based on the membership of the input variables obtained through fuzzification, each rule in the fuzzy rule base is matched and the activation strength of each rule is calculated. The activation strength of a rule is the minimum of the membership of each input variable in the rule's premise. Based on the activation strength of the rule and the conclusion of the rule, the fuzzy set and its membership corresponding to each output variable are determined. E6) Defuzzification: Select the center of gravity method, maximum membership method or weighted average method to convert the fuzzy set of output variables into precise control variables; E7) Parameter update: Apply the precise control quantity obtained by defuzzification to the control system of the amphibious search and rescue vehicle and adjust the corresponding control parameters; E8) Iterative loop: Continuously collect real-time data from sensors and repeat the above steps of fuzzification, fuzzy inference, defuzzification, and control parameter adjustment. Dynamically adjust the control parameters in real time according to environmental changes and the operating status of the amphibious search and rescue vehicle to achieve adaptive control.

6. The control method of the amphibious search and rescue device based on multi-sensor fusion and intelligent algorithm according to claim 5 is characterized in that: The environmental factors include wind speed, water flow speed, temperature, humidity, and light intensity. The control objectives include maintaining the stable operation of the amphibious search and rescue vehicle and improving the accuracy and reliability of the operation. The control quantities include the output power of the motor, the speed of the propeller, and the steering angle of the servo.

7. The control method of the amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm according to claim 1 is characterized in that: During the mode switching process of the amphibious search and rescue vehicle, it is judged whether the mode switching conditions are met based on the fused environmental perception information and the decision results of the intelligent algorithm. If so, the amphibious search and rescue vehicle is controlled to switch between flight, land driving, and surface navigation modes. At the same time, the control parameters are adjusted according to the characteristics of different modes during the switching process.

8. The control method of the amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm according to claim 1 is characterized in that: The method also includes a path planning step, which uses intelligent algorithms and fused environmental perception information to plan the travel path of the amphibious search and rescue vehicle. During the planning process, factors such as obstacle distribution, target location, and energy consumption are considered to generate the optimal path.

9. The control method of the amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm according to claim 1, characterized in that: The method also includes real-time monitoring and evaluation of the operating status of the amphibious search and rescue vehicle. When an abnormal situation is detected in the amphibious search and rescue vehicle, an intelligent algorithm is used to generate and execute a corresponding emergency response strategy based on the type and severity of the abnormal situation.

10. The control method of the amphibious search and rescue vehicle based on multi-sensor fusion and intelligent algorithm according to claim 1, characterized in that: The multi-sensor fusion system is scalable and can easily add new types of sensors without changing the core algorithm architecture to adapt to different rescue scenario requirements, and the data of the newly added sensors can be seamlessly connected to the data fusion algorithm for processing.