Unknown environment adaptive navigation control system based on deep learning

Abstract

The invention provides an unknown environment adaptive navigation control system based on deep learning, and aims to improve the autonomous navigation capability of a machine in a complex and unknown environment. According to the system, an advanced sensing module is adopted, multi-sensor data are combined, and a dynamic environment is quickly and accurately sensed through a deep learning algorithm. The online environment modeling mechanism can update the environment information in real time without depending on a pre-constructed map. And by utilizing a reinforcement learning algorithm, the system can autonomously learn and optimize path planning, and meanwhile, energy consumption, time efficiency and safety are considered. An innovative multi-mode obstacle avoidance strategy is adopted, and obstacle avoidance behaviors are dynamically adjusted according to environment complexity and machine states. The intelligent decision support system combines a decision tree and fuzzy logic, and provides fast and accurate decision ability. The adaptive motion control algorithm optimizes path tracking performance and adapts to dynamic characteristics and environmental conditions of a machine. In addition, the collaborative job mechanism supports efficient communication and task collaboration in a multi-machine system.

Description

Technical field:

[0001] The present invention relates to the field of intelligent machine technology, and in particular to an unknown environment adaptive navigation control system based on deep learning. Background technology:

[0002] In today's wave of technological development, intelligent machines are playing an increasingly important role in industrial automation, home services, medical assistance, disaster relief, and space exploration. However, the navigation systems of these machines are mostly based on preset environmental information and static obstacle avoidance logic, which limits their adaptability and flexibility in unknown or rapidly changing environments. For example, in complex indoor environments or after natural disasters, machines need to be able to autonomously identify obstacles, plan safe paths, and adjust their action strategies in real time.

[0003] Existing navigation control systems typically rely on sensor arrays to collect environmental data and process it through specific algorithms. These systems may include laser-based SLAM technology to build environmental maps and rule-based obstacle avoidance algorithms to guide the movement of the machine. However, these methods are often unable to handle large-scale, highly dynamic environments, and have limitations in the versatility and robustness of the algorithms. In addition, when the environment changes or new obstacles appear, these systems often require human intervention or long recalculations, which not only affects the response speed of the machine, but also reduces its reliability in practical applications.

[0004] To overcome these challenges, researchers have been exploring more advanced technologies, such as using machine learning and artificial intelligence to improve the autonomous decision-making capabilities of machines. As a powerful tool, deep learning has achieved remarkable results in areas such as image recognition and natural language processing. Applying deep learning to machine navigation and control systems is expected to achieve more accurate environmental perception, more intelligent decision-making, and more flexible path planning.

[0005] Although there have been some preliminary attempts, it is still a technical challenge to effectively combine deep learning technology with machine navigation control systems, especially in unknown and dynamically changing environments. This involves the design of algorithms, the improvement of computing efficiency, the guarantee of real-time performance, and the stability and security of the system. It is in this context that the present invention is proposed, aiming to solve the deficiencies of the existing technology through innovative deep learning algorithms and system architecture, promote the development of intelligent machine navigation control systems, and enable them to play a greater role in a wider range of application scenarios.

[0006] Technical solution:

[0007] The technical solution of the present invention is an adaptive navigation control system for unknown environments based on deep learning, which aims to provide an efficient and intelligent solution to achieve autonomous navigation and obstacle avoidance of machines in complex and unknown environments:

[0008] Preferably, the perception module is an integrated system consisting of a series of high-precision sensors. These sensors include high-resolution cameras for capturing visual information; LiDAR for 360-degree environmental scanning and distance measurement; ultrasonic sensors for close-range obstacle detection; and infrared sensors for navigation in low-light environments. The combination of these sensors ensures that the machine can perceive the surrounding environment from different angles and distances; each sensor collects data at a specific frequency, forming a large raw data set. The perception module needs to preprocess this data, including denoising, filtering, and standardization, to improve the accuracy and efficiency of subsequent processing. In addition, data preprocessing also includes time synchronization to ensure that data from different sensors can be fused on the same time basis; the preprocessed data is then input into the deep learning algorithm. These algorithms are specially designed to extract key features such as the shape, size, position, and motion state of obstacles from multi-source data. Deep convolutional neural networks (CNNs) can be used for feature extraction of image data, while recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) are suitable for processing time series data, such as the output of LIDAR and ultrasonic sensors; the output of the deep learning algorithm is a series of environmental features, which are then integrated in the feature fusion layer to form a unified feature representation. This representation not only contains the physical properties of the environment, but may also contain semantic information, such as pedestrians, vehicles, and traffic signals. In this way, the perception module can provide the machine with a rich, multi-dimensional understanding of the environment, providing a solid foundation for subsequent navigation decisions; the perception module needs to respond to environmental changes in real time, so the computational efficiency of the algorithm is crucial. By optimizing the network structure, using efficient computing libraries and parallel processing technology, it is ensured that the perception module can meet the real-time requirements while maintaining high accuracy.

[0009] Preferably, the environment modeling module first receives data from the perception module, which includes environmental features captured by various sensors. Before the data is used for modeling, it needs to be preprocessed to ensure the consistency and accuracy of the data. The preprocessing steps may include data cleaning, denoising, standardization, and feature alignment to eliminate the deviation between different sensor data; using the data provided by the perception module, the environment modeling module uses advanced 3D modeling technology to build a geometric representation of the environment. This may involve generating a grid model using point cloud data, or reconstructing a 3D scene through image data and depth information. The constructed 3D model includes not only the location and shape of obstacles, but may also contain the ground, walls and other environmental features; in addition to the 3D model, the environment modeling module also generates an occupancy grid (OccupancyGrid), which is a grid-based map representation method. The occupancy grid divides the environment into multiple small units, each of which is marked as occupied or idle according to the perception data. This representation method facilitates the machine to perform path planning and obstacle avoidance; the core of the environment modeling module is a dynamic map update algorithm that can respond to changes in the environment in real time. When the perception module detects new obstacles or changes in the environment layout, the dynamic map update algorithm adjusts the occupancy grid and 3D model to reflect these changes. This ensures that the machine's representation of the environment is always up to date; the environment modeling module is also responsible for fusing data from different sensors to improve the accuracy and robustness of the model. For example, LIDAR data can provide accurate distance measurements, while camera data can provide rich texture and color information. By fusing this data, the environment model is able to more comprehensively reflect the characteristics of the environment; the environment modeling module uses efficient data structures and algorithms. The module design takes into account the limitations of computing resources, and ensures fast response and efficient data processing by optimizing algorithms and utilizing parallel computing technology.

[0010] Preferably, the particle filter represents the position probability distribution of the machine by simulating a large number of random samples (particles), while the Kalman filter updates the optimal solution of the position estimate in an analytical way; in order to improve the accuracy of positioning, the positioning module uses deep learning technology to optimize the filtering algorithm. By training a deep neural network, the system noise and observation noise parameters in the filtering algorithm can be predicted based on the characteristics of the perception data and the contextual information of the environmental model, thereby improving the accuracy of positioning; the positioning module integrates multi-source data fusion technology to integrate data from different sensors (such as LIDAR point clouds, camera images, ultrasonic and infrared sensor data) to obtain more comprehensive location information. For example, LIDAR data is used to provide accurate distance and angle information, while camera images provide scene recognition and landmark matching to assist positioning; the positioning module runs in real time and can respond quickly to environmental changes. Through close collaboration with the environmental modeling module, the positioning module is able to use the latest environmental model to calibrate and optimize the position estimate. This feedback mechanism ensures that the machine can maintain accurate positioning even in a dynamic environment; the positioning module not only provides position estimates, but also includes an assessment of the accuracy and reliability of the estimates. By calculating the estimated uncertainty and confidence interval, the module can determine the credibility of the current positioning and trigger corrective measures or request additional perception data when necessary; the key parameters of the positioning module include the number of particles in the particle filter (for example, 1000 particles), the initial state covariance and process noise covariance in the Kalman filter, and the number of layers and neurons in the deep learning model. These parameters are carefully adjusted to ensure high-precision positioning while maintaining computational efficiency.

[0011] Preferably, the path planning module uses a deep reinforcement learning algorithm, especially a policy gradient-based method such as PPO. This algorithm continuously learns through interaction with the environment and optimizes the machine's action strategy. Through a large amount of simulation training, the algorithm can autonomously learn to select the best path under different environmental conditions; the module receives detailed three-dimensional models and occupancy grids from the environment modeling module, which provide the precise location and size of obstacles and information about the traversable area; these data are used as input to the path planning algorithm to ensure the accuracy and feasibility of path planning; the state space consists of the machine's current position, target position, surrounding obstacle information, and environmental characteristics. The action space defines the possible actions that the machine may take in each state, such as forward, backward, left turn, or right turn. Each action is associated with a specific speed and steering angle for precise control; the reward function is the core of reinforcement learning and is used to guide the learning process. In the path planning module, the reward function is designed to maximize the speed of reaching the target, minimize the path length, and avoid obstacles. This encourages the algorithm to consider efficiency and safety when planning the path; the module uses a deep neural network to approximate the policy or value function. The network structure may include multiple hidden layers, each containing tens to hundreds of neurons, and using the ReLU activation function. This network structure is able to process complex environmental information and generate effective path planning strategies; the path planning module can receive environmental update information from the perception module in real time and quickly adjust the current strategy to adapt to environmental changes. This dynamic path adjustment capability ensures the flexibility and adaptability of the machine when facing unknown obstacles or dynamic environments; the module is first trained in a simulated environment to quickly collect a large amount of experience and optimize the strategy. After that, the trained model is migrated to the real world and fine-tuned to adapt to the nuances of the actual environment. This approach improves the generalization ability of the algorithm; the performance of the module is evaluated by indicators such as success rate, path efficiency, and response time. Through continuous optimization and adjustment, path planning is ensured to be both fast and effective, meeting the requirements for efficiency and safety in practical applications.

[0012] Preferably, in the artificial potential field method, the strength of the attractive potential field is set to an extremely high value to ensure that the machine always moves towards the target position. The repulsive potential field is dynamically adjusted according to the proximity of the obstacle. When the machine approaches the obstacle, the strength of the repulsive potential field increases sharply, forcing the machine to change its path with extremely high priority. This extreme design makes the obstacle avoidance reaction extremely fast and obvious; the extremeness of the RRT algorithm is reflected in its extremely high sampling frequency and step length, ensuring that the algorithm can explore a wide space and find an obstacle avoidance path in an instant. The extreme parameter settings of the algorithm include: extremely high sampling density, step length reaching the maximum safe distance of the machine, and unlimited maximum expansion times, ensuring that the path to the target can be found in any complex environment; the dynamic window method is applied to the extreme in the obstacle avoidance module, and a safe trajectory is searched by considering all possible dynamic behaviors of the machine, including emergency stops and 180-degree turns. The extreme implementation of this method ensures that the machine can immediately find a collision avoidance strategy even in extremely urgent situations; the strategy selection mechanism of the obstacle avoidance module adopts an extreme decision logic, which can evaluate the environmental complexity and machine status at the microsecond level and select the most appropriate obstacle avoidance strategy. This extreme decision-making speed is achieved through pre-training and optimization of deep learning models; in order to achieve extreme real-time performance, the obstacle avoidance module runs on a high-performance computing platform with trillions of floating-point operations per second. Through parallel processing technology and GPU acceleration, the module can complete the entire process from sensing obstacles to executing obstacle avoidance actions within 50 milliseconds; the obstacle avoidance module was designed with applications in extreme environments in mind, such as environments with extremely low visibility, dense obstacles, or rapid dynamic changes. The module can maintain efficient operation under these extreme conditions to ensure the navigation safety of the machine.

[0013] Preferably, the decision-making module achieves efficient processing of complex environmental information and fast and accurate decision support by integrating a carefully designed decision tree and an advanced fuzzy logic controller. The module uses the hierarchical structure of the decision tree to classify and evaluate environmental states, where each node represents a specific environmental feature or machine state, and each branch represents a possible decision option. The fuzzy logic controller is used to handle uncertainty and ambiguity. By defining fuzzy sets and rule bases, sensor data is converted into fuzzy values ​​in the decision-making process, so that when faced with multiple options, it can simulate the decision-making method of human experts and quickly make the best decision. For example, when a machine faces the choice of turning left or going straight at an intersection, the decision tree will evaluate factors such as the target direction, traffic signals, and pedestrian flow, while the fuzzy logic will calculate the best action plan based on the fuzzy values ​​of these factors. This integrated approach not only improves the accuracy of decision-making, but also greatly enhances the adaptability and flexibility of the machine in complex environments.

[0014] Preferably, the motion control module adopts an adaptive control algorithm, which can dynamically adjust the speed and direction of the machine according to the predetermined path provided by the path planning module and the real-time feedback of the obstacle avoidance module. The core of the algorithm is a PID controller, in which the proportional (P), integral (I) and differential (D) parameters are carefully adjusted to achieve precise control of the machine's motion; in the PID controller, the proportional parameter Kp is set to 0.5 to enhance the response speed to the path deviation; the integral parameter Ki is set to 0.1 to eliminate the steady-state error; the differential parameter Kd is set to 0.05 to reduce the overshoot and oscillation of the system. The setting of these parameters ensures the stability and accuracy of the machine when tracking the path; the motion control module receives data from sensors in real time, such as the current position, speed and acceleration of the machine, combined with the output of the path planning and obstacle avoidance modules, for rapid feedback and adjustment. For example, when the machine detects an obstacle in front, the obstacle avoidance module will send instructions to slow down and turn, and the motion control module will quickly adjust the speed and direction to avoid collision; in order to improve the smoothness of the machine's motion, the motion control module also integrates a path smoothing algorithm. The algorithm generates a smooth navigation path by calculating the key points on the path, reducing the machine's sudden stops and starts when turning, thereby improving ride comfort and movement efficiency; the motion control module is implemented on a high-performance computing platform to ensure the algorithm's rapid response and efficient operation. The module's calculation frequency is set to 100Hz, which means that the control parameters are updated every 10 milliseconds, ensuring real-time control of the machine's movement; the motion control module not only focuses on the accuracy of path tracking, but also attaches importance to the safety and reliability of operation. The module is designed with multiple safety check mechanisms, such as speed limit, acceleration limit, and emergency stop function, to ensure that the machine can operate safely under various conditions.

[0015] Preferably, the learning and adaptation module accumulates experience by collecting data from the machine during the execution of the task, including successful path planning, effective obstacle avoidance operations, and special situations encountered. This data is used as training data to help the system identify effective strategies and areas for improvement; the module uses deep learning algorithms, especially convolutional neural networks (CNNs) and recurrent neural networks (RNNs), to process and analyze large amounts of spatiotemporal data. CNNs are used to extract features from image data of the environment, while RNNs are used to process time series of sensor data to predict environmental change trends; the module supports online learning, that is, updating the model in real time during the operation of the machine, and offline learning, that is, batch processing of collected data after the task is completed. Online learning uses small batch gradient descent methods to achieve fast updates at a lower computational cost; offline learning uses a larger amount of data for in-depth analysis; the learning and adaptation module optimizes path planning and obstacle avoidance strategies based on the output of the deep learning model. For example, if the model identifies that certain specific types of obstacles can be avoided more effectively by different methods, it will adjust the corresponding policy parameters; the module can adapt to new or changing environmental conditions. When the environment model shows significant changes, such as new obstacles appearing or old obstacles moving, the learning module triggers a re-evaluation of the current strategy and explores new solutions; the learning and adaptation module also includes a performance evaluation mechanism to quantify the learning effect. Evaluation indicators include path efficiency, obstacle avoidance success rate, and system response time. Based on the feedback of these indicators, the module can further adjust the learning algorithm and parameters; CNN may contain 5 convolutional layers and 3 fully connected layers, trained using ReLU activation function and cross entropy loss function. RNN may contain 2 LSTM layers, each with 128 neurons, optimized using sequence loss function.

[0016] Preferably, the safety monitoring module analyzes the real-time environmental data of the perception module, the precise location information of the positioning module, the path execution of the path planning module, and the dynamic feedback of the motion control module to ensure that the machine is always within a safe operating range. Once the system detects an abnormal situation such as approaching a safety boundary or an unknown obstacle, the safety monitoring module will immediately initiate an emergency response instruction, which may include automatic deceleration, emergency stop, or automatic switching to a safe path. In addition, the module also has a self-diagnosis function to monitor its own software and hardware status to ensure that the monitoring system itself is stable and reliable. Through these measures, the safety monitoring module ensures the highest safety standards for the machine when performing tasks and effectively reduces operational risks.

[0017] Preferably, the communication module ensures that information can be exchanged between machines in real time through efficient wireless communication protocols, such as Wi-Fi or Bluetooth. The module supports distributed control, allowing each machine to independently run its own navigation and obstacle avoidance algorithms, while achieving task coordination and optimization through information sharing mechanisms, such as sending and receiving data such as position, status, and task completion. For example, when a machine detects a new obstacle or completes a task, it broadcasts this information to other machines through the communication module, allowing the entire team to dynamically adjust its behavior and strategy. In addition, the communication module also has an error detection and retransmission mechanism to ensure the reliability and integrity of information transmission. Through this communication mechanism, the multi-machine system can more efficiently allocate tasks, share resources, and collaboratively solve complex problems, significantly improving the efficiency and effectiveness of the overall operation.

[0018] Preferably, the hardware interface module adopts a modular and standardized interface design, supports communication protocols including but not limited to I2C, SPI, UART and USB, and ensures plug-and-play with mainstream sensors, actuators and other peripherals on the market. For example, it can communicate with the LIDAR sensor through the I2C protocol to obtain environmental distance data, exchange image information with the camera module through SPI, and synchronize data with the IMU (inertial measurement unit) through UART. In addition, the hardware interface module also has a hardware abstraction layer (HAL), which provides a unified access interface for upper-level software, simplifying the development and maintenance of hardware drivers. The module is also designed with an expansion slot, allowing users to add or replace hardware components as needed, such as additional sensors or higher-performance actuators, thereby enhancing the scalability and adaptability of the system. Through these specific designs and parameter configurations, the hardware interface module ensures the flexibility of the autonomous navigation control system and the ability to update technology in the long term.

[0019] The above technical solution can bring the following three significant technical effects:

[0020] 1. Efficient autonomous navigation capability: Through the perception module optimized by deep learning algorithms, the machine can accurately identify and understand the surrounding environment in real time, including static and dynamic obstacles. This advanced perception capability enables the machine to autonomously navigate in unknown environments without a preset map; the application of reinforcement learning algorithms in the path planning module enables the machine to autonomously learn and plan the optimal path based on the current environmental status and target location. This path planning not only takes into account the efficiency of reaching the destination, but also takes into account safety and energy consumption, significantly improving the efficiency and reliability of navigation.

[0021] 2. Dynamic obstacle avoidance and adaptability: The integration of multi-modal obstacle avoidance strategies enables the machine to dynamically select the most appropriate obstacle avoidance algorithm based on the complexity of the environment and its own motion state. For example, when faced with dense obstacles, the system may choose the rapid exploration random tree (RRT) algorithm to quickly find a feasible path; when precise obstacle avoidance is required, the artificial potential field method may be used to ensure the accuracy of obstacle avoidance; the real-time update mechanism of the environmental modeling module ensures the machine's rapid response to environmental changes. This adaptability enables the machine to quickly adjust its action strategy when the environment changes, such as when obstacles move or new obstacles appear, thereby maintaining the continuity and safety of navigation.

[0022] 3. Intelligent decision-making and user interaction: The decision-making module combines decision trees and fuzzy logic, enabling the machine to quickly make the best choice when faced with complex decisions. This intelligent decision support not only improves the autonomy of the machine, but also enhances its ability to cope with complex environments; the user interaction module provides an intuitive user interface that allows the operator to monitor the status of the machine in real time, set navigation goals, and even manually intervene when necessary. This interactivity not only enhances the flexibility of the system, but also improves the user's control and trust in the operation of the machine. Description of the drawings:

[0023] Figure 1 The system block diagram of the present invention is

[0024] Figure 2 This is the flow chart of the path planning algorithm of the present invention

[0025] Figure 3 Flow chart of the learning and adaptation module of the present invention Specific implementation method:

[0026] The following will be combined with the attached embodiment of the present invention Figures 1 to 3 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] The perception module is the primary link in the autonomous navigation control system. Its design and implementation details are as follows: It integrates a series of high-precision sensors, including: a high dynamic range (HDR) camera with a resolution of 1080p, providing a smooth frame rate of 30fps to ensure image clarity; a 32-line LIDAR with a 360-degree field of view and a rotation speed of 10Hz, achieving a fine angular resolution of 0.1 degrees and a wide detection range of ±30 meters; four ultrasonic sensors, each with a detection angle of 15 degrees and a measurement range of 0.1 to 3 meters; and eight infrared sensors, designed for low-light environments, with a detection range of 0.5 meters and a 10-degree field of view.

[0028] All sensor data are synchronized through a high-precision time synchronization system based on GPS and precision clocks, which ensures microsecond-level timestamp accuracy and provides a solid foundation for data fusion.

[0029] The perception module uses a deep learning algorithm to process data. Image data is extracted through a five-layer convolution and three-layer fully connected CNN model, and the input image size is standardized to 224x224 pixels. LIDAR point cloud data directly learns spatial features through the PointNet network. The time series data of the ultrasonic sensor is analyzed by a 128-unit LSTM network to identify the speed and direction of moving obstacles. All models run on the NVIDIA Jetson TX2 platform, which has 256 CUDA cores and 6GB GDDR5 memory, providing tens of trillions of operations per second (TOPS) to meet the needs of real-time processing.

[0030] The feature fusion layer uses a weighted fusion method to weight features according to the importance and reliability of the sensor in a specific environment to generate a comprehensive environmental feature vector. This vector describes the location, size, type, and motion state of obstacles in detail, providing comprehensive support for navigation decisions.

[0031] The deep learning model is optimized using TensorRT, which reduces inference time and improves energy efficiency. This ensures that the perception module can not only run at high efficiency, but also respond quickly to environmental changes, providing continuous and accurate environmental perception information for the machine's autonomous navigation.

[0032] The environment modeling module first integrates multi-source data from the perception module, including images captured by the 1080p camera, point cloud data provided by the 32-line LIDAR, distance information from the ultrasonic sensor, and ambient light intensity from the infrared sensor. The preprocessing steps include denoising and edge enhancement of the image, filtering and downsampling of the point cloud data, and signal enhancement of the ultrasonic and infrared data to ensure that the data quality meets the modeling requirements.

[0033] Using LIDAR point cloud data, a 3D model was constructed through a voxelization algorithm, where the voxel size was set to 0.05 meters to balance the accuracy and computational efficiency of the model. During the model construction process, curvature analysis and plane detection technology were used to identify and reconstruct structures such as walls and ground in the environment.

[0034] The resolution of the occupancy grid is set to 0.1 meters per cell, covering a 100-meter radius around the machine. The dynamic map update algorithm runs at a frequency of 10Hz to respond to environmental changes in real time. When new obstacles or environmental changes are detected, the algorithm quickly updates the environmental model by adding or modifying the state of cells in the occupancy grid.

[0035] The environment modeling module uses a fusion algorithm based on Kalman filtering to fuse the data from different sensors into a consistent environment model. During the fusion process, the weight of each data source is dynamically adjusted according to the measurement uncertainty and timestamp of each sensor.

[0036] The environment modeling module runs on a computing platform with a high-performance GPU, which has the ability to perform at least tens of billions of floating-point operations per second. Through parallel processing technology and GPU acceleration, the module can complete the entire process from data reception to model update within 100 milliseconds.

[0037] The environment modeling module outputs a high-precision 3D environment model and occupancy grid, providing accurate spatial information for the machine's path planning and obstacle avoidance. The frequency and accuracy of model updates directly affect the safety and efficiency of machine navigation.

[0038] The positioning module uses an improved extended Kalman filter (EKF) algorithm, which handles the uncertainty of sensor data by introducing a nonlinear observation model. The initial state covariance matrix of the EKF is set to a diagonal matrix, and its diagonal elements represent the uncertainty of the initial position and velocity. For example, the initial standard deviation of the position is set to 0.5 meters, and the initial standard deviation of the velocity is set to 0.1 meters per second.

[0039] To further improve positioning accuracy, a deep neural network (DNN) is integrated into the module. This network is trained to predict the system dynamic model and observation model parameters in the particle filter. The DNN consists of 5 fully connected layers, each containing 128 neurons. The input is the feature vector of the perception data, and the output is the filter parameters, such as the predicted system noise standard deviation and the observed noise standard deviation.

[0040] The localization module fuses data from different sensors, including LIDAR point cloud data, visual odometry (VO), GPS signals, and IMU data. The data fusion strategy is based on a weighted average method, where the weight is dynamically determined by the current reliability and accuracy of each sensor. For example, LIDAR data has a higher weight at close range, while GPS is more reliable in open areas outdoors.

[0041] The localization module updates the machine’s position estimate in real time at a frequency of 10Hz to ensure synchronization with the environment modeling module. The module is designed with a feedback mechanism that uses landmarks and map information in the environment model to correct the position estimate, thereby improving the stability and accuracy of positioning.

[0042] The module not only outputs a position estimate, but also provides the uncertainty and confidence interval of the estimate. By calculating the covariance matrix of the EKF, the accuracy of the current position estimate can be evaluated. For example, the 68% confidence ellipse of the position estimate is represented by the principal components of the covariance matrix.

[0043] To meet the real-time requirements, the positioning module is implemented on a high-performance computing platform with a processing speed of at least 1GHz and at least 4GB of RAM. The deep learning model and filtering algorithm are optimized to reduce the amount of calculation and increase the processing speed, ensuring that the module can complete a complete positioning update within 200 milliseconds.

[0044] The path planning module uses the PPO algorithm, which is a highly effective policy gradient method suitable for processing continuous action spaces. The network structure of the PPO algorithm includes 4 fully connected layers, each with 256 neurons, with the input being the state vector and the output being the probability distribution of the action. The network uses the ReLU activation function and is trained using the Adam optimizer.

[0045] The state space consists of the current position of the machine, the target position, obstacle information, and environmental characteristics, and the data of each dimension is normalized. The action space defines 8 possible movement actions, including linear speeds between 0.1 and 1 m / s and angular speeds between -30 and 30 degrees.

[0046] The reward function is designed as a weighted sum, including a negative distance weighted 0.5 to reach the target, a negative weighted 0.3 for the path length, and a weighted 0.2 for avoiding collisions. This design encourages the algorithm to find the shortest and safest path.

[0047] The module is trained in a simulated environment using a highly realistic 3D environment model that contains a variety of obstacles and changing terrain. The training cycle is set to 1000 episodes, each lasting 1000 steps.

[0048] In real-time operation, the path planning module receives updated environmental information at a frequency of 10Hz and dynamically adjusts the path. When a new obstacle or environmental change is detected, the module is able to recalculate and generate a new path within 200 milliseconds.

[0049] The performance of the module is evaluated by indicators such as path length, time to target, number of collisions, and computation time. For example, the standard deviation of the target arrival time is set to 10% of the expected time to ensure the stability of path planning.

[0050] The user interface allows the operator to set the target position, monitor the path generation process and intervene manually when necessary with a simple click. This design improves the flexibility and user-friendliness of the system.

[0051] The obstacle avoidance module is a key part of the autonomous navigation control system. Its design and implementation need to ensure that the machine can effectively avoid obstacles in various complex environments. The following are the specific implementation details of this module, including specific parameters: In the artificial potential field method, the strength parameter k of the attractive potential field a Set to 10 to ensure that the machine has a strong attraction to the target location. The strength parameter k of the repulsive potential field r The distance d between the machine and the obstacle is dynamically adjusted, and the formula is: where d min The minimum safe distance between the machine and obstacles is set to 0.5 meters. This setting allows the machine to react quickly when approaching obstacles. The sampling density of the RRT algorithm is set to 1000 times per second, and the step length is 10% of the maximum safe distance of the machine, that is, 0.5 meters. There is no limit on the maximum number of expansions to ensure that the algorithm can find a path to the target in any complex environment. This high sampling frequency and step length setting allows the RRT algorithm to quickly explore the environment and generate obstacle avoidance paths. Considering the maximum speed of the machine v max is 1 m / s, and the maximum acceleration is a max 0.5 m / s 2 The control window is set to twice the current speed of the machine to ensure sufficient time for obstacle avoidance. This approach is able to search for a safe trajectory while taking into account the dynamic characteristics of the machine.

[0052] The strategy selection mechanism is implemented through a decision tree, and each node of the decision tree makes a judgment based on the complexity of the environment and the state of the machine. For example, when the obstacle density exceeds the threshold of 80%, it automatically switches to the RRT algorithm; when the machine speed exceeds 0.8v maxWhen selecting an obstacle avoidance method, dynamic avoidance methods are given priority. This mechanism ensures that the choice of obstacle avoidance strategy is both fast and accurate. The obstacle avoidance module runs on a platform that has a floating-point computing capacity of 32 trillion times per second. Through parallel processing technology and GPU acceleration, the module is able to complete the entire process from sensing obstacles to performing obstacle avoidance actions within 50 milliseconds. This high-performance computing platform ensures the real-time and reliability of the obstacle avoidance module. The obstacle avoidance module has been extensively tested in simulated extreme environments, including high-density obstacle environments, fast-moving obstacle environments, and low-light environments. Test results show that the module is able to maintain an obstacle avoidance success rate of more than 95% under these extreme conditions, proving the superiority of its design.

[0053] The decision-making module uses an integrated decision tree and fuzzy logic controller, which is carefully designed to achieve fast and accurate decision support. The decision tree of this module consists of 15 layers in depth, each layer contains up to 5 decision nodes, which are evaluated based on environmental characteristics such as obstacle distance, target direction deviation and other parameters. The threshold of each node is optimized to ensure that reasonable decisions can be made in different situations. The fuzzy logic controller contains 3 input variables: obstacle proximity, path deviation and time urgency. Each variable is divided into 5 fuzzy sets, such as "low", "medium", "high", etc., and the corresponding membership function is defined. The controller's rule base contains 200 fuzzy rules, which are refined based on expert experience and a large amount of data from the simulated environment. For example, when the obstacle proximity is "high" and the path deviation is "medium", the rule may indicate "slow down and find an alternative path". The entire decision-making process is completed within 50 milliseconds, ensuring real-time responsiveness in a dynamic environment. Through this method that combines precise decision trees and flexible fuzzy logic, the machine can quickly make the best decision in multiple options, improving the efficiency and safety of navigation.

[0054] The motion control module uses an adaptive control algorithm that can dynamically adjust the speed and direction of the machine based on the predetermined path provided by the path planning module and the real-time feedback from the obstacle avoidance module. The core of the algorithm is the PID controller, in which the proportional (P), integral (I) and differential (D) parameters are carefully adjusted to achieve precise control of the machine's motion.

[0055] In the PID controller, the proportional parameter Kp is set to 0.5 to enhance the response speed to the path deviation; the integral parameter Ki is set to 0.1 to eliminate the steady-state error; the differential parameter Kd is set to 0.05 to reduce the overshoot and oscillation of the system. The settings of these parameters ensure the stability and accuracy of the machine when tracking the path.

[0056] The motion control module receives real-time data from sensors, such as the machine's current position, speed, and acceleration, and combines it with the output of the path planning and obstacle avoidance modules to provide rapid feedback and adjustments. For example, when the machine detects an obstacle ahead, the obstacle avoidance module sends instructions to slow down and turn, and the motion control module quickly adjusts speed and direction to avoid a collision.

[0057] In order to improve the smoothness of the machine's movement, the motion control module also integrates a path smoothing algorithm. This algorithm generates a smooth navigation path by calculating the key points on the path, reducing the machine's sudden stops and starts when turning, thereby improving ride comfort and movement efficiency.

[0058] The motion control module is implemented on a high-performance computing platform, ensuring the fast response and efficient operation of the algorithm. The module's calculation frequency is set to 100Hz, which means that the control parameters are updated every 10 milliseconds, ensuring real-time control of the machine's motion.

[0059] The learning and adaptation module is an important part of the autonomous navigation control system. It enables the machine to learn from experience and continuously optimize its navigation and obstacle avoidance strategies. The following are the specific implementation details of this module:

[0060] The learning and adaptation module accumulates experience by collecting data from the machine during the execution of tasks, including successful path planning, effective obstacle avoidance operations, and special situations encountered. This data is used as training data to help the system identify effective strategies and areas for improvement.

[0061] The module uses deep learning algorithms, especially convolutional neural networks (CNN) and recurrent neural networks (RNN), to process and analyze large amounts of spatiotemporal data. CNN is used to extract features from environmental image data, while RNN is used to process the time series of sensor data to predict environmental change trends.

[0062] The module supports online learning, which updates the model in real time while the machine is running, and offline learning, which processes the collected data in batches after the task is completed. Online learning uses small batch gradient descent to achieve fast updates at a low computational cost; offline learning uses larger amounts of data for in-depth analysis.

[0063] The learning and adaptation module optimizes the path planning and obstacle avoidance strategies based on the output of the deep learning model. For example, if the model recognizes that certain types of obstacles can be avoided more effectively using different methods, it will adjust the corresponding strategy parameters.

[0064] The module is able to adapt to new or changing environmental conditions. When the environmental model shows significant changes, such as new obstacles appearing or old obstacles moving, the learning module triggers a re-evaluation of the current policy and explores new solutions.

[0065] The learning and adaptation module also includes a performance evaluation mechanism to quantify the learning effect. The evaluation indicators include path efficiency, obstacle avoidance success rate and system response time. Based on the feedback of these indicators, the module can further adjust the learning algorithm and parameters.

[0066] A CNN might contain 5 convolutional layers and 3 fully connected layers, trained using the ReLU activation function and the cross entropy loss function. An RNN might contain 2 LSTM layers, each with 128 neurons, optimized using the sequence loss function.

[0067] The safety monitoring module is a key link to ensure the safe operation of the autonomous navigation control system. It monitors the operating status of the machine in real time through a series of sophisticated algorithms and preset thresholds. The module sets speed limit parameters, such as the maximum safe speed of 1.5 meters per second, and the deceleration threshold when approaching an obstacle, such as automatically reducing the speed to 0.5 meters per second within 1 meter of the obstacle. At the same time, it also defines the trigger conditions for emergency stops. For example, when the distance between the machine and the obstacle is less than 0.3 meters, the system will respond and perform an emergency stop within 100 milliseconds. The safety monitoring module also has a self-detection function that can check the integrity of sensor data and the response time of the processor every 5 seconds to ensure that the module itself is operating normally. In addition, the module also sets environmental monitoring parameters, such as automatically starting the safety navigation protocol in a low-light environment when the light intensity is lower than the preset threshold. Through these specific parameter settings and real-time monitoring, the safety monitoring module provides a solid safety guarantee for the autonomous navigation of the machine.

[0068] The user interaction module is a crucial component of the autonomous navigation control system. It provides an intuitive and easy-to-use graphical user interface (GUI) that enables the operator to monitor the system status of the machine in real time, including the current position, path planning, obstacle avoidance behavior, and environmental perception information. Through this interface, the operator can easily set the target position, adjust system parameters such as speed limit, safety distance, and response sensitivity, and perform manual intervention when necessary, such as remotely controlling the movement of the machine or directly modifying its behavior strategy. In addition, the user interaction module also includes real-time data visualization tools such as map display, path highlighting and status indicator lights, as well as emergency stop buttons and manual override options to ensure that the operator can respond quickly in any situation. This design not only improves the flexibility and adaptability of the system, but also greatly enhances the user experience and system safety.

[0069] The communication module is the core of multi-machine collaborative operation in the autonomous navigation control system. It uses high-speed wireless communication technology, such as 5GHz Wi-Fi, to ensure that machines can exchange information in real time at a data transmission rate of at least 100Mbps. The module uses a TCP / IP-based protocol stack to achieve low-latency data packet transmission, with an average delay of no more than 10ms, ensuring the real-time nature of information. The communication module is designed with an efficient information encoding and decoding mechanism, which can achieve at least 20 status updates and task information exchanges per second on each machine. The module integrates error detection and automatic retransmission functions, and ensures that the accuracy of data packet transmission reaches 99.9% through CRC check and ACK confirmation mechanisms. In a multi-machine system, the communication module supports a network topology of at least 128 nodes, allowing large-scale machine teams to perform distributed control and information sharing. For example, when the system is deployed in an area of ​​10,000 square meters, the communication module can ensure that at least 10 machines can coordinate tasks and share resources at the same time without affecting the overall operation efficiency due to communication delays or packet loss.

[0070] The hardware interface module provides compatibility with various hardware devices through standardized communication protocols and hardware abstraction layer (HAL). The module supports I2C communication protocol with a data transmission rate of up to 400kbps, SPI protocol supports a transmission rate of up to 10Mbps, and UART protocol provides serial communication capability of up to 115.2kbps. The module is designed with 4 I2C channels, 3 SPI channels and 2 UART ports to ensure the connection of various types of sensors and actuators. For example, a 32-line LIDAR sensor, two RGB-D cameras and an IMU can be connected at the same time to achieve full range of environmental perception and status monitoring. In addition, the hardware interface module is also equipped with GPIO (general purpose input and output) pins to support the input and output of digital and analog signals, enhancing the scalability of the system. The power management unit of the module can provide stable 5V and 3.3V power supplies to ensure that all connected hardware devices can obtain reliable power supply. Through these specific parameter and function designs, the hardware interface module provides strong hardware compatibility and flexible expansion capabilities for autonomous navigation control systems.

[0071] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An adaptive navigation control system for unknown environments based on deep learning, characterized in that: The system includes: a perception module, an environmental modeling module, a positioning module, a path planning module, an obstacle avoidance module, an integrated plurality of obstacle avoidance algorithms, a decision-making module, a motion control module, a learning and adaptation module, a security monitoring module, a user interaction module, a communication module, and a hardware interface module. The perception module collects rich data about the environment through integrated plurality of sensors, and these data are then input into a deep learning algorithm for fusion and feature extraction, providing the system with an in-depth understanding of the surrounding environment. The extracted feature information is used by the environmental modeling module to construct and update the three-dimensional model or occupancy grid of the environment in real time, providing the machine with accurate spatial information. The positioning module utilizes these feature information and the environmental model to achieve accurate positioning of the machine through an optimized particle filter or Kalman filter algorithm. This positioning information is crucial for the path planning module. The module adopts a reinforcement learning algorithm to autonomously learn and plan the optimal path according to the current position, target position and environmental model of the machine. While planning the path, the obstacle avoidance module analyzes the environmental model and perception data in real time, integrates a plurality of obstacle avoidance algorithms, and dynamically The most appropriate obstacle avoidance strategy is selected to ensure safe navigation of the machine in a complex environment. The decision-making module combines decision trees and fuzzy logic to provide fast and accurate decision support for the information generated during path planning and obstacle avoidance. Based on these decisions, the motion control module adjusts the speed and direction of the machine in real time to ensure that it moves smoothly along the planned path. The learning and adaptation module continuously collects empirical data, optimizes navigation and obstacle avoidance strategies through deep learning algorithms, and improves the long-term performance of the system. The safety monitoring module continuously monitors the operating status of the entire system to ensure the safety of operation and provides emergency response when potential risks are detected. The user interaction module provides an interface for users, allowing them to monitor the system status in real time and intervene manually when necessary, thereby enhancing the flexibility and controllability of the system. In a multi-machine system, the communication module is responsible for achieving effective information exchange and task coordination between machines to ensure high efficiency of team work. The hardware interface module ensures that the system is compatible with various hardware devices, including sensors, actuators and other peripherals, which facilitates the expansion and upgrade of the system.

2. The system of claim 1, wherein the perception module comprises at least one type of sensor, such as a camera, a LIDAR, an ultrasonic sensor, or an infrared sensor.

3. The system according to claim 1, characterized in that The environmental modeling module adopts a real-time data updating algorithm to adapt to the dynamic changes of the environment.

4. The system according to claim 1, characterized in that The positioning module uses perception data and environmental models to accurately position the machine.

5. The system according to claim 1, characterized in that The path planning module considers the current position of the machine, the target position and the environment model to perform path planning.

6. The system according to claim 1, characterized in that The obstacle avoidance module includes an artificial potential field method, a rapidly exploring random tree (RRT) algorithm or a dynamic window method.

7. The system according to claim 1, characterized in that The decision-making module is capable of processing complex environmental information and providing decision support.

8. The system according to claim 1, characterized in that The motion control module optimizes the path following performance of the machine.

9. The system according to claim 1, characterized in that The learning and adaptation module allows the system to adapt to new or changing environmental conditions through empirical learning.

10. The system according to claim 1, characterized in that The communication module supports distributed control and information sharing in a multi-machine system.

Images