CN-121979192-A - Unknown environment self-adaptive navigation control system based on deep learning

CN121979192ACN 121979192 ACN121979192 ACN 121979192ACN-121979192-A

Abstract

The system comprises a plurality of modules, wherein a heterogeneous sensing array interface acquires spatial characteristics, dynamic target information and self running state data of an unknown environment in real time, a time sequence environment dynamic modeling module is responsible for understanding a current environment structure and dynamic elements, a multi-mode perception optimization module is responsible for deeply fusing multi-source perception information, an unknown environment risk prediction module predicts potential collision risk and sensor failure risk, an adaptive navigation decision module generates an initial optimal path, a dynamic path planning control module realizes dynamic correction and accurate execution of the path, a running error self correction module corrects accumulated errors of modules such as a visual odometer in real time, a navigation state visualization module presents a system state through a visual interface, and supports manual monitoring and intervention, and a system operation management module is responsible for system management.

Inventors

GUO FENG
ZHANG CHENYAN
HUANG CHENGXUAN
LIU XINYI
Song Feiyang
Xu Yingsha
HUANG JUNDONG
YU ZHENG
WANG YIXI
Xia Shengdong
LI GUORUI
ZHANG YONG
YAO TIANLU
FENG HAO
QIU SHUANG
JIAO HANLIN
ZHANG XIANFEI
ZHANG XIONG
Tong Yongfei
ZHOU YUTING

Assignees

国网湖北省电力有限公司信息通信公司
湖北华中电力科技开发有限责任公司

Dates

Publication Date: 20260505
Application Date: 20251217

Claims (10)

1. The unknown environment self-adaptive navigation control system based on deep learning is characterized by comprising a heterogeneous sensing array interface, a time sequence environment dynamic modeling module, a multi-mode perception optimization module, an unknown environment risk prediction module, a self-adaptive navigation decision module, an operation error self-correction module and a dynamic path planning control module; the heterogeneous sensing array interface integrates multi-type sensor equipment, acquires spatial characteristics of unknown environments, dynamic target information and self running state data in real time, and acquires multi-mode data streams; The time sequence environment dynamic modeling module distinguishes the current environment structure and the dynamic elements based on the multi-mode data flow, and builds an environment map updated in real time; The multi-mode sensing optimization module is used for calibrating contribution degrees of the sensors on line based on the environment map updated in real time, the multi-mode data stream and the historical navigation data, and isolating or degrading the sensors with the operation contribution degrees lower than a set value; the unknown environmental risk prediction module predicts potential collision risk and sensor failure risk based on the multi-mode data stream and the historical navigation data to obtain a comprehensive risk evaluation value; the self-adaptive navigation decision module dynamically adjusts a navigation strategy according to the environment map updated at the time and the comprehensive risk evaluation value to generate an optimal path and a navigation strategy instruction; The dynamic path planning control module receives the optimal path and the navigation strategy instruction, and adjusts the executing mechanism in real time to realize dynamic correction and accurate execution of the path; And the running error self-correction module corrects the long-term drift error of each sensor in real time through loop detection.
2. The unknown environment self-adaptive navigation control system based on deep learning of claim 1, wherein the heterogeneous sensing array interface is used as a system data input front end, a solid-state laser radar array, a 4D millimeter wave radar module, a panoramic vision sensing unit, a high-precision IMU/GNSS integrated navigation module and an environment meteorological sensor are integrated, point cloud, images, millimeter wave echoes, inertial data and environment parameters are acquired in real time through a hardware timestamp synchronization and flexible sampling interface, and a standardized multi-mode data stream with time-space alignment is output.
3. An unknown environment adaptive navigation control system based on deep learning as claimed in claim 1, wherein: The time sequence environment dynamic modeling module performs time sequence alignment pretreatment and dynamic feature extraction on the multi-mode data stream, separates a static environment structure and dynamic obstacles in real time based on the utilization of a dynamic SLAM technology, performs dynamic point cloud filtering, and builds an environment map updated in real time, wherein the update frequency is more than or equal to 10Hz; the calculation formula of the dynamic point cloud filtering in the time sequence environment dynamic modeling module is as follows: ; in the formula (i), A dynamic point probability mask representing the current frame t, For the number of point clouds, each element represents the probability that the corresponding three-dimensional point belongs to a dynamic target; representing a three-dimensional point cloud dynamic target segmentation network, inputting multi-mode data for end-to-end training, Representing the laser point cloud of the current frame, Representing the visual image of the current frame, Representing the aggregate data of the historical frames, For the number of points of the history map, Representing network learnable parameters.
4. The unknown environmental adaptive navigation control system based on deep learning of claim 1, wherein the multi-modal sensing optimization module calibrates the contribution degree of each sensor on line based on the environmental map updated in real time, the multi-modal data stream and the historical navigation data, and the fusion weight calculation formula is: ; in the formula (i), Representation pair in a system The composite scores of the individual sensors are summed, Representing the number of sensors; representing the final fusion weight of the ith sensor at time t, A quality score for sensor i representing the depth Q network assessment, Representing the kalman filter confidence of sensor i, Representing a predicted value of the failure probability of the sensor i, The probability of failure compensation is indicated, Representing the impact of controlling the empirical quality score, Indicating the impact of controlling the real-time statistical confidence, Representing the impact of controlling the predictive failure prediction, Representing an exponential function; The multi-mode perception optimization module recognizes a solid-state laser radar array, a 4D millimeter wave radar module, a panoramic vision perception unit or an environmental meteorological sensor When the sensor failure isolation and system degradation operation protection are automatically started by the module, and the perception master right and the dynamic switching of the navigation mode are performed; The multi-mode perception optimization module identifies a high-precision IMU/GNSS integrated navigation module or more than two sensors When the driver or remote operator is requested to take over.
5. The unknown environmental adaptive navigation control system based on deep learning of claim 1, wherein the unknown environmental risk prediction module predicts potential collision risk and sensor failure risk based on multi-modal data flow and historical navigation data to obtain a comprehensive risk assessment value; the unknown environmental risk prediction module is internally provided with a space-time diagram neural network, and calculates a comprehensive risk evaluation value according to the multi-mode data stream and the historical data The calculation formula is as follows: ; in the formula (i), The comprehensive risk assessment value is indicated, Representing the total number of predicted scenarios, i.e. the number of all possible future development scenarios considered by the system, Represents the occurrence probability of the kth scene, takes a value range of 0-1, Representing a risk value assessment of the kth scenario, Representing a risk sensitivity coefficient for adjusting the sensitivity of the system to risk, Representing a risk attenuation factor.
6. The unknown environmental adaptive navigation control system based on deep learning of claim 1, wherein the adaptive navigation decision module dynamically adjusts the navigation strategy according to the environmental map and the comprehensive risk assessment value updated at the time: the self-adaptive navigation decision module filters the dynamic point cloud and then performs the analysis according to the updated environment map and the comprehensive risk evaluation value Dynamically adjusting a navigation strategy: When (when) When the performance priority strategy is adjusted; When (when) When the method is used, a conservative passing strategy is adjusted; When (when) When the method is used, active safety strategy adjustment is carried out, and a driver or a remote operator is requested to take over; and generating an optimal path and a navigation strategy instruction based on the corresponding navigation strategy.
7. An unknown environment adaptive navigation control system based on deep learning as claimed in claim 1, wherein: The running error self-correction module calculates loop detection confidence coefficient The calculation formula is as follows: ; in the formula (i), The confidence of loop-back detection is indicated, ∈(0,1), Representing ORB feature similarity, and calculating the ratio of the successful matching number to the total feature number by violently matching ORB feature points of the current frame and the candidate closed-loop frame, wherein the range is 0, 1; Representing pose consistency scores, solving the relative pose of candidate closed-loop frames by using a PnP algorithm, evaluating the deviation between the candidate closed-loop frames and the predicted value of the pose of the odometer, wherein the score is higher when the deviation is smaller; The time difference between the current time and the candidate closed-loop frame time is represented, and the longer the time is, the lower the closed-loop reliability is; 、、 Penalty weight coefficients respectively representing feature similarity, pose consistency scores and time differences; Representing a sigmoid function; When (when) And when the drift error is more than 0.85%, automatically triggering global pose diagram optimization, and inhibiting the long-term drift error to be within 0.5%.
8. The unknown environment self-adaptive navigation control system based on deep learning of claim 1, wherein the navigation control system further comprises a navigation state visualization module and a system operation and maintenance management module; The navigation state visualization module presents a heterogeneous sensing array interface module, a time sequence environment dynamic modeling module, a multi-mode perception optimization module, an unknown environment risk prediction module, a self-adaptive navigation decision module, an operation error self-correction module and a dynamic path planning control module in a visual interface, and simultaneously supports manual monitoring and intervention; the system operation and maintenance management module is responsible for equipment state monitoring, data storage and backup, algorithm model iteration update and authority management.
9. An unknown environment self-adaptive navigation control method based on deep learning is characterized in that: the unknown environment self-adaptive navigation control method based on the deep learning is realized based on the unknown environment self-adaptive navigation control system based on the deep learning according to any one of claims 1 to 8; the heterogeneous sensing array interface integrates multi-type sensor equipment, acquires spatial characteristics of unknown environments, dynamic target information and self running state data in real time, and acquires multi-mode data streams; distinguishing a current environment structure from dynamic elements based on the multi-mode data stream, and constructing an environment map updated in real time; based on the real-time updated environment map, the multi-mode data stream and the historical navigation data, calibrating the contribution degree of each sensor on line, and isolating or degrading the sensor with the operation contribution degree lower than a set value; based on the multi-mode data stream and the historical navigation data, predicting potential collision risk and sensor failure risk to obtain a comprehensive risk assessment value; dynamically adjusting a navigation strategy according to the updated environment map and the comprehensive risk evaluation value to generate an optimal path and a navigation strategy instruction; Receiving an optimal path and a navigation strategy instruction, and adjusting an executing mechanism in real time to realize dynamic correction and accurate execution of the path; And the running error self-correction module corrects the long-term drift error of each sensor in real time through loop detection.
10. An unknown environment self-adaptive navigation control device based on deep learning is characterized in that: The unknown environment self-adaptive navigation control device based on deep learning comprises a memory and a processor, wherein the memory is used for storing computer program codes and transmitting the computer program codes to the processor; The processor is configured to execute the flow of the deep learning-based unknown environment adaptive navigation control system according to any one of claims 1 to 8 according to instructions in the computer program code.

Description

Unknown environment self-adaptive navigation control system based on deep learning Technical Field The invention relates to the technical field of autonomous navigation control, in particular to an unknown environment self-adaptive navigation control system based on deep learning. Background As autonomous systems extend to highly complex, strongly uncertain scenarios, traditional navigation methods face multiple technical bottlenecks. Firstly, the adaptability defect of dynamic environment is obvious, the existing algorithm generally depends on a static environment assumption, when the system is faced with moving obstacles or sudden interference, the system response is lagged and manual intervention is needed to recalibrate, so that the positioning accuracy is suddenly reduced, secondly, the multi-mode sensing Fusion has structural defects, the space-time asynchronous characteristics and the measuring noise difference of a laser radar, a millimeter wave radar and a camera are caused, the traditional Early/Late Fusion strategy is difficult to effectively extract complementary characteristics, especially, the systematic failure is easily caused under extreme working conditions (such as vision shielding), thirdly, the accumulated error problem in a long-term task is outstanding, the deep learning modules such as a vision odometer and the like are influenced by iterative drift, and the path redundancy rate exceeds an industry threshold value by being matched with an unreliable closed-loop detection mechanism. The current research is limited by high cost of real scene data acquisition, most of results only stay in a simulation verification stage, and an intelligent navigation framework with real-time environment modeling, cross-modal feature co-evolution and incremental topology learning needs to be constructed so as to break through core bottlenecks of insufficient dynamic scene generalization capability, low fault tolerance rate of heterogeneous sensors, long-period task error divergence and the like. Disclosure of Invention The invention aims to overcome the problems in the prior art, provides an unknown environment self-adaptive navigation control system based on deep learning, has the advantages of real-time environment dynamic understanding, multi-mode perception co-evolution and long-term running error self-correction, and solves the problems of poor adaptability, low fusion efficiency of heterogeneous sensors and navigation misalignment caused by accumulated errors in long-term tasks in a dynamic scene in the traditional method. In order to achieve the above object, the technical solution of the present invention is: The invention provides an unknown environment self-adaptive navigation control system based on deep learning, which specifically comprises a heterogeneous sensing array interface, a time sequence environment dynamic modeling module, a multi-mode perception optimization module, an unknown environment risk prediction module, a self-adaptive navigation decision module, an operation error self-correction module and a dynamic path planning control module; the heterogeneous sensing array interface integrates multi-type sensor equipment, acquires spatial characteristics of unknown environments, dynamic target information and self running state data in real time, and acquires multi-mode data streams; The time sequence environment dynamic modeling module distinguishes the current environment structure and the dynamic elements based on the multi-mode data flow, and builds an environment map updated in real time; The multi-mode sensing optimization module is used for calibrating contribution degrees of the sensors on line based on the environment map updated in real time, the multi-mode data stream and the historical navigation data, and isolating or degrading the sensors with the operation contribution degrees lower than a set value; the unknown environmental risk prediction module predicts potential collision risk and sensor failure risk based on the multi-mode data stream and the historical navigation data to obtain a comprehensive risk evaluation value; the self-adaptive navigation decision module dynamically adjusts a navigation strategy according to the environment map updated at the time and the comprehensive risk evaluation value to generate an optimal path and a navigation strategy instruction; The dynamic path planning control module receives the optimal path and the navigation strategy instruction, and adjusts the executing mechanism in real time to realize dynamic correction and accurate execution of the path; And the running error self-correction module corrects the long-term drift error of each sensor in real time through loop detection. The heterogeneous sensing array interface is used as a system data input front end, integrates a solid-state laser radar array, a 4D millimeter wave radar module, a panoramic vision sensing unit, a high-precision IMU/GNSS integrated navigation module and