Guidance optimization method for AI self-adaptive unmanned aerial vehicle for high-speed long voyage

Abstract

The invention discloses a high-speed long-flight AI adaptive unmanned aerial vehicle guidance optimization method which comprises the following steps that multi-source sensing data of a target area are collected in real time through a multi-mode sensor, and the multi-mode sensor comprises a thermal imaging module, a visible light camera module, an inertial measurement unit and a global positioning system; performing dynamic target identification and tracking on the multi-source sensing data acquired in the step S100 based on a deep neural network; generating an optimal attack path in real time through a reinforcement learning algorithm according to a motion track prediction result of the target in combination with flight state data of the unmanned aerial vehicle and environmental constraint conditions; and when the target enters the preset strike range, triggering an AI terminal guidance function. The method has the following advantages and effects: the autonomous combat ability of the unmanned aerial vehicle in a complex dynamic environment is greatly improved, real-time target identification, path planning and attack optimization of the unmanned aerial vehicle in a complex battlefield environment are realized, and the combat efficiency and precision are improved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to an AI adaptive guidance optimization method for high-speed, long-distance UAVs. Background Art

[0002] In existing drone technology, the inaccurate multimodal data fusion and poor real-time path planning in complex dynamic environments are key issues hindering autonomous guidance capabilities. Traditional drone systems rely on single sensors or static models for target tracking and path planning, making them difficult to cope with the challenges of high-speed maneuvering targets, dense obstacles, and environmental interference. This is specifically manifested in the following two aspects: First, the lack of spatiotemporal synchronization between multiple sensors results in data timestamp misalignment due to frame rate differences between thermal and visible light sensors, as well as hardware latency. This leads to cumulative errors in target trajectory prediction, affecting the accuracy of subsequent path planning. Second, dynamic path planning is inefficient. Traditional algorithms, with their high computational complexity under complex constraints, are unable to generate feasible paths that conform to the drone's dynamic model in real time, resulting in response delays and collision risks. These issues directly lead to inaccurate target tracking, unreliable attack paths, and biased strike effectiveness assessments in complex battlefield environments, severely limiting their practical application value in highly dynamic scenarios. Summary of the Invention

[0003] The purpose of the present invention is to provide an AI adaptive UAV guidance optimization method for high-speed and long-distance navigation to solve the problems raised in the background technology.

[0004] The above technical objectives of the present invention are achieved through the following technical solutions:

[0005] A high-speed, long-distance AI adaptive UAV guidance optimization method includes the following steps:

[0006] S100, collecting multi-source perception data of the target area in real time through a multimodal sensor, the multimodal sensor including a thermal imaging module, a visible light camera module, an inertial measurement unit, and a global positioning system; wherein the thermal imaging module is used to obtain target thermal radiation characteristics, the visible light camera module is used to obtain target morphological characteristics, the inertial measurement unit is used to monitor the attitude angle and acceleration of the UAV in real time, and the global positioning system is used to locate the geographic coordinates of the UAV;

[0007] S200, performing dynamic target recognition and tracking on the multi-source perception data collected in step S100 based on a deep neural network, including: performing spatiotemporal alignment and fusion of thermal radiation features and target morphological features to generate a target feature vector; performing multi-scale feature extraction on the target feature vector using a preset convolutional recurrent neural network, and outputting a target motion trajectory prediction result;

[0008] S300: Based on the target's trajectory prediction results, combined with the drone's flight status data and environmental constraints, an optimal attack path is generated in real time through a reinforcement learning algorithm;

[0009] S400: When the target enters the preset strike range, the AI ​​terminal guidance function is triggered, including: verifying the target lock status through the data closed loop of the multi-modal sensor.

[0010] By adopting the above technical solution, by combining multimodal sensors, deep neural networks and reinforcement learning algorithms, the optimal attack path can be generated dynamically in real time; through multimodal sensors, multi-source perception data of the target area can be collected in real time, which provides a rich data source for target identification and tracking, thereby improving the adaptability and response speed of the drone to dynamic targets; using deep neural networks to perform dynamic target identification and tracking on perception data improves the accuracy and robustness of target detection; at the same time, based on the target's motion trajectory prediction, drone flight status data and environmental constraints, the optimal attack path is generated through a reinforcement learning algorithm, which can ensure that accurate strike missions can be completed even in complex environments; the advantage of the present invention lies in its dynamic adaptability and autonomy. During the execution of the mission, the strategy can be adjusted in real time without human intervention to cope with the ever-changing battlefield environment.

[0011] Further, in step S200, dynamic target recognition and tracking are performed on the multi-source perception data collected in step S100 based on a deep neural network, including the following sub-steps:

[0012] S201, performing multimodal preprocessing on raw data of the thermal imaging module and the visible light camera module;

[0013] S202, performing cross-modal feature fusion and target feature vector generation based on the output of step S201, specifically including:

[0014] S203: Based on the output of step S202, multi-scale feature extraction and trajectory prediction are implemented through a preset convolutional recurrent neural network, specifically including:

[0015] S204: Perform physical constraint verification and optimization of the trajectory prediction result based on the output of step S203, specifically including:

[0016] S205. Based on the outputs of step S203 and step S204, the convolutional recurrent neural network feature extraction strategy is updated through an online adaptive learning module.

[0017] By adopting the above technical solution, based on steps S201 to S205, the target recognition and tracking capabilities are further enhanced; the multimodal preprocessing technology optimizes the data fusion effect of the thermal imaging module and the visible light camera module, ensuring the accuracy and real-time performance of target recognition; in addition, multi-scale feature extraction and trajectory prediction of target features are performed through the convolutional recurrent neural network, which not only improves the accuracy of target motion trajectory prediction, but also effectively responds to complex battlefield changes; the physical constraint verification and optimization steps ensure the feasibility and reliability of the prediction results, and provide accurate information for subsequent path planning; through the online adaptive learning module, the feature extraction strategy of the convolutional recurrent neural network can be continuously optimized to adapt to the ever-changing environment and target characteristics, thereby enhancing the self-learning ability and adaptability of the system.

[0018] Further configuration is that the step S201 specifically includes:

[0019] S2011. Eliminate inter-sensor acquisition delays through a spatiotemporal alignment compensation algorithm. The algorithm calculates dynamic interpolation coefficients based on the timestamp difference between thermal imaging frames and visible light frames and the target's motion speed, generating a time-synchronized multimodal data stream.

[0020] S2012. Perform radiation correction and noise suppression on thermal radiation characteristics, including: constructing a thermal radiation mapping model based on the ambient temperature gradient, and eliminating high-frequency interference through wavelet threshold denoising;

[0021] S2013. Perform illumination adaptive enhancement on visible light morphological features, use a multi-scale Retinex algorithm to separate the target reflection component and the illumination component, and use histogram equalization to improve the detail resolution of low-contrast areas.

[0022] By adopting the above technical solutions, the technical details of multimodal data preprocessing in step S201 are further clarified, including spatiotemporal alignment compensation, radiation correction and noise suppression, and adaptive illumination enhancement. The application of the above technologies can effectively reduce the acquisition delay and interference between sensors, thereby improving the quality and accuracy of the data; the spatiotemporal alignment compensation algorithm can eliminate the data inconsistency caused by the time difference between different sensors, and ensure the temporal synchronization of the data of the thermal imaging module and the visible light camera module, which is crucial for the accurate identification of dynamic targets; the radiation correction and noise suppression of thermal radiation characteristics eliminate high-frequency interference in the image through wavelet threshold denoising technology, further improving the clarity and reliability of the thermal imaging image; the adaptive illumination enhancement technology can automatically adjust the brightness and contrast of the image, and can still maintain the accuracy of target recognition under poor lighting conditions, especially in complex low-light environments, ensuring the stability of the visual system.

[0023] Further configuration is that the step S202 specifically includes:

[0024] S2021. Build a spatiotemporal pyramid fusion network, input the denoised thermal radiation feature map from step S2012 and the enhanced visible light morphology feature map from step S2013 into parallel convolution branches, respectively, and extract multi-scale spatial features through cascaded dilated convolution layers.

[0025] S2022. Introduce a bidirectional long short-term memory network to model temporal features, correlate the radiation change trend of the thermal imaging sequence with the morphological evolution trajectory of the visible light sequence, and generate a spatiotemporal joint encoding vector.

[0026] S2023. Nonlinearly weight the multi-scale spatial features extracted in step S2021 and the spatiotemporal joint coding vector generated in step S2022 through a dynamic weight matrix. The weight matrix is ​​adaptively adjusted based on the target motion state, and finally outputs a target feature vector with unified dimension.

[0027] By adopting the above technical solutions, a technical solution for deep fusion and feature extraction of multimodal data is proposed through a spatiotemporal pyramid fusion network, a bidirectional long short-term memory network, and a dynamic weight matrix. The spatiotemporal pyramid fusion network extracts spatial features of different scales through a dilated convolutional layer, which can effectively capture the different levels of features of the target and fuse them into a unified target feature vector. The introduction of a bidirectional long short-term memory network helps the model better capture the movement trends and changes of the target in the time series, enhancing the model's ability to perceive the dynamic changes of the target. The use of a dynamic weight matrix makes the fusion of different features more flexible, and can adjust the weights of each feature in real time according to the target's movement state, thereby generating a more accurate target feature vector. Ultimately, it can improve the target recognition and tracking capabilities in complex backgrounds and dynamic environments, and improve the accuracy and reliability of strikes.

[0028] Further configuration is that the step S203 specifically includes:

[0029] S2031. Embed a multimodal attention mechanism in the convolution module of the convolutional recurrent neural network. The multimodal attention mechanism generates spatial attention mask parameters based on the denoised thermal radiation feature map output from step S2012 and the enhanced visible light morphological feature map output from step S2013, and dynamically focuses on the local features of high-threat targets. Specifically, the multimodal attention mechanism includes:

[0030] S2031.1. Normalize the denoised thermal radiation feature map to generate a thermal radiation saliency map.

[0031] Calculate the edge gradient of the enhanced visible light morphological feature map to generate a morphological saliency map;

[0032] S2031.2. Perform pixel-by-pixel weighted fusion of the thermal radiation saliency map and the morphological saliency map, with the weight ratio dynamically adjusted based on the target motion state. Next, perform nonlinear activation on the fused weight map using a Sigmoid function to generate spatial attention mask parameters.

[0033] S2031.3. Perform a dot product operation on the spatial attention mask and the convolutional feature map of the convolutional recurrent neural network to enhance the response value of the high-threat area and suppress background interference;

[0034] S2032. Adopting a hierarchical feature pyramid structure, shallow detail features are integrated with deep semantic features across layers to enhance the model’s robustness in detecting small-scale targets.

[0035] S2033. Output the target motion trajectory through the time series prediction module, specifically: input the feature vector extracted by the convolutional recurrent neural network into the gated recurrent unit, combine it with the real-time posture data of the drone, predict the latitude and longitude coordinates and speed change curve of the target in the next 5 seconds, and generate a predicted trajectory.

[0036] By adopting the above technical solutions, the robustness of the model is further enhanced by proposing a multimodal attention mechanism and a hierarchical feature pyramid structure. The multimodal attention mechanism can dynamically focus on the high-threat areas of the target and perform weighted processing on key features, effectively improving the accuracy of target detection. The pixel-by-pixel weighted fusion of the thermal radiation saliency map and the morphological saliency map enables the model to adaptively adjust its attention under different states of the target, ensuring efficient recognition under high-speed movement or complex backgrounds. The hierarchical feature pyramid structure can effectively fuse shallow details with deep semantic features, allowing the model to maintain high detection accuracy when facing small-scale targets, thereby improving the robustness and adaptability of target detection. Ultimately, it can accurately identify and track high-speed moving targets in highly dynamic and complex battlefield environments.

[0037] Further configuration is that the step S204 specifically includes:

[0038] S2041: The feasibility of the predicted trajectory generated in step S2033 is verified based on the UAV dynamics model. If the trajectory curvature exceeds the maximum turning angle of the UAV or the altitude change rate exceeds the limit, the trajectory smoothing correction mechanism is triggered to generate an alternative path that meets the kinematic constraints through cubic spline interpolation. If the deviation between the corrected path and the original predicted path exceeds 10%, step S2033 is called again to generate a new predicted trajectory.

[0039] S2042: Use a Kalman filter to suppress the noise of the predicted trajectory, and dynamically compensate for the position error of the trajectory point in combination with the acceleration data output by the inertial measurement unit in step S100.

[0040] By adopting the above technical solution, the feasibility of the predicted trajectory is verified by combining the UAV dynamics model and the trajectory noise is suppressed by the Kalman filter, thereby enhancing the path optimization capability of the system; by performing physical constraint verification on the predicted trajectory, it is ensured that the generated path meets the dynamic constraints of the UAV, thereby avoiding flight errors caused by unreasonable paths; the Kalman filter can effectively eliminate noise in the trajectory prediction process, and dynamically compensate for the position error of the trajectory point in combination with the acceleration data output by the inertial measurement unit in step S100, thereby greatly improving the accuracy and reliability of path planning; this process ensures the stability and executability of the path, thereby improving the success rate of mission execution.

[0041] Further configuration is that the step S205 specifically includes:

[0042] S2051. When the mean Euclidean distance deviation between the optimized trajectory in step S2042 and the predicted trajectory output in step S2033 continues to exceed 2 meters, collect the multi-source perception data and deviation data from steps S201 to S204 as incremental training samples, and fine-tune the spatial attention mask parameters generated in step S2031 using a lightweight back-propagation algorithm.

[0043] S2052. If the mean Euclidean distance deviation between the optimized trajectory and the predicted trajectory in step S2051 continues to exceed 2 meters, and the deviation decrease rate is less than 5% after fine-tuning through the lightweight backpropagation algorithm, the parameter reset mechanism is triggered and the spatial attention mask parameters of step S2031 are reinitialized to the default weight distribution to restore the feature focusing capability.

[0044] By adopting the above technical solution, a method of fine-tuning the parameters of the spatial attention mask through incremental training samples is proposed, so that the convolutional recurrent neural network can be continuously optimized; by collecting multi-source perception data and deviation data from steps S201 to S204 in real time and fine-tuning through a lightweight backpropagation algorithm, the accuracy of target recognition and path planning are further improved; thus, experience can be continuously accumulated during the execution process, and changes in different environments can be adapted, thereby improving the system's adaptability and long-term stability.

[0045] Further configuration is that the step S300 specifically includes:

[0046] S301: Based on the target motion trajectory prediction result output in step S200, the UAV flight state data and environmental constraints are integrated to construct a dynamic flight state matrix; wherein:

[0047] The curvature radius and acceleration change rate of the target trajectory are extracted from the results of the trajectory smoothing correction mechanism in step S2041, and the matching degree is calculated with the maximum maneuverability of the UAV to generate a trajectory feasibility coefficient. The trajectory feasibility coefficient ranges from 0 to 1.

[0048] Based on the Kalman filter compensation result of step S2042, dynamic threat modeling is performed on the obstacle distribution to generate an obstacle threat probability distribution map;

[0049] The terrain elevation data is rasterized using a digital elevation model to generate a three-dimensional terrain constraint matrix;

[0050] S302: Call the pre-trained reinforcement learning model and initialize the Q-learning network parameters; based on the trajectory feasibility coefficient, obstacle threat probability distribution map, and three-dimensional terrain constraint matrix from step S301, design a multi-dimensional reward function, including a path safety reward item, a fuel efficiency reward item, and a trajectory feasibility reward item;

[0051] S303: Output the initial attack path through the reinforcement learning model, and verify the feasibility of the path in combination with the UAV dynamics model in step S2041;

[0052] S304: During the path execution process, the multi-source perception data updated in step S100 is received in real time, and the state space of the reinforcement learning model is updated through online incremental learning:

[0053] If the wind speed change exceeds a preset threshold, the wind resistance offset of the path node is calculated based on the trajectory feasibility bonus item in step S302. The offset calculation must satisfy the corrected trajectory feasibility coefficient ≥ 0.65.

[0054] If the predicted collision time between the new obstacle and the current path is less than the set time, emergency obstacle avoidance is initiated, a local detour path is generated based on the Dijkstra algorithm, and the maneuverability matching degree of the detour path is verified through the trajectory feasibility coefficient of step S301.

[0055] By adopting the above technical solution, the optimal attack path can be generated based on the reinforcement learning algorithm, and online incremental learning can be performed according to the actual environmental conditions; by receiving the multi-source perception data updated in step S100 in real time, the path can be adjusted according to real-time changes during flight to ensure the optimality and safety of the path; especially in the face of emergencies such as wind speed changes and the appearance of obstacles, the emergency obstacle avoidance mechanism can be automatically activated, and a local detour path can be generated through the Dijkstra algorithm. This adaptive mechanism can efficiently cope with complex battlefield environments.

[0056] Further configuration is that the step S303 specifically includes:

[0057] If the curvature of the initial attack path exceeds the maximum turning angle of the drone, the trajectory feasibility driven correction mechanism is triggered:

[0058] Dynamically adjust the Bezier curve interpolation parameters based on the trajectory feasibility coefficient of step S301, and increase the interpolation control point density by 20% for every 0.1 decrease in the feasibility coefficient;

[0059] Verify whether the trajectory feasibility coefficient of the corrected path is ≥ 0.7. If not, re-call the reinforcement learning model in step S302 to generate a new path;

[0060] If the fuel efficiency bonus weight of the revised path decreases by more than 20% and the trajectory feasibility coefficient is ≥ 0.8, the revised path is retained.

[0061] By adopting the above technical solution, dynamic adjustments can be made when the path deviates. The path correction method based on Bezier curve interpolation can effectively improve the stability and accuracy of the path, ensuring that the drone can successfully complete the mission. This mechanism can quickly generate a corrected path when responding to dynamically changing environments and find a balance between fuel efficiency and trajectory feasibility, greatly improving the flexibility of path optimization.

[0062] Further configuration is that the step S400 specifically includes:

[0063] S401: Based on the spatial attention mask parameters generated in step S2031, perform real-time analysis on the time-synchronized multimodal data stream in step S2011:

[0064] Extracting the high-temperature focus area of ​​the denoised thermal radiation feature map in step S2012 as a candidate thermal radiation locking frame;

[0065] Extracting the edge gradient peak area of ​​the visible light morphological feature map enhanced in step S2013 as a morphological locking candidate frame;

[0066] S402, adjusting the target lock logic based on the terminal parameters in the optimal attack path generated in step S300 and the terminal velocity of the trajectory predicted in step S2033;

[0067] S403. After the weapon system is activated, verify the strike effect through multimodal data:

[0068] S4031. Compare the thermal radiation characteristic diagrams before and after the strike to calculate the temperature difference;

[0069] S4032: Calculate the structural similarity index of the morphological features before and after the impact based on the enhanced visible light morphological feature image obtained in step S2013;

[0070] S4033, determine the attack effect:

[0071] If the temperature difference is greater than 50°C and the structural similarity index is less than 0.3, it is determined to be a valid strike, triggering the incremental training sample collection in step S2051;

[0072] If the temperature difference is ≤30°C or the structural similarity index is ≥0.5, it is determined that the attack has failed, and the trajectory feasibility driven correction mechanism in step S3003 is called to generate a secondary attack path;

[0073] S404: Input the effective attack data into the online adaptive learning module of step S205 to update the spatial attention mask parameters of step S2031.

[0074] By adopting the above-mentioned technical solution, the introduction of AI terminal guidance function, combined with the closed-loop data verification of the strike effect by multimodal sensors, ensures accurate strike of the target; by comparing the thermal radiation characteristic maps and visible light morphological characteristic maps before and after the strike, the strike effect can be evaluated in real time, and the incremental training sample collection is triggered according to the results to achieve continuous optimization of the convolutional recurrent neural network; ultimately, the strike accuracy is further improved, and at the same time, the self-learning ability can be further enhanced based on the feedback of the strike effect, making the UAV more intelligent and autonomous in actual combat.

[0075] In summary, the present invention has the following beneficial effects:

[0076] Through efficient multimodal data fusion, dynamic target tracking, reinforcement learning path optimization and adaptive adjustment mechanism, the present invention greatly improves the autonomous combat capability of UAVs in complex dynamic environments, realizes real-time target recognition, path planning and attack optimization of UAVs in complex battlefield environments, and improves combat efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 Schematic diagram of the process of the embodiment. DETAILED DESCRIPTION

[0078] The present invention will be further described in detail below with reference to the accompanying drawings.

[0079] As attached Figure 1 As shown;

[0080] This embodiment discloses an AI adaptive guidance optimization method for high-speed and long-distance UAVs, including the following steps:

[0081] S100, collecting multi-source perception data of the target area in real time through a multimodal sensor, the multimodal sensor including a thermal imaging module, a visible light camera module, an inertial measurement unit, and a global positioning system; wherein the thermal imaging module is used to obtain target thermal radiation characteristics, the visible light camera module is used to obtain target morphological characteristics, the inertial measurement unit is used to monitor the attitude angle and acceleration of the UAV in real time, and the global positioning system is used to locate the geographic coordinates of the UAV;

[0082] S200, performing dynamic target recognition and tracking on the multi-source sensory data collected in step S100 based on a deep neural network, including: performing spatiotemporal alignment and fusion of thermal radiation features and target morphological features to generate a target feature vector; performing multi-scale feature extraction on the target feature vector using a preset convolutional recurrent neural network, and outputting prediction results of the target's category, position, and motion trajectory;

[0083] S300, based on the target's trajectory prediction results, combined with the UAV's flight status data and environmental constraints, generate an optimal attack path in real time through a reinforcement learning algorithm, wherein the flight status data includes the UAV's remaining fuel, flight speed, and altitude, and the environmental constraints include obstacle distribution, wind speed, and terrain elevation;

[0084] S400: When the target enters the preset strike range, the AI ​​terminal guidance function is triggered, including: verifying the target lock status through the data closed loop of the multi-modal sensor.

[0085] In step S200, dynamic target recognition and tracking are performed on the multi-source perception data collected in step S100 based on a deep neural network, including the following sub-steps:

[0086] S201, performing multimodal preprocessing on raw data of the thermal imaging module and the visible light camera module;

[0087] S202, performing cross-modal feature fusion and target feature vector generation based on the output of step S201, specifically including:

[0088] S203: Based on the output of step S202, multi-scale feature extraction and trajectory prediction are implemented through a preset convolutional recurrent neural network, specifically including:

[0089] S204: Perform physical constraint verification and optimization of the trajectory prediction result based on the output of step S203, specifically including:

[0090] S205. Based on the outputs of step S203 and step S204, the convolutional recurrent neural network feature extraction strategy is updated through an online adaptive learning module.

[0091] Wherein, step S201 specifically includes:

[0092] S2011. Eliminate inter-sensor acquisition delays through a spatiotemporal alignment compensation algorithm. The algorithm calculates dynamic interpolation coefficients based on the timestamp difference between thermal imaging frames and visible light frames and the target's motion speed, generating a time-synchronized multimodal data stream.

[0093] S2012. Perform radiation correction and noise suppression on thermal radiation characteristics, including: constructing a thermal radiation mapping model based on the ambient temperature gradient, and eliminating high-frequency interference through wavelet threshold denoising;

[0094] S2013. Perform illumination adaptive enhancement on visible light morphological features, use a multi-scale Retinex algorithm to separate the target reflection component and the illumination component, and use histogram equalization to improve the detail resolution of low-contrast areas.

[0095] Wherein, step S202 specifically includes:

[0096] S2021. Build a spatiotemporal pyramid fusion network, input the denoised thermal radiation feature map from step S2012 and the enhanced visible light morphology feature map from step S2013 into parallel convolution branches, respectively, and extract multi-scale spatial features through cascaded dilated convolution layers.

[0097] S2022. Introduce a bidirectional long short-term memory network to model temporal features, correlate the radiation change trend of the thermal imaging sequence with the morphological evolution trajectory of the visible light sequence, and generate a spatiotemporal joint encoding vector.

[0098] S2023. Nonlinearly weight the multi-scale spatial features extracted in step S2021 and the spatiotemporal joint coding vector generated in step S2022 through a dynamic weight matrix. The weight matrix is ​​adaptively adjusted based on the target motion state, and finally outputs a target feature vector with unified dimension.

[0099] Wherein, step S203 specifically includes:

[0100] S2031. Embed a multimodal attention mechanism in the convolution module of the convolutional recurrent neural network. The multimodal attention mechanism generates spatial attention mask parameters based on the denoised thermal radiation feature map output from step S2012 and the enhanced visible light morphological feature map output from step S2013, and dynamically focuses on the local features of high-threat targets. Specifically, the multimodal attention mechanism includes:

[0101] S2031.1. Normalize the denoised thermal radiation feature map to generate a thermal radiation saliency map, where the weight of the high-temperature area is increased to 0.8-1.0, and the weight of the low-temperature area is limited to 0-0.2;

[0102] The edge gradient of the enhanced visible light morphological feature map is calculated to generate a morphological saliency map, in which the gradient amplitude of the target contour area is mapped to the weight value;

[0103] S2031.2. Perform pixel-by-pixel weighted fusion of the thermal radiation saliency map and the morphological saliency map. The weight ratio is dynamically adjusted based on the target's motion state: if the target is accelerating, the thermal radiation weight ratio is increased to 70%; if the target is stationary, the morphological weight ratio is increased to 80%. Next, perform nonlinear activation on the fused weight map using a sigmoid function to generate the spatial attention mask parameters.

[0104] S2031.3. Perform a dot product operation on the spatial attention mask and the convolutional feature map of the convolutional recurrent neural network to enhance the response value of the high-threat area and suppress background interference;

[0105] S2032. Adopting a hierarchical feature pyramid structure, shallow detail features are integrated with deep semantic features across layers to enhance the model’s robustness in detecting small-scale targets.

[0106] S2033. Output the target motion trajectory through the time series prediction module, specifically: input the feature vector extracted by the convolutional recurrent neural network into the gated recurrent unit, combine it with the real-time posture data of the drone, predict the latitude and longitude coordinates and speed change curve of the target in the next 5 seconds, and generate a predicted trajectory.

[0107] Wherein, step S204 specifically includes:

[0108] S2041: The feasibility of the predicted trajectory generated in step S2033 is verified based on the UAV dynamics model. If the trajectory curvature exceeds the maximum turning angle of the UAV or the altitude change rate exceeds the limit, the trajectory smoothing correction mechanism is triggered to generate an alternative path that meets the kinematic constraints through cubic spline interpolation. If the deviation between the corrected path and the original predicted path exceeds 10%, step S2033 is called again to generate a new predicted trajectory.

[0109] S2042: Use a Kalman filter to suppress the noise of the predicted trajectory, and dynamically compensate for the position error of the trajectory point in combination with the acceleration data output by the inertial measurement unit in step S100.

[0110] Wherein, step S205 specifically includes:

[0111] S2051. When the mean Euclidean distance deviation between the optimized trajectory in step S2042 and the predicted trajectory output in step S2033 continues to exceed 2 meters, collect the multi-source perception data and deviation data from steps S201 to S204 as incremental training samples, and fine-tune the spatial attention mask parameters generated in step S2031 using a lightweight back-propagation algorithm.

[0112] S2052. If the mean Euclidean distance deviation between the optimized trajectory and the predicted trajectory in step S2051 continues to exceed 2 meters, and the deviation decrease rate is less than 5% after fine-tuning through the lightweight backpropagation algorithm, the parameter reset mechanism is triggered and the spatial attention mask parameters of step S2031 are reinitialized to the default weight distribution to restore the feature focusing capability.

[0113] Wherein, step S300 specifically includes:

[0114] S301: Based on the target motion trajectory prediction result output in step S200, the UAV flight state data and environmental constraints are integrated to construct a dynamic flight state matrix; wherein:

[0115] The curvature radius and acceleration change rate of the target trajectory are extracted from the results of the trajectory smoothing correction mechanism in step S2041, and the matching degree is calculated with the maximum maneuverability of the UAV to generate a trajectory feasibility coefficient. The trajectory feasibility coefficient ranges from 0 to 1.

[0116] Based on the Kalman filter compensation result of step S2042, dynamic threat modeling is performed on the obstacle distribution to generate an obstacle threat probability distribution map;

[0117] The terrain elevation data is rasterized using a digital elevation model to generate a three-dimensional terrain constraint matrix;

[0118] S302: Call the pre-trained reinforcement learning model and initialize the Q-learning network parameters. Based on the trajectory feasibility coefficient, obstacle threat probability distribution map, and three-dimensional terrain constraint matrix from step S301, design a multi-dimensional reward function, including a path safety reward item, a fuel efficiency reward item, and a trajectory feasibility reward item. Specifically:

[0119] Path safety bonus: Calculates the minimum Euclidean distance between a path node and an obstacle, combined with the threat level from the obstacle threat probability distribution map. The greater the distance and the lower the threat probability, the higher the bonus value.

[0120] Fuel efficiency bonus: Dynamically adjusts the bonus weight based on the ratio of remaining fuel percentage to path length, where the path length is calculated using the actual elevation distance from the 3D terrain constraint matrix.

[0121] Trajectory matching reward: The trajectory feasibility coefficient from step S301 is used as a weighting factor, multiplied by the path safety reward and the fuel efficiency reward. A reward penalty mechanism is triggered when the feasibility coefficient is lower than 0.6.

[0122] S303: Output the initial attack path through the reinforcement learning model, and verify the feasibility of the path in combination with the UAV dynamics model in step S2041;

[0123] S304: During the path execution process, the multi-source perception data updated in step S100 is received in real time, and the state space of the reinforcement learning model is updated through online incremental learning:

[0124] If the wind speed change exceeds a preset threshold, set to ±5 m / s, the wind resistance offset of the path node is calculated based on the trajectory feasibility bonus item in step S302. The offset calculation must meet the requirement that the corrected trajectory feasibility coefficient is ≥ 0.65.

[0125] If the predicted collision time between the new obstacle and the current path is less than 3 seconds, emergency obstacle avoidance is initiated, a local detour path is generated based on the Dijkstra algorithm, and the maneuverability matching degree of the detour path is verified using the trajectory feasibility coefficient of step S301.

[0126] Wherein, step S303 specifically includes:

[0127] If the curvature of the initial attack path exceeds the maximum turning angle of the drone, the trajectory feasibility driven correction mechanism is triggered:

[0128] Dynamically adjust the Bezier curve interpolation parameters based on the trajectory feasibility coefficient of step S301, and increase the interpolation control point density by 20% for every 0.1 decrease in the feasibility coefficient;

[0129] Verify whether the trajectory feasibility coefficient of the corrected path is ≥ 0.7. If not, re-call the reinforcement learning model in step S302 to generate a new path;

[0130] If the fuel efficiency bonus weight of the revised path decreases by more than 20% and the trajectory feasibility coefficient is ≥ 0.8, the revised path is retained.

[0131] Wherein, step S400 specifically includes:

[0132] S401: Based on the spatial attention mask parameters generated in step S2031, perform real-time analysis on the time-synchronized multimodal data stream in step S2011:

[0133] Extracting the high-temperature focus area of ​​the denoised thermal radiation feature map in step S2012 as a candidate thermal radiation locking frame;

[0134] Extracting the edge gradient peak area of ​​the visible light morphological feature map enhanced in step S2013 as a morphological locking candidate frame;

[0135] S402: Based on the terminal parameters of the optimal attack path generated in step S300 and the terminal velocity of the trajectory predicted in step S2033, adjust the target locking logic:

[0136] Extract the terminal coordinates from the optimal attack path generated in step S300, and obtain the terminal height using the GPS data from step S100. If the terminal height is less than 100 meters and the remaining fuel input in step S300 is less than 20%, reduce the confidence threshold of the thermal radiation lock candidate box from 0.8 to 0.7;

[0137] If the terminal velocity of the predicted trajectory in step S2033 is greater than 10 m / s, the weapon pitch angle compensation is calculated based on the trajectory smoothing correction result in step S2041: compensation amount = velocity deviation × 0.05 rad / (m / s);

[0138] S403. After the weapon system is activated, verify the strike effect through multimodal data:

[0139] S4031. Compare the thermal radiation characteristic diagrams before and after the strike to calculate the temperature difference;

[0140] S4032: Calculate the structural similarity index of the morphological features before and after the impact based on the enhanced visible light morphological feature image obtained in step S2013;

[0141] S4033, determine the attack effect:

[0142] If the temperature difference is greater than 50°C and the structural similarity index is less than 0.3, it is determined to be a valid strike, triggering the incremental training sample collection in step S2051;

[0143] If the temperature difference is ≤30°C or the structural similarity index is ≥0.5, it is determined that the attack has failed, and the trajectory feasibility driven correction mechanism in step S3003 is called to generate a secondary attack path;

[0144] S404: Input the effective attack data into the online adaptive learning module of step S205 to update the spatial attention mask parameters of step S2031.

[0145] Example 1

[0146] This simulates a high-speed enemy vehicle moving at 120 km / h in a complex mountainous environment, requiring a drone to conduct covert tracking and precision strikes. Environmental conditions include a wind speed of 8 m / s, visibility of 5 km, and a terrain elevation standard deviation of 200 m.

[0147] In step S100, the thermal imaging module model is FLIR Tau2, with a frame rate of 30Hz, and detects an abnormal temperature of 85°C in the vehicle engine area; the visible light camera module model is Sony IMX677 4K HDR camera, which recognizes the vehicle's outline size of 4.8×2.1m and matches the satellite image database with a confidence level of 92%; the inertial measurement unit model is BMI0886 axis sensor.

[0148] In step 2011, based on the target speed of 32 m / s and the delay compensation of 12 ms, the dynamic interpolation coefficient is calculated to be 0.78, and a synchronous data stream is generated.

[0149] In step S2012, radiation compensation is performed based on the temperature gradient model, and the signal-to-noise ratio is improved by 18 dB after wavelet threshold denoising.

[0150] In step S2013, after processing by the multi-scale Retinex algorithm, the detail resolution of the low-contrast area is improved by 40%, and the SSIM index of the vehicle tire texture reaches 0.89.

[0151] In step S2021, the spatiotemporal pyramid fusion network uses three layers of dilated convolution, a bidirectional long short-term memory network with 256 hidden nodes, and a dynamic weight matrix generated by Softmax with a thermal radiation weight of 0.63 vs. a visible light weight of 0.37.

[0152] In step S203, the thermal radiation saliency map threshold is set to 75°C, the morphological gradient threshold is greater than 30 pixels / m, and the fused spatial mask increases the feature response value of the vehicle engine area by 3.2 times;

[0153] In step S204, a 5-second trajectory is predicted. After Kalman filtering, the position error is reduced from 3.2m to 1.1m. The original trajectory has a curvature radius of 85m, and an alternative path with a curvature radius of 135m is generated through cubic spline interpolation, with a deviation rate of 7.3%.

[0154] In step S300, the trajectory feasibility coefficient is 0.82, the obstacle threat map marks three air defense radars, and the terrain constraint matrix elevation standard deviation limits the flight altitude to ≥ 300m. The total length of the output path is 18.7km, and the fuel consumption accounts for 23%. After Bezier curve correction, the feasibility coefficient is increased to 0.89.

[0155] In step S400, the IoU between the thermal radiation candidate frame and the morphological frame is 0.91, with a locking confidence of 98%; the temperature difference after the impact ΔT is 112°C, and the structural similarity SSIM is 0.21, indicating that the judgment is valid;

[0156] Collect deviation data with an average of 1.4m, fine-tune the attention mask parameters, and use a weight update rate of 0.032.

[0157] Example 2

[0158] A high-fidelity forest fire digital twin system was built to simulate the spread of fire, the behavior of trapped people, and complex meteorological conditions to test the robustness of this method in extreme environments.

[0159] In step S100, the temperature field is generated by COMSOL Multiphysics, with the human body area set to 37±0.3°C, the fire core area>400°C, and the edge gradient 5°C / m;

[0160] Using the Unreal Engine 5 Nanite virtualization engine to simulate near-infrared imaging under thick smoke, the rendering resolution is 4096×2160 and the frame rate is 60Hz;

[0161] White noise is injected into the inertial measurement unit, with a standard deviation of angular velocity of 0.02 rad / s and acceleration noise of 0.05 m / s 2 ;

[0162] The GPS signal delay is adjustable, and the horizontal positioning error obeys N(0,0.12)m.

[0163] In step S200, multimodal data is hard-synchronized based on the simulation clock, with a delay jitter of <1ms. Adversarial samples are added during dynamic weight matrix training to adaptively switch the weights of thermal radiation and visible light at the edge of the fire scene, with a switching threshold temperature difference of 15°C.

[0164] In the simulated densely populated area, that is, >5 people / 100m 2 , enabling the regional suppression strategy to reduce the non-target human response value to 10% through spatial masking.

[0165] In step S204, the simulated drone model is a 6-DOF Quadrotor with a maximum thrust / weight ratio of 2.5:1. The original path pitch angle exceeds the limit, triggering cubic spline interpolation to generate a smooth path. After the path correction, the thrust fluctuation is reduced by 37% and the energy consumption is reduced by 12%.

[0166] In step S300, the fire threat map generates 100 sets of diffusion paths through Monte Carlo simulation, and the probability of fire in each grid is calculated; the terrain constraint imports DEM data, and areas with slopes greater than 30° are marked as no-fly zones;

[0167] 5,000 training iterations were deployed in the PyBullet physics engine. The state space included 32-dimensional parameters such as drone pose, fire probability, and personnel coordinates.

[0168] The optimal path avoids all areas with a fire probability greater than 0.3, has a total length of 2.4 km, and a delivery time of 3 minutes and 17 seconds. When simulating gust disturbances, the online replanning response time is less than 80ms, and the trajectory offset compensation error is less than 0.4m.

[0169] In step S400, the virtual laser radar verifies the landing point of the materials: the horizontal error is 0.32m, and the vertical error is 0.15m. The thermal imaging simulation module detects the temperature change of the trapped person: the target area after delivery has a ΔT of +0.8°C, and the first aid kit heating patch is activated.

[0170] If the delivery error is greater than 1m, incremental learning is triggered: simulation data is collected to fine-tune the forget gate weights of the trajectory prediction GRU. After the simulator records 100 trials, the average delivery accuracy is improved from 0.82m to 0.49m, with an optimization rate of 40.2%.

[0171] The above two embodiments verify the core technologies in the claims through specific scenarios: high-precision fusion of multimodal data, physical constraint optimization of dynamic trajectory prediction, environmental adaptability of reinforcement learning path planning, and online learning capability driven by closed-loop feedback; data show that high-precision positioning and real-time response in seconds can be achieved in complex environments.

[0172] This specific embodiment is merely an explanation of the present invention and is not intended to limit the present invention. After reading this specification, those skilled in the art may make non-creative modifications to this embodiment as needed. However, as long as such modifications are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A high-speed, long-distance AI adaptive UAV guidance optimization method, characterized in that: The following steps are included: S100, collecting multi-source perception data of the target area in real time through a multimodal sensor, the multimodal sensor including a thermal imaging module, a visible light camera module, an inertial measurement unit, and a global positioning system; wherein the thermal imaging module is used to obtain target thermal radiation characteristics, the visible light camera module is used to obtain target morphological characteristics, the inertial measurement unit is used to monitor the attitude angle and acceleration of the UAV in real time, and the global positioning system is used to locate the geographic coordinates of the UAV; S200, performing dynamic target recognition and tracking on the multi-source perception data collected in step S100 based on a deep neural network, including: performing spatiotemporal alignment and fusion of thermal radiation features and target morphological features to generate a target feature vector; performing multi-scale feature extraction on the target feature vector using a preset convolutional recurrent neural network, and outputting a target motion trajectory prediction result; S300: Based on the target's trajectory prediction results, combined with the drone's flight status data and environmental constraints, an optimal attack path is generated in real time through a reinforcement learning algorithm; S400: When the target enters the preset strike range, the AI ​​terminal guidance function is triggered, including: verifying the target lock status through the data closed loop of the multi-modal sensor.

2. The high-speed, long-distance AI adaptive UAV guidance optimization method according to claim 1, characterized in that: In step S200, dynamic target recognition and tracking are performed on the multi-source perception data collected in step S100 based on a deep neural network, including the following sub-steps: S201, performing multimodal preprocessing on raw data of the thermal imaging module and the visible light camera module; S202, performing cross-modal feature fusion and target feature vector generation based on the output of step S201, specifically including: S203: Based on the output of step S202, multi-scale feature extraction and trajectory prediction are implemented through a preset convolutional recurrent neural network, specifically including: S204: Perform physical constraint verification and optimization of the trajectory prediction result based on the output of step S203, specifically including: S205. Based on the outputs of step S203 and step S204, the convolutional recurrent neural network feature extraction strategy is updated through an online adaptive learning module.

3. The high-speed, long-distance AI adaptive UAV guidance optimization method according to claim 2, characterized in that: The step S201 specifically includes: S2011. Eliminate inter-sensor acquisition delays through a spatiotemporal alignment compensation algorithm. The algorithm calculates dynamic interpolation coefficients based on the timestamp difference between thermal imaging frames and visible light frames and the target's motion speed, generating a time-synchronized multimodal data stream. S2012. Perform radiation correction and noise suppression on thermal radiation characteristics, including: constructing a thermal radiation mapping model based on the ambient temperature gradient, and eliminating high-frequency interference through wavelet threshold denoising; S2013. Perform illumination adaptive enhancement on visible light morphological features, use a multi-scale Retinex algorithm to separate the target reflection component and the illumination component, and use histogram equalization to improve the detail resolution of low-contrast areas.

4. The high-speed, long-distance AI adaptive UAV guidance optimization method according to claim 3 is characterized by: The step S202 specifically includes: S2021. Build a spatiotemporal pyramid fusion network, input the denoised thermal radiation feature map from step S2012 and the enhanced visible light morphology feature map from step S2013 into parallel convolution branches, respectively, and extract multi-scale spatial features through cascaded dilated convolution layers. S2022. Introduce a bidirectional long short-term memory network to model temporal features, correlate the radiation change trend of the thermal imaging sequence with the morphological evolution trajectory of the visible light sequence, and generate a spatiotemporal joint encoding vector. S2023. Nonlinearly weight the multi-scale spatial features extracted in step S2021 and the spatiotemporal joint coding vector generated in step S2022 through a dynamic weight matrix. The weight matrix is ​​adaptively adjusted based on the target motion state, and finally outputs a target feature vector with unified dimension.

5. The high-speed, long-distance AI adaptive UAV guidance optimization method according to claim 4 is characterized by: The step S203 specifically includes: S2031. Embed a multimodal attention mechanism in the convolution module of the convolutional recurrent neural network. The multimodal attention mechanism generates spatial attention mask parameters based on the denoised thermal radiation feature map output from step S2012 and the enhanced visible light morphological feature map output from step S2013, and dynamically focuses on the local features of high-threat targets. Specifically, the multimodal attention mechanism includes: S2031.

1. Normalize the denoised thermal radiation feature map to generate a thermal radiation saliency map. Calculate the edge gradient of the enhanced visible light morphological feature map to generate a morphological saliency map; S2031.

2. Perform pixel-by-pixel weighted fusion of the thermal radiation saliency map and the morphological saliency map, with the weight ratio dynamically adjusted based on the target motion state. Next, perform nonlinear activation on the fused weight map using a Sigmoid function to generate spatial attention mask parameters. S2031.

3. Perform a dot product operation on the spatial attention mask and the convolutional feature map of the convolutional recurrent neural network to enhance the response value of the high-threat area and suppress background interference; S2032. Adopting a hierarchical feature pyramid structure, shallow detail features are integrated with deep semantic features across layers to enhance the model’s robustness in detecting small-scale targets. S2033. Output the target motion trajectory through the time series prediction module, specifically: input the feature vector extracted by the convolutional recurrent neural network into the gated recurrent unit, combine it with the real-time posture data of the drone, predict the latitude and longitude coordinates and speed change curve of the target in the next 5 seconds, and generate a predicted trajectory.

6. The high-speed, long-distance AI adaptive UAV guidance optimization method according to claim 5, characterized in that: The step S204 specifically includes: S2041: The feasibility of the predicted trajectory generated in step S2033 is verified based on the UAV dynamics model. If the trajectory curvature exceeds the maximum turning angle of the UAV or the altitude change rate exceeds the limit, the trajectory smoothing correction mechanism is triggered to generate an alternative path that meets the kinematic constraints through cubic spline interpolation. If the deviation between the corrected path and the original predicted path exceeds 10%, step S2033 is called again to generate a new predicted trajectory. S2042: Use a Kalman filter to suppress the noise of the predicted trajectory, and dynamically compensate for the position error of the trajectory point in combination with the acceleration data output by the inertial measurement unit in step S100.

7. The high-speed, long-distance AI adaptive UAV guidance optimization method according to claim 6, characterized in that: The step S205 specifically includes: S2051. When the mean Euclidean distance deviation between the optimized trajectory in step S2042 and the predicted trajectory output in step S2033 continues to exceed 2 meters, collect the multi-source perception data and deviation data from steps S201 to S204 as incremental training samples, and fine-tune the spatial attention mask parameters generated in step S2031 using a lightweight back-propagation algorithm. S2052. If the mean Euclidean distance deviation between the optimized trajectory and the predicted trajectory in step S2051 continues to exceed 2 meters, and the deviation decrease rate is less than 5% after fine-tuning through the lightweight backpropagation algorithm, the parameter reset mechanism is triggered and the spatial attention mask parameters of step S2031 are reinitialized to the default weight distribution to restore the feature focusing capability.

8. The high-speed, long-distance AI adaptive UAV guidance optimization method according to claim 6, characterized in that: The step S300 specifically includes: S301: Based on the target motion trajectory prediction result output in step S200, the UAV flight state data and environmental constraints are integrated to construct a dynamic flight state matrix; wherein: The curvature radius and acceleration change rate of the target trajectory are extracted from the results of the trajectory smoothing correction mechanism in step S2041, and the matching degree is calculated with the maximum maneuverability of the UAV to generate a trajectory feasibility coefficient. The trajectory feasibility coefficient ranges from 0 to 1. Based on the Kalman filter compensation result of step S2042, dynamic threat modeling is performed on the obstacle distribution to generate an obstacle threat probability distribution map; The terrain elevation data is rasterized using a digital elevation model to generate a three-dimensional terrain constraint matrix; S302: Call the pre-trained reinforcement learning model and initialize the Q-learning network parameters; based on the trajectory feasibility coefficient, obstacle threat probability distribution map, and three-dimensional terrain constraint matrix from step S301, design a multi-dimensional reward function, including a path safety reward item, a fuel efficiency reward item, and a trajectory feasibility reward item; S303: Output the initial attack path through the reinforcement learning model, and verify the feasibility of the path in combination with the UAV dynamics model in step S2041; S304: During the path execution process, the multi-source perception data updated in step S100 is received in real time, and the state space of the reinforcement learning model is updated through online incremental learning: If the wind speed change exceeds a preset threshold, the wind resistance offset of the path node is calculated based on the trajectory feasibility bonus item in step S302. The offset calculation must meet the requirement that the corrected trajectory feasibility coefficient is ≥ 0.

65. If the predicted collision time between the new obstacle and the current path is less than the set time, emergency obstacle avoidance is initiated, a local detour path is generated based on the Dijkstra algorithm, and the maneuverability matching degree of the detour path is verified through the trajectory feasibility coefficient of step S301.

9. The high-speed, long-distance AI adaptive UAV guidance optimization method according to claim 8, characterized in that: The step S303 specifically includes: If the curvature of the initial attack path exceeds the maximum turning angle of the drone, the trajectory feasibility driven correction mechanism is triggered: Dynamically adjust the Bezier curve interpolation parameters based on the trajectory feasibility coefficient of step S301, and increase the interpolation control point density by 20% for every 0.1 decrease in the feasibility coefficient; Verify whether the trajectory feasibility coefficient of the corrected path is ≥ 0.

7. If not, re-call the reinforcement learning model in step S302 to generate a new path; If the fuel efficiency bonus weight of the revised path decreases by more than 20% and the trajectory feasibility coefficient is ≥ 0.8, the revised path is retained.

10. The high-speed, long-distance AI adaptive UAV guidance optimization method according to claim 9, characterized in that: The step S400 specifically includes: S401: Based on the spatial attention mask parameters generated in step S2031, perform real-time analysis on the time-synchronized multimodal data stream in step S2011: Extracting the high-temperature focus area of ​​the denoised thermal radiation feature map in step S2012 as a candidate thermal radiation locking frame; Extracting the edge gradient peak area of ​​the visible light morphological feature map enhanced in step S2013 as a morphological locking candidate frame; S402, adjusting the target lock logic based on the terminal parameters in the optimal attack path generated in step S300 and the terminal velocity of the trajectory predicted in step S2033; S403. After the weapon system is activated, verify the strike effect through multimodal data: S4031. Compare the thermal radiation characteristic diagrams before and after the strike to calculate the temperature difference; S4032: Calculate the structural similarity index of the morphological features before and after the impact based on the enhanced visible light morphological feature image obtained in step S2013; S4033, determine the attack effect: If the temperature difference is greater than 50°C and the structural similarity index is less than 0.3, it is determined to be a valid strike, triggering the incremental training sample collection in step S2051; If the temperature difference is ≤30°C or the structural similarity index is ≥0.5, it is determined that the attack has failed, and the trajectory feasibility driven correction mechanism in step S3003 is called to generate a secondary attack path; S404: Input the effective attack data into the online adaptive learning module of step S205 to update the spatial attention mask parameters of step S2031.

Images