Multi-sensor data adaptive weight distribution method and system in dynamic environment

Abstract

The invention discloses a multi-sensor data adaptive weight distribution method and system in a dynamic environment, and the method comprises the steps: building a sensing degradation compensation spectrum through obtaining visual features and point cloud information in an environment perception process, and recognizing a signal distortion propagation path; generating an excitation anchor point position by adopting a reverse sequence tracking analysis technology, and establishing a weight fluctuation suppression mechanism to form an independent regulation domain; performing contribution degree reverse evaluation on the independent regulation domain, identifying a dominant sensor and a subordinate sensor, and executing weight cross modulation; constructing a cross influence node through fusion strategy cross analysis, and implementing adaptive parameter weight adjustment to generate a sensor delay spectrum; and finally, generating a time anchoring sequence by adopting sensing time reverse analysis, constructing a window activation matrix and carrying out time interleaving processing to form a multi-source sensing interleaving graph so as to extract the modulation weight of each sensor. According to the method, dynamic compensation and intelligent modulation of sensor performance can be realized, and the accuracy, robustness and adaptive capability of a multi-modal sensing system of the unmanned aerial vehicle are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle environment perception technology, in particular to a multi-sensor data adaptive weight distribution method and system in dynamic environment. BACKGROUND

[0002] When the unmanned aerial vehicle performs tasks in complex environment, it needs to rely on the cooperation of multiple sensors to obtain comprehensive environmental information. However, the performance of sensors will degrade over time, resulting in decreased perception accuracy and signal distortion, which seriously affects the environmental perception ability and flight safety of the unmanned aerial vehicle. Traditional sensor calibration methods mainly focus on static compensation for single sensors, and are difficult to handle the dynamic coupling relationship between multiple sensors and real-time performance changes.

[0003] In the prior art, multi-modal sensor data fusion mainly adopts a fixed weight distribution strategy, which cannot be dynamically adjusted according to the real-time state of the sensors. When some sensors experience performance degradation or signal distortion, the system cannot identify and adjust the weight configuration in a timely manner, resulting in reduced reliability of the fusion result. At the same time, traditional methods lack in-depth analysis of the signal distortion propagation path, and cannot solve the mutual interference problem between sensors from the root. SUMMARY

[0004] The present application provides a multi-sensor data adaptive weight distribution method and system in dynamic environment, which identifies the signal distortion propagation path by establishing a sensor degradation compensation spectrum, generates a sensor contribution index using reverse evaluation technology, realizes weight cross modulation of dominant sensors and subordinate sensors, and constructs a multi-source perception interleaving graph through time interleaving technology, finally extracts the optimal modulation weight of each sensor, improving the perception accuracy and system robustness of the unmanned aerial vehicle in complex environment.

[0005] The first aspect of the present application provides a multi-sensor data adaptive weight distribution method in dynamic environment, comprising the following steps: Obtain multi-modal sensor data in the environment perception process, the multi-modal sensor data including visual features and point cloud information, and perform feature analysis and processing on the multi-modal sensor data to construct a sensor degradation compensation spectrum; Identify the signal distortion propagation path based on the sensor degradation compensation spectrum, perform reverse tracking analysis on the signal distortion propagation path to generate excitation anchor points, establish a weight fluctuation suppression mechanism at the excitation anchor points to generate a fluctuation control signal, and use the fluctuation control signal to perform mutual exclusion distribution analysis to determine an independent adjustment domain; The contribution degree of the independent adjustment domain is inversely evaluated to generate an inverse index, dominant sensors and subordinate sensors are identified from the independent adjustment domain based on the inverse index, the dominant sensors and the subordinate sensors are weight cross-modulated to generate a modulated weight configuration, and a channel coding matrix is generated by re-encoding a perception channel based on the modulated weight configuration; The channel coding matrix is cross-analyzed based on a fusion strategy to construct a cross-influence node, adaptive parameter weight adjustment is implemented at the cross-influence node to generate a step adjustment sequence, and a sensor delay spectrum is generated by setting different adjustment delays for the step adjustment sequence; The sensor delay spectrum is inversely analyzed based on perception time to generate a time anchoring sequence, sensor activation windows are determined based on the time anchoring sequence to generate a window activation matrix, a multi-source perception interweaving graph is generated by time interweaving the channel coding matrix based on the window activation matrix, and modulation weight values of each sensor are extracted based on the multi-source perception interweaving graph.

[0006] The second aspect of the present application proposes a multi-sensor data adaptive weight distribution system in a dynamic environment, comprising: A data acquisition module is configured to acquire multi-modal sensing data in an environmental perception process, wherein the multi-modal sensing data comprises visual features and point cloud information, and the multi-modal sensing data is processed by feature analysis to construct a sensing degradation compensation spectrum. A path analysis module is configured to identify a signal distortion propagation path based on the sensing degradation compensation spectrum, inversely track and analyze the signal distortion propagation path to generate an excitation anchor point position, establish a weight fluctuation suppression mechanism at the excitation anchor point position to generate a fluctuation control signal, and determine independent adjustment domains by mutual exclusion distribution analysis using the fluctuation control signal. A weight evaluation module is configured to inversely evaluate the contribution degree of the independent adjustment domain to generate an inverse index, identify dominant sensors and subordinate sensors from the independent adjustment domain based on the inverse index, cross-modulate the dominant sensors and the subordinate sensors to generate a modulated weight configuration, and generate a channel coding matrix by re-encoding a perception channel based on the modulated weight configuration. A strategy adjustment module is configured to cross-analyze the channel coding matrix based on a fusion strategy to construct a cross-influence node, implement adaptive parameter weight adjustment at the cross-influence node to generate a step adjustment sequence, and generate a sensor delay spectrum by setting different adjustment delays for the step adjustment sequence. An execution fusion module is configured to inversely analyze the sensor delay spectrum based on perception time to generate a time anchoring sequence, determine sensor activation windows based on the time anchoring sequence to generate a window activation matrix, time interweave the channel coding matrix based on the window activation matrix to generate a multi-source perception interweaving graph, and extract modulation weight values of each sensor based on the multi-source perception interweaving graph.

[0007] The beneficial effects of the present application are embodied in the following points: 1. By constructing a sensing degradation compensation spectrum through feature analysis and processing of multi-modal sensing data, the spatial distribution and frequency domain characteristics of sensor performance degradation can be accurately identified, and the propagation path of signal distortion can be accurately identified from the compensation spectrum, realizing a technical breakthrough from passive acceptance of sensor degradation to active analysis and compensation, and improving the sensing ability and response speed of the system to changes in sensor performance. 2. The contribution degree reverse evaluation technology is adopted to generate a reverse index, which can objectively quantify the actual contribution degree of each sensor in cooperative sensing, and through identification of dominant sensors and subordinate sensors and execution of weight cross modulation, the potential contribution ability of subordinate sensors is stimulated, realizing optimization of sensing resources and improvement of overall system performance, and avoiding the problem of resource waste caused by traditional fixed weight allocation. 3. Through time interleaving technology, the window activation matrix and the channel coding matrix are deeply fused to form a spiral scheduling diagram and determine the weight deployment rhythm, and a multi-source sensing interleaving diagram is constructed, realizing the cooperative optimization of multi-sensor in time and space dimensions, not only improving the time sequence consistency and spatial coordination of sensing data, but also enhancing the adaptive ability and anti-interference ability of the system in complex environments through a dynamic weight modulation mechanism.

[0008] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 is a flow diagram of a dynamic environment multi-sensor data adaptive weight distribution method of the present application.

[0010] Figure 2 is a structural block diagram of a dynamic environment multi-sensor data adaptive weight distribution system of the present application. DETAILED DESCRIPTION

[0011] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0012] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly.

[0013] It should also be noted that when an element is referred to as being "fixed on" or "disposed on" another element, it may be directly on the other element or there may be an intermediate element. When an element is referred to as being "connected to" another element, it may be directly connected to the other element or there may be an intermediate element.

[0014] In addition, the descriptions of "first", "second", etc. in this application are for descriptive purposes only and should not be understood as indicating or implying their relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that they can be implemented by ordinary technicians in this field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0015] The technical solutions of the embodiments of this application are introduced below.

[0016] like Figure 1 As shown, an embodiment of the present invention provides a method for adaptively allocating weights of multi-sensor data in a dynamic environment, comprising the following steps S110 to S150: Step S110 , acquiring multimodal sensing data in the process of environmental perception, the multimodal sensing data including visual features and point cloud information, performing feature analysis on the multimodal sensing data to construct a sensing degradation compensation spectrum.

[0017] Specifically, multimodal sensor data is acquired during the environmental perception process. An integrated sensor array synchronously collects multi-source information during environmental perception. Multimodal sensor data primarily consists of two types of data: visual features captured by vision sensors and point cloud information acquired by lidar. Visual features, including image brightness distribution, edge gradients, texture features, and color space information, are acquired via a camera array using multispectral imaging technology covering the visible and near-infrared wavelengths, with a wavelength range of 400-1000 nm. The camera array is arranged in a stereoscopic configuration, with a baseline spacing set to a standard 60cm spacing to ensure accurate depth information acquisition, with a depth measurement accuracy of 0.1 meter. Point cloud information is acquired through lidar scanning, including 3D spatial coordinates, reflection intensity values, and timestamp information. The scanning frequency is set to 10Hz, with an angular resolution of 0.1 degrees. During data acquisition, GPS clock synchronization is used to ensure spatial correspondence between visual features and point cloud information at the same moment, with a time synchronization accuracy of 1 millisecond. Multimodal sensor data undergoes preprocessing, including Gaussian denoising filtering, coordinate transformation, and data registration, to form a unified world coordinate system.

[0018] In some embodiments, the feature analysis processing on the multi-modal sensor data constructs a sensor degradation compensation spectrum, including: generating a sensor mutual interference feature based on the visual features; identifying a sensor blind area boundary by performing a singular point capture on the sensor mutual interference feature; constructing a noise immune curve based on the sensor blind area boundary and the point cloud information; and establishing a sensor degradation compensation spectrum based on the noise immune curve.

[0019] The sensor mutual interference feature is generated based on the visual features. The acquired visual features are classified according to spatial distribution and frequency domain characteristics, and are organized into a feature vector matrix to represent the visual information of different regions. The sensor mutual interference feature reflects the characteristic performance of the mutual influence between different sensors. For example, when a multi-camera array captures the same target, the field of view overlap of adjacent cameras will produce redundant information, and the imaging differences of each camera under different lighting conditions will form a mutual interference pattern. The field of view overlap degree between sensors is calculated using the stereo vision configuration of the camera array. The correlation coefficient between visual features is calculated to identify the feature consistency of the field of view overlap region of adjacent sensors. The spatial distribution pattern of signal interference between sensors is detected by analyzing the spatial gradient change of visual features. In combination with the spectral dimension information obtained by multi-spectral imaging technology, the relationship between mutual interference strength and sensor distance, field of view angle, signal wavelength, etc. is established. The mutual interference strength is calculated at different wave bands respectively through comparative analysis of multi-spectral channels. The mutual interference harmonic components are identified by spectral analysis, which reflect the coupling effect between sensors. The pure signal and mutual interference signal components are separated from the mixed visual features. The spatial distribution of the mutual interference feature is calculated according to the geometric configuration of the sensor to generate a sensor mutual interference feature vector.

[0020] The sensor mutual interference characteristic is identified by singular point capture to recognize the sensor blind area boundary. The sensor mutual interference characteristic matrix is analyzed, and the main characteristic component is extracted by singular value decomposition method. The singular points with numerical anomaly in the sensor mutual interference characteristic distribution are searched, and the abnormal detection threshold is set as 3 times of the standard deviation of the average value. The jump degree S of singular value is calculated as S = σi / σi+1, where σi is the ith singular value and σi+1 is the i+1th singular value. When S > 2.0, it is determined as a significant singular point. Combined with the mutual interference strength function parameters, the regions with sensor spacing d less than 1 meter and field of view overlap angle θ higher than 50% are focused on, which are more likely to have singular points. The intensity of the 20-50Hz harmonic component in the mutual interference spectrum is used as the weight factor for singular point detection, and the sensitivity of singular point detection is improved by 20% at positions with high harmonic intensity. The detected singular points are connected by a cubic spline interpolation method to form a continuous boundary curve, and the interpolation node spacing is set to 0.2 meters. The singular points are subjected to median filtering to remove noise-induced pseudo singular points. According to the number density of singular points, the sensor blind area boundary is divided into high density area (density > 10 points / m2), medium density area (5-10 points / m2) and low density area (< 5 points / m2). The geometric characteristics of the sensor blind area boundary are fitted using a Bézier curve, and the curve fitting accuracy is 0.05 meters.

[0021] Based on the sensor blind area boundary and point cloud information, a noise immunity curve is constructed. The identified sensor blind area boundary is spatially aligned with the simultaneously acquired point cloud information to establish an accurate mapping relationship between the blind area and the three-dimensional space, with an alignment accuracy of 0.1 meters. The spatial point set corresponding to the blind area position is extracted from the point cloud information, and the density distribution of these points is analyzed. The normal area point cloud density is 100-200 points / m3, and the density in the blind area is reduced to 20-50 points / m3. The noise level of the point cloud information near the blind area boundary is calculated, and the standard deviation method is used to evaluate the variability of the point cloud density. A relationship model between noise intensity and distance from the blind area boundary is established, and the noise intensity within 1 meter from the boundary is 2-3 times that of the normal area. The periodic fluctuations of noise are identified by analyzing the time series of point cloud information in a 10-second time window, and it is found that the noise intensity shows a periodic change of 3-5 seconds. The signal and noise components are separated from the point cloud information using statistical filtering method, and the filtering threshold is set to 2 times the standard deviation. The signal-to-noise ratio at different spatial positions is calculated, and the normal area signal-to-noise ratio is 15-25 dB, while the blind area boundary decreases to 5-10 dB. According to the signal-to-noise ratio distribution and the sensor blind area boundary, a noise immunity curve is constructed, which describes the noise immunity of the system at different spatial positions. The noise immunity curve is fitted by a polynomial, and a 5th order polynomial is used to achieve a fitting accuracy of 99%.

[0022] The sensor degradation compensation spectrum is established based on the noise immunity curve. The noise immunity curve is converted into a frequency domain representation, and the frequency spectrum characteristics of the curve are analyzed using discrete Fourier transform to identify the main frequency components in the range of 0.1-1 Hz. By comparing and analyzing the spatial variation characteristics of the noise immunity curve, a sensor degradation function is established to describe the decay law of sensor performance over time and usage conditions, with a decay rate of about 2-5% per 1000 hours. By comparing the current noise immunity curve with the ideal reference curve, the degradation degree index D = ||C_current-C_ideal|| is calculated, where D is the degradation degree, C_current is the current measured noise immunity curve data, and C_ideal is the ideal reference curve data. According to the numerical values of the noise immunity curve at different spatial positions, the corresponding compensation function is designed, and the compensation amplitude is 1.2 times the degradation degree to achieve overcompensation effect. A three-dimensional compensation spectrum matrix is established, with dimensions of 100x100x50 corresponding to the x, y, and z coordinates of the space, and the matrix elements are the compensation coefficients of the corresponding positions. The numerical values of the compensation spectrum are optimized and adjusted by the least squares method to reduce the overall perception error of the system to 30% of the original. A three-layer processing structure is constructed for the compensation spectrum: the basic compensation layer handles large-scale degradation with a compensation range of ±50%; the fine compensation layer handles moderate degradation with a compensation range of ±20%; and the fine-tuning layer handles slight degradation with a compensation range of ±5%. The sensor degradation compensation spectrum is normalized to unify the compensation coefficient range to 0-2, forming a standard API interface for system calls.

[0023] In step S120, the signal distortion propagation path is identified based on the sensor degradation compensation spectrum, and the excitation anchor point position is generated by inverse order tracking analysis. A weight fluctuation suppression mechanism is established at the excitation anchor point position to generate a fluctuation control signal, and the independent adjustment domain is determined by mutual exclusion distribution analysis using the fluctuation control signal.

[0024] Specifically, the signal distortion propagation path is identified based on the sensor degradation compensation spectrum. The sensor degradation compensation spectrum of step S110 is analyzed in the frequency domain, and the characteristic abnormal mode in the compensation spectrum is extracted to trace the propagation trajectory of the distorted signal. The sensor degradation compensation spectrum is decomposed layer by layer according to a 0.1 Hz frequency interval, and abnormal peak values, frequency shifts, and amplitude mutations in the spectrum are identified as distortion characteristic markers. The abnormal peak value is represented by a local frequency energy exceeding 3 times the average value, the frequency shift is represented by a spectrum line position deviation exceeding ±5%, and the amplitude mutation corresponds to a spectrum density change rate exceeding 50% / second. For example, when the unmanned aerial vehicle flies over the electromagnetic interference source, the compensation spectrum of the forward-looking radar will have an abnormal peak value near 2.4 GHz. By analyzing the time difference of the appearance of the peak value between different sensors, the propagation path of the interference signal from the forward-looking sensor to the lateral sensor can be determined. The frequency spectrum anomaly index A_freq = |S_measured-S_reference| / S_reference is calculated, where A_freq represents the frequency spectrum anomaly, S_measured represents the measured spectrum data, and S_reference represents the ideal reference spectrum data. The abnormality threshold is set to 0.2, and the frequency points exceeding the threshold are marked as candidate positions of the distortion source. By analyzing the energy distribution difference of different frequency bands in the compensation spectrum, the propagation direction of the distortion energy is identified, and the energy density gradient points to the downstream direction of the distortion propagation. Each abnormal frequency component is mapped to a specific sensor node, and a frequency-sensor mapping relationship table is established. By analyzing the time delay of the compensation spectrum of adjacent sensors, the continuous path of the distortion propagation is tracked, and a complete signal distortion propagation path network is formed.

[0025] In some embodiments, the reverse-order tracking analysis of the signal distortion propagation path generates an excitation anchor point position, including: obtaining an abnormal aggregation point of the signal distortion propagation path; deriving a cross-sensor distortion feedback loop based on the abnormal aggregation point; generating a compensation preferred point by reverse-order tracking through the feedback loop; and determining the excitation anchor point position based on the compensation preferred point.

[0026] An abnormal aggregation point of signal distortion propagation path is obtained. In the identified signal distortion propagation path network, search for the key node position where the distortion energy appears abnormally concentrated. Calculate the distortion energy inflow and outflow balance of each node in the network, and mark it as an energy aggregation point when the inflow energy exceeds the outflow energy by more than 20%. Establish an energy balance detection mechanism, calculate the average energy flow using a 5-second sliding window, and the window update frequency is 10 Hz. The aggregation point shows a sudden jump or continuous drift phenomenon in the sensor data. When the forward radar encounters the aggregation point, the ranging accuracy deteriorates from ±2 cm to ±8 cm, the camera image appears local blur or color distortion, and the laser radar point cloud density decreases by 30-50% in the aggregation area. Merge the abnormal nodes with a spatial distance less than 1 meter and a distortion intensity difference less than 10% to form an abnormal aggregation point area. Automatically group the abnormal nodes by K-means clustering method, and the clustering number is set to 3-5, and the iterative convergence threshold is 0.01. According to the size of the aggregated distortion energy, the abnormal aggregation points are divided into high-level (energy>80%), medium-level (energy 40-80%) and low-level (energy<40%) three levels. High-level aggregation points will cause multiple sensors to be abnormal at the same time, medium-level aggregation points will affect 2-3 adjacent sensors, and low-level aggregation points will only cause slight decline in the performance of a single sensor. The stability of the aggregation point is judged by monitoring for 10 seconds continuously, and the aggregation point is considered stable if the duration is more than 5 seconds. Assign a unique identifier to each abnormal aggregation point, record its spatial coordinates, energy level, stability index and associated sensor list.

[0027] Based on the identified abnormal aggregation point, the closed-loop propagation path of the distortion signal among different sensor nodes is tracked. The propagation relationship with the surrounding sensors within a radius of 2 meters is analyzed from the abnormal aggregation point, and the directed connection of the propagation is determined by the energy flow direction. It is found that the distortion signal will propagate along the metal structure of the fuselage. The abnormal signal of the forward radar is transmitted to the lateral sensor through the fuselage frame, and then transmitted back to the processor through the cable bundle, finally affecting the next measurement of the forward radar. In the connection graph, the closed path starting from the abnormal aggregation point and finally returning to the point is found, and the path length is limited within 5 nodes. The total gain of each feedback loop G_loop = ΠG_segment is calculated, where G_loop represents the total gain of the loop, and G_segment represents the local gain of each propagation segment in the loop. A typical feedback loop is: forward radar -> fuselage vibration -> IMU jitter -> flight control compensation -> motor speed change -> fuselage vibration aggravation -> forward radar measurement deviation, forming an oscillation loop with a period of 5 seconds. The local gain is calculated by the ratio of the output distortion intensity to the input distortion intensity of the propagation segment, and the -6dB / octave attenuation of the propagation distance is considered. When the total gain of the loop is greater than or equal to 1.0, it is determined that the feedback loop may cause distortion amplification and needs to be focused on. Three levels are distinguished: sensor internal loop (<0.5 meters), sensor inter-local loop (0.5-2 meters), and global long loop (>2 meters). Identify key nodes in the feedback loop network with a degree greater than 3, which are the preferred locations for breaking the loop.

[0028] Compensation optimal points are generated by reverse tracing along feedback loops. Reverse tracing is performed along the identified distortion feedback loops to find the optimal compensation positions that can effectively break the loop or reduce the loop gain. Reverse tracing starts from the propagation segment with the largest loop gain and searches the control nodes in the opposite direction of signal propagation. Forward tracing is performed from the IMU with the most severe oscillation, and it is found that the main disturbance comes from the unbalanced rotation of the motor. By adding a damper to the motor support, 85% of the vibration transmission can be cut off. Adding a vibration isolation rubber pad at the connection between the fuselage and the forearm can block the propagation path of the vibration to the front radar. The reverse tracing algorithm is established, and the depth-first search strategy is adopted, with the search depth limited to twice the loop length. Considering the gain reduction effect, control power demand and implementation difficulty, the compensation efficiency index is calculated, with the efficiency value ranging from 0 to 1. The position with the best compensation effect is usually at the weak link of the loop, such as adding a magnetic ring at the cable joint to block electromagnetic coupling, and adding a shield to the sensor shell to reduce external interference. The globally optimal combination of compensation positions is searched among all the nodes of the feedback loop, and the genetic algorithm is used for optimization with a population size of 50. The optimization process considers the constraints of spatial position limitation (installation space > 10 cm x 10 cm), power capacity limitation (< 50 W) and response time limitation (< 0.1 second). The simultaneous influence of compensation positions on multiple feedback loops is analyzed, and positions that can effectively control 2-3 loops at the same time are selected. It is ensured that the selected compensation positions still have effective compensation ability within the range of ± 20% of the system parameters.

[0029] The position of the excitation anchor point is determined based on the compensation preferred point. The virtual coordinates of the compensation preferred point are converted into specific physical positions on the UAV platform, and the coordinate conversion accuracy reaches ±2 cm. For example, if the compensation preferred point is located at a virtual position 1.5 meters in front of the fuselage and 0.8 meters to the right, the excitation device is installed at the corresponding position of the front arm of the UAV. The excitation device is actually a small electromagnetic generator, with a size of 5 cm x 3 cm x 2 cm and a weight of 120 g. It exerts a compensation effect on nearby sensors by generating an electromagnetic field of a specific frequency. Considering the spatial geometric constraints and load distribution requirements of the UAV, the optimal installation position of the excitation device is determined within a radius of 20 cm around the compensation preferred point. The geometric constraints include avoiding the propeller sweep area, maintaining the center of gravity balance, and minimizing aerodynamic interference. The excitation anchor point is preferentially installed at the middle segment of the arm, where the vibration is smaller and does not affect the propeller airflow. It is fixed by carbon fiber buckles and does not change the appearance of the machine after installation. The spatial decay characteristics of the excitation signal are determined, with the signal strength decaying inversely proportional to the square of the distance, and the effective action distance is set to 2 meters. The collaborative layout of multiple excitation anchor points is optimized to achieve optimal coverage of the entire sensor network, with the coverage overlap controlled within the range of 10-20%. The front arm anchor point mainly compensates for the forward sensors, the rear arm anchor point is responsible for the rearward sensors, and the central anchor point handles the IMU and flight control system, forming a three-point coverage layout. The electromagnetic interference effect between excitation anchor points is analyzed, and a spatial isolation distance of at least 1 meter is set to avoid signal crosstalk.

[0030] A weight fluctuation suppression mechanism is established at the excitation anchor point position to generate fluctuation control signals. A three-level suppression mechanism consisting of fluctuation detection, suppression decision, and signal generation is established at the determined excitation anchor point position. Fluctuation detection continuously monitors the real-time change state of each sensor weight, identifies abnormal fluctuations by calculating the variation degree and trend of the weight sequence, and automatically triggers the suppression process when the weight deviates from the normal range by more than the threshold value. After receiving the trigger signal, the suppression decision makes pattern matching according to the amplitude, frequency, and duration characteristics of the weight fluctuation, matches the fast recovery mode for pulse-type fluctuations, matches the damping attenuation mode for oscillatory fluctuations, and matches the reverse compensation mode for drift-type fluctuations. Signal generation automatically retrieves the corresponding control parameter template and fills in the specific values according to the suppression mode determined by the decision, forming a standardized fluctuation control signal. The entire suppression mechanism works in a closed-loop feedback mode. After the control signal is sent, the weight change effect is continuously monitored. When the weight has not reached the expected stable state, the mechanism automatically adjusts the control strength and generates a revised signal. When the weight returns to the stable range, the mechanism stops signal generation and enters the monitoring standby state. The generated fluctuation control signal contains complete information such as target sensor identification, weight adjustment instruction, adjustment parameter, and execution priority. Through the excitation anchor point, the control instruction is sent to the corresponding sensor, realizing the automatic suppression of weight fluctuation and the stable operation of the sensor network.

[0031] In some embodiments, the determining independent regulation domains by utilizing the fluctuation control signals comprises: resolving the fluctuation control signals into dynamic weight sequences; extracting time offset and amplitude drift rate of the dynamic weight sequences; constructing a mutual exclusion node distribution map among sensors according to the time offset; and circumscribing independent regulation domains by using the amplitude drift rate and the mutual exclusion node distribution map.

[0032] The fluctuation control signals are resolved into dynamic weight sequences. The fluctuation control signals are sampled and converted, and the sampling frequency is set to 500 Hz. The continuous control signals are converted into discrete weight state sequences. A multi-channel synchronous sampling system is established, and the time synchronization accuracy is ±1 microsecond. The control signals are divided by using a sliding window technology, the window length is 0.1 second, and the overlap degree is 50%. The signal in each window corresponds to a weight state snapshot. A Hanning window function is used to reduce spectral leakage, and the attenuation coefficient of the window edge is 0.5. The instantaneous envelope and instantaneous phase of the control signals are extracted, and the envelope is detected by using Hilbert transform. The phase detection accuracy is ±1 degree. The dynamic weight sequence includes three basic fields of time index, weight value and confidence, and the data format is 32-bit floating point number. A data quality evaluation mechanism is established, and the confidence of the data with a signal-to-noise ratio greater than 20 dB is set to 1.0. The weight sequence is subjected to 3-order Butterworth low-pass filtering, and the cutoff frequency is set to 50 Hz. The sequence quality and analysis accuracy are improved.

[0033] The time offset and amplitude drift rate of the dynamic weight sequence are extracted. The time delay relationship R_xy(τ)=Σx(n)y(n+τ) between the weight sequences of different sensors is calculated by cross-correlation analysis, where R_xy(τ) represents the cross-correlation function value at lag τ time, and x(n) and y(n) represent the dynamic weight sequence data of two sensors, respectively. The time offset is determined as the lag time at which the correlation function reaches the maximum value, and the time delay estimation accuracy is improved to 0.1 millisecond by parabolic interpolation. The slope of the regression straight line is calculated as the drift rate index by performing least squares linear fitting on the weight sequence, and the drift rate unit is % / second. The mutation time nodes with a drift rate change of more than 50% in the weight sequence are identified, and the segmented characteristics of the drift characteristics are analyzed. The offset and drift characteristics are extracted on three time scales of 1 second, 10 seconds and 60 seconds, respectively, to provide multi-level feature description.

[0034] A mutual exclusion node distribution map is constructed between sensors according to the time sequence offset. A weighted network graph is formed with sensors as vertices and time sequence offsets as edge weights, and the edge weight value is equal to the absolute value of the time sequence offset. A network graph data structure is established, represented by an adjacency matrix, with a matrix dimension of N x N (N is the number of sensors). The threshold for judging mutual exclusion is set to 50 milliseconds, and when the absolute value of the time sequence offset between two sensors exceeds the threshold, it is marked as a mutual exclusion node pair. For example, when the weight sequence time sequence offset of the forward-looking radar and the downward-looking camera exceeds 50 milliseconds, it indicates that the adjustment responses of the two sensors are significantly inconsistent and should be classified into different adjustment domains. Spectral clustering analysis is performed on the sensor network, and the eigenvalue decomposition of the Laplacian matrix is used to identify sensor groups with similar time sequence characteristics. The number of clusters is set to 2-4, and the convergence threshold is 0.001. A mutual exclusion strength matrix is constructed, with matrix elements being the normalized result of the absolute value of the time sequence offset, with a value range of 0-1.

[0035] Independent adjustment domains are circled using amplitude drift rate and mutual exclusion node distribution map. Preliminary grouping is performed based on the similarity of the amplitude drift rate, and sensors with a drift rate difference of less than 0.1% / second are classified into the same adjustment candidate area. The topological structure of the mutual exclusion node distribution map is combined to determine the area boundary, and node pairs with a mutual exclusion strength greater than 0.7 are used as the boundary line of different adjustment domains. The K-means clustering algorithm is used to finally group the sensors, and the clustering features include two dimensions of drift rate and time sequence offset. The in-domain drift rate similarity threshold is set to 0.05% / second, and the inter-domain mutual exclusion strength threshold is set to 0.5, and the region that meets the double conditions is determined as an independent adjustment domain. Each independent adjustment domain contains 2-4 sensors, and the weight adjustment response time difference of the sensors within the domain is less than 20 milliseconds. The spatial boundary of the adjustment domain is calculated to determine the geometric center and coverage radius of each domain. The adjustment capacity of the sensors within the domain is integrated to calculate the adjustment domain capacity, with a capacity range of ±30% weight adjustment amplitude. Finally, 3-5 independent adjustment domains are circled, each domain has internal coordination and consistency, and external mutual independence adjustment characteristics, providing partitioned precise control capability for unmanned aerial vehicle sensor systems.

[0036] In step S130, the contribution degree of the independent adjustment domain is inversely evaluated to generate an inverse index, the dominant sensor and the subordinate sensor are identified from the independent adjustment domain based on the inverse index, the weight cross modulation is performed on the dominant sensor and the subordinate sensor to generate a modulated weight configuration, and the perception channel is re-encoded based on the modulated weight configuration to generate a channel encoding matrix.

[0037] Specifically, the contribution degree of the independent regulation domain is inversely evaluated to generate the reverse index. For the independent regulation domain, the actual contribution degree of each sensor to the overall sensing performance in the domain is tracked using the reverse analysis method. Starting from the overall output of the independent regulation domain, the specific contribution component of each sensor node is inversely decomposed. The contribution degree of a single sensor is calculated as C_i=(P_total-P_without_i) / P_total, where C_i represents the contribution degree of sensor i, P_total represents the total sensing performance of the independent regulation domain, and P_without_i represents the domain sensing performance after removing sensor i. For example, in the forward obstacle avoidance regulation domain of a UAV, if it contains a forward-looking camera, a radar, and an ultrasonic sensor, the actual contribution degree of each sensor can be quantified by removing each sensor one by one and measuring the degree of decline in obstacle avoidance performance. By reverse analysis, the interdependence between sensors is identified, and two types of direct contribution (performance provided by the sensor independently) and indirect contribution (performance generated through cooperation with other sensors) are distinguished. A multi-dimensional contribution evaluation framework is established, including signal quality contribution, coverage contribution, redundancy contribution, and stability contribution, with a score range of 0-1 for each dimension. The stability index S_i of the sensor is measured by calculating the coefficient of variation of the sensor output, and the stability index of the sensor with a coefficient of variation less than 0.1 is 1.0. The dependency index D_i of the sensor is evaluated by analyzing the correlation between the sensor and other sensors, and the dependency index of the sensor with strong independence is close to 0. The reverse index R_i is calculated as R_i=w1XC_i+w2XS_i+w3XD_i, where R_i represents the reverse index of sensor i, C_i represents the contribution degree, S_i represents the stability index, and D_i represents the dependency index, and the weight coefficients are w1=0.5, w2=0.3, and w3=0.2. The numerical range of the reverse index is determined by statistical analysis, and it is usually distributed between 0.2 and 0.9, forming a quantitative evaluation system for the importance of sensors.

[0038] The dominant sensor and the subordinate sensor are identified from the independent regulation domain based on the reverse index. The reverse index provides a quantitative standard for the role positioning and weight allocation of the sensor by comprehensively evaluating the actual contribution ability, working stability and independence degree of the sensor. The higher the index value, the more important the sensor is in the system. The generated reverse index is used as a classification standard to classify and grade the role of the sensor in each independent regulation domain. The sensor classification criterion is set, and when the reverse index of the sensor is more than 1.5 times the average value in the domain, it is classified as a dominant sensor. The dominant sensor identification threshold is equal to the average value of the reverse index plus 1.5 times the standard deviation, and the scientific nature of the classification is ensured by statistical methods. For example, in a regulation domain containing 5 sensors, if the reverse indexes are [0.8, 0.6, 0.5, 0.4, 0.3], the average value is 0.52, the standard deviation is 0.19, and the dominant sensor threshold is 0.81, so the sensor with a reverse index of 0.8 is identified as a dominant sensor. The sensors with reverse indexes lower than the average value in the domain are classified as subordinate sensors, and the remaining sensors are classified as medium sensors. The characteristic pattern of the dominant sensor is analyzed, which usually has obvious advantages in terms of sensing accuracy, response speed, coverage range, etc. The potential value of the subordinate sensor is analyzed, although the current contribution is low, but it still has compensation ability and auxiliary role under certain conditions. A sensor classification database is established to record the reverse index, classification result, advantage characteristics and application scene of each sensor. The role of the sensor is periodically re-evaluated, and the reverse index is recalculated and the classification is adjusted every 100 working periods.

[0039] In some embodiments, the weight cross modulation of the dominant sensor and the subordinate sensor generates a modulated weight configuration, including: weight fluctuation feature extraction of the dominant sensor to obtain weight evolution trend; mining potential contribution ability reserve from the subordinate sensor; coupling analysis of the weight evolution trend and the contribution ability reserve to obtain modulation benefit index; and performing weight cross modulation according to the modulation benefit index to generate a modulated weight configuration.

[0040] Weight fluctuation feature extraction is performed on the dominant sensor to obtain the weight evolution trend. For the identified dominant sensor, the time sequence variation law of the weight historical data is analyzed, and the basic trend feature of the weight change is extracted. A 10-second time window is used to slide analyze the weight sequence of the dominant sensor, and the weight mean and change rate in each window are calculated. A three-class system of weight change trend is established: upward trend (change rate > 0.1 / sec), downward trend (change rate <-0.1 / sec), and stable trend (absolute value of change rate ≤ 0.1 / sec). The fluctuation intensity of the weight change is calculated, which is quantified by the standard deviation of the weight sequence. It is considered to be high fluctuation when the standard deviation is greater than 0.05, and it is considered to be low fluctuation when it is less than 0.02. The autocorrelation characteristics of the weight sequence are analyzed, and the correlation coefficients of 1 second, 5 seconds, and 10 seconds delay are calculated to evaluate the periodicity characteristics of the weight change. The direction vector of the weight evolution is extracted, and the development direction of the weight sequence is fitted by linear regression. A positive slope indicates a growth trend, and a negative slope indicates a decay trend. The weight evolution trend parameters of each dominant sensor are recorded, including trend type, change rate, fluctuation intensity, and correlation coefficient.

[0041] Potential contribution ability reserves are mined from the subordinate sensors. The sensing ability and performance potential of the subordinate sensors that are not fully utilized under the current weight configuration are analyzed. The hardware performance reserve of the subordinate sensor is evaluated by comparing the theoretical maximum performance of the sensor with the current actual use performance to calculate the performance utilization rate. The signal processing ability reserve of the subordinate sensor is analyzed by measuring the ratio of the maximum data processing speed of the sensor to the current processing load. The difference between the theoretical maximum contribution degree and the current actual contribution degree of the sensor is calculated as a quantitative indicator of the potential contribution ability reserve. The ability reserve ratio is equal to the difference between the theoretical maximum performance and the current performance divided by the theoretical maximum performance. The higher the reserve ratio, the greater the unused potential. The potential ability is divided into four categories: signal acquisition reserve, data processing reserve, anti-interference reserve, and cooperative reserve. The reserve amount of each category is expressed as a percentage. For example, a subordinate ultrasonic sensor currently uses only 60% of the detection range and 40% of the sampling frequency, so its signal acquisition reserve is 40% and its data processing reserve is 60%. The performance of the subordinate sensor under different environmental conditions is analyzed to identify its superior ability in specific scenarios. Detailed ability reserve information of each subordinate sensor is recorded to form a complete ability profile containing four types of reserve amount, applicable scene, and activation condition.

[0042] For example, the coupling analysis of the weight evolution trend and the contribution ability reserve to obtain the modulation benefit index includes: coupling strength analysis of the weight evolution trend to determine alignment accuracy, the coupling strength includes change rate, jitter amplitude and reserve saturation; setting weight conversion parameters according to the alignment accuracy; redistributing the contribution ability reserve using the weight conversion parameters to generate a modulation benefit index.

[0043] The coupling strength analysis of the weight evolution trend determines the alignment accuracy. For the extracted weight evolution trend, multi-dimensional coupling strength analysis is performed to evaluate the matching degree between the trend characteristics and system requirements. The weight change rate characteristics are analyzed, and the change rate is standardized to the range of 0-1, and the faster the change rate, the more difficult the alignment. The weight jitter amplitude characteristics are measured, and the jitter degree is quantified by the coefficient of variation of the weight sequence, and the coefficient of variation greater than 0.1 is considered to be high jitter. The reserve saturation characteristics are evaluated, and the proportion of the current weight level relative to the theoretical maximum weight is calculated, and the higher the saturation, the smaller the adjustment space. The comprehensive coupling strength is calculated by the weighted average of the change rate, the jitter amplitude and the reserve saturation, and the weight distribution is 0.4, 0.3 and 0.3 respectively. According to the coupling strength, the matching degree is divided into high matching (coupling strength>0.7), medium matching (0.4≤coupling strength≤0.7) and low matching (coupling strength<0.4) three levels. The alignment accuracy is calculated by the product of the coupling strength and the direction cosine value, where the direction angle is the included angle between the weight evolution direction and the expected direction. The closer the alignment accuracy to 1 indicates that the matching degree between the weight evolution trend and the system requirements is higher, and it is more suitable for weight modulation operation.

[0044] The weight conversion parameter is set according to the alignment accuracy. Based on the calculated alignment accuracy, the adaptive adjustment mechanism of the weight conversion parameter is set. The alignment accuracy reflects the matching degree between the dominant sensor weight evolution trend and the actual system requirements. When the alignment accuracy is high, it means that the weight change direction of the dominant sensor is correct but the strength may be insufficient, and when the alignment accuracy is low, it means that the weight change direction is deviated and needs to be corrected. The weight conversion parameter is adjusted linearly based on the alignment accuracy, and the basic value is 0.1, and the adjustment range is 0.4, i.e. the parameter range is 0.1-0.5. When the alignment accuracy is 1.0, the weight conversion parameter is 0.5, which means that 50% of the weight transfer is allowed; when the alignment accuracy is 0, the conversion parameter is 0.1, which only allows 10% of the fine adjustment transfer. The selection of the weight conversion parameter directly affects the degree of release of the subordinate sensor capability, and the parameter is too large, which will cause the subordinate sensor to undertake too much task beyond the capability range, and the parameter is too small, which cannot fully tap the potential value of the subordinate sensor. The parameter boundary constraint is set, and the value range of the weight conversion parameter is limited to 0.05-0.6, which ensures the system stability.

[0045] The modulation benefit index is generated by reallocating the contribution capability reserves using the weight conversion parameter. The process of reallocating the contribution capability reserves is embodied as the dynamic upgrading of the subordinate sensor functional roles. The ultrasonic sensor, which originally only undertook the task of auxiliary monitoring, may undertake the key task of accurate ranging at close range after reallocation. The standby camera, which was originally in standby state, may be activated to participate in the main visual perception task. The reallocation process adopts a gradual strategy. First, part of the reserve capacity of the subordinate sensor is released for experimental task allocation. The allocation proportion is gradually expanded according to the execution effect, avoiding the system inadaptation caused by one-time large-scale adjustment. The modulation benefit index M = Σ (P × R_i × U_i) is calculated, where M represents the total modulation benefit index, P represents the weight conversion parameter, R_i represents the reserve amount of the i-th type of reserve, and U_i represents the utilization efficiency of the i-th type of reserve. The utilization efficiencies of various reserves are set as follows: the signal acquisition reserve efficiency is 0.8, the data processing reserve efficiency is 0.9, the anti-interference reserve efficiency is 0.6, and the cooperative reserve efficiency is 0.7. The reserve allocation optimization problem is solved by a linear programming method. The objective function is to maximize the modulation benefit index, and the constraint condition is that the total reserve amount does not exceed 100%.

[0046] The modulated weight configuration is generated by executing weight cross modulation according to the modulation benefit index. The calculated modulation benefit index is segmented and divided according to the benefit level. Different numerical intervals correspond to different intensity modulation strategies. When the modulation benefit index M > 0.8, high-intensity cross modulation is executed, and the dominant sensor transfers 30% weight to the subordinate sensor; when 0.5 < M ≤ 0.8, medium-intensity modulation is executed, and 15% weight is transferred; when M ≤ 0.5, low-intensity modulation is executed, and 5% weight is transferred. The weight transfer formula W_new = W_old × (1 ± λ × M) is set, where W_new is the modulated weight, W_old is the pre-modulation weight, λ is the transfer coefficient set to 0.5, and M is the modulation benefit index. A negative sign is used for the dominant sensor to reduce the weight, and a positive sign is used for the subordinate sensor to increase the weight. The weight configuration is gradually optimized by the iterative modulation method. The modulation benefit index is recalculated after each modulation. When the improvement amplitude of the index is less than 0.01 for three consecutive rounds, the modulation process is terminated. The safety boundary of weight modulation is set, and the weight variation amplitude of a single sensor does not exceed ± 40%, ensuring that the basic functions of the system are not affected. The generated modulated weight configuration realizes the optimal reallocation of the weight resources of the dominant sensor and the subordinate sensor, and improves the perception efficiency of the entire regulation domain.

[0047] The channel encoding matrix is generated based on the re-encoding of the perception channels with the modulated weight configuration. A mapping mechanism of the weight configuration to the channel encoding is established, and the modulated weight of each sensor corresponds to a specific encoding parameter setting. A three-layer encoding system is designed: the main channel encoding is responsible for the core perception task and allocates 70% of the encoding resources; the auxiliary channel encoding processes the secondary perception task and allocates 25% of the encoding resources; the redundant channel encoding provides backup guarantee and allocates 5% of the encoding resources. The element E_ij of the channel encoding matrix is calculated as E_ij = W_i × P_j × T_ij, where E_ij represents the encoding value of sensor i for perception task j, W_i represents the modulated weight of sensor i, P_j represents the priority of perception task j, and T_ij represents the adaptability of sensor i to task j. The priority allocation of the perception task is set as follows: the obstacle avoidance task priority is 1.0, the navigation task priority is 0.8, the monitoring task priority is 0.6, and the communication task priority is 0.4. The adaptability of the sensor to the task is evaluated by calculating the matching degree between the sensor technical specifications and the task requirements, and the adaptability of complete matching is 1.0. The finally generated channel encoding matrix has a dimension of N × M (N is the number of sensors and M is the number of perception tasks), and the matrix completely describes the optimal perception channel allocation scheme under the modulated weight configuration.

[0048] In step S140, a cross-influence node is constructed by cross-analysis of the fusion strategy of the channel encoding matrix, and a step adjustment sequence is generated by implementing adaptive parameter weight adjustment at the cross-influence node. The sensor delay spectrum is generated by setting different adjustment delays for the step adjustment sequence.

[0049] The fusion strategy cross analysis of the channel coding matrix is used to construct the cross-influence node. Different fusion strategy modes are identified from the row and column structure of the channel coding matrix, mainly including two basic strategies of parallel fusion and serial fusion. Parallel fusion strategy represents the synchronous activation mode of multiple column elements in the same row of the matrix, which is suitable for the scene of multiple sensors sensing the same target at the same time. Serial fusion strategy reflects the time sequence dependence relationship of cross-row elements in the matrix, which is suitable for the scene of sensors processing complex sensing tasks in sequence. For example, in the unmanned aerial vehicle obstacle avoidance system, the front-view camera and radar use parallel fusion strategy to detect obstacles at the same time, while the ultrasonic sensor uses serial fusion strategy to provide accurate ranging at close range. By analyzing the intersection of different fusion strategies in space position, the key position of strategy interaction is identified. The strength of strategy interaction is calculated, and the overlap degree of different strategy modes in the matrix is used to quantify the interaction strength. The interaction strength threshold is set to 0.6, and the positions exceeding the threshold are marked as potential cross-influence nodes. Through cluster analysis, the spatially adjacent potential nodes are merged to form the actual cross-influence node, and the cluster radius is set to 50% of the distance between sensors. The characteristic attributes of each cross-influence node are analyzed, including the number of sensors involved, the combination of strategy types, and the size of the influence range. The importance rating of the cross-influence node is established, and the score is based on the number of sensors involved and the complexity of the strategy. The score range is 1-10 points. Nodes with a score greater than 7 points are selected as key cross-influence nodes, which are given priority for parameter adjustment.

[0050] In some embodiments, the adaptive parameter weight adjustment in the cross-influence node generates a step adjustment sequence, including: detecting a sensor delay accumulation mode in the cross-influence node; obtaining an adaptive parameter through elastic window configuration based on the delay accumulation mode; generating an exponential decay adjustment interval according to the adaptive parameter; and constructing a step adjustment sequence using the exponential decay adjustment interval.

[0051] The cross-influence node detects the delay accumulation mode of the sensor. The time delay change of multiple sensor signals passing through the node is monitored in real time at each cross-influence node position. Four basic delay types of data are collected: signal transmission delay, data processing delay, queue waiting delay and synchronization coordination delay. Through continuous monitoring, it is found that the delay will superimpose effect at the cross-influence node, and the delay of multiple sensors will not simply add up when converging at the node, but will show a complex growth mode. Through time window calculation, the change trend of cumulative delay is divided into three basic types: stable accumulation mode shows slow linear growth of delay, fast accumulation mode shows sharp nonlinear growth of delay, and oscillation accumulation mode shows periodic fluctuation of delay in a fixed range. It is found that different accumulation modes correspond to different node working states. Stable accumulation mode usually appears when the node load is moderate, fast accumulation mode appears when the node is overloaded, and oscillation accumulation mode appears when the node processing capacity and load demand are critically matched. Analyze the periodic characteristics of delay accumulation, identify the repeated pattern of delay change, and find that the typical period is 30-60 seconds.

[0052] The adaptive parameters are obtained by the delay accumulation mode for elastic window configuration. According to the detected delay accumulation mode, an elastic time window configuration scheme is designed to realize adaptive response to different delay modes. The elastic window adopts a variable window length design, and the window base length is set to 5 seconds, which is dynamically adjusted according to the delay accumulation speed. For stable accumulation mode, the window length remains the base value; for fast accumulation mode, the window length is shortened to 2 seconds to improve response speed; for oscillation accumulation mode, the window length is extended to 8 seconds to smooth fluctuations. The window overlap rate is set to 50% to ensure the continuity between adjacent windows. Adaptive parameters are extracted through statistical analysis of window data, including response sensitivity (range 0.1-1.0), adjustment intensity (range 0.2-0.8) and stability factor (range 0.3-0.9) and other key parameters. Establish a parameter adaptive adjustment mechanism, increase the response sensitivity when the delay accumulation accelerates, and increase the stability factor when the system tends to be stable. The exponential moving average method is used to smooth the adaptive parameters, and the smoothing coefficient is set to 0.3 to avoid sharp changes in parameters. Ensure that the adaptive parameters change within the preset range to prevent parameter abnormalities from causing system instability.

[0053] The adaptive parameter is used to generate the exponentially decaying adjustment interval. The exponentially decaying strategy is adopted to ensure the reasonable distribution of adjustment frequency. The exponentially decaying strategy provides high-frequency adjustment at the beginning of system startup and gradually reduces the adjustment frequency when the system tends to be stable. The initial adjustment interval is set to 0.5 seconds, and the decay coefficient is determined according to the system stability factor. The higher the stability factor, the faster the decay. The incremental rule of the adjustment interval is calculated, and the interval increment ratio is controlled by the adjustment intensity parameter. The higher the adjustment intensity, the greater the increment amplitude. The basic increment rate is 20%, and the actual increment rate is 20% × adjustment intensity. The maximum interval is 10 seconds. The decay rate is controlled by the response sensitivity in the adaptive parameter. When the sensitivity is high, the decay is slowed down to maintain fast response. When the sensitivity is low, the decay is accelerated to save resources. The lower limit of the adjustment interval is set to 0.1 seconds and the upper limit is set to 30 seconds, ensuring that the system can respond quickly and will not be adjusted too much. When the system anomaly is detected, the decay process is reset, and high-frequency adjustment is restored to quickly stabilize the system.

[0054] The exponentially decaying adjustment interval is used to construct the step adjustment sequence. The exponentially decaying adjustment interval generated is used as a time reference to organize and arrange the weight adjustment operation sequence of each sensor. The step adjustment sequence adopts the adjustment mode of discrete time points, which divides the continuous adjustment process into a series of independent adjustment steps. The adjustment execution order is determined according to the importance level of the sensor: high importance sensors (related to obstacle avoidance) are adjusted first, medium importance sensors (related to navigation) are adjusted second, and low importance sensors (related to monitoring) are adjusted last. The adjustment interval is discretized into specific adjustment time, and an accurate time schedule is generated. The weight change amplitude of each adjustment step is set, and the single-step change amplitude is set to ±5% for high-frequency adjustment and up to ±15% for low-frequency adjustment. A conflict detection mechanism for adjustment operations is established. When the adjustment times of multiple sensors overlap, the conflict is solved through priority sorting and time fine-tuning.

[0055] The sensor delay spectrum is generated by setting different adjustment delays for the step adjustment sequence. According to the generated step adjustment sequence, the corresponding time delay is assigned to different types of adjustment operations to build the sensor delay spectrum. According to the importance and urgency of the adjustment, the adjustment operations are divided into four delay levels: immediate adjustment (delay 0 seconds, used for emergency obstacle avoidance), short delay adjustment (delay 1-3 seconds, used for path correction), medium delay adjustment (delay 5-10 seconds, used for parameter optimization), and long delay adjustment (delay 15-30 seconds, used for system maintenance). The calculation rules for delay time are established, and the specific delay values are determined by considering the adjustment complexity and the influence range. The adjustment complexity is evaluated by the number of sensors involved and the algorithm complexity, and the complexity score ranges from 1 to 5. The influence range is evaluated by the degree of influence of the adjustment operation on the system performance, and the influence score ranges from 1 to 3. The delay time calculation formula is set as the base delay time multiplied by the weighted coefficients of complexity and influence. The calculated delay time is classified and organized according to sensor type and adjustment type to form a two-dimensional delay distribution table. The data structure of the sensor delay spectrum is established, including fields such as sensor identifier, adjustment type, delay time, priority, and dependency relationship.

[0056] In step S150, the sensor delay spectrum is analyzed in reverse to generate a time anchor sequence, a window activation matrix is generated based on the time anchor sequence to determine the activation window of each sensor, and a multi-source perception interweaving graph is generated by time interleaving the channel coding matrix according to the window activation matrix, and the modulation weight of each sensor is extracted according to the multi-source perception interweaving graph.

[0057] In some embodiments, the sensor delay spectrum is analyzed in reverse to generate a time anchor sequence, including: using the sensor delay spectrum to distinguish between immediate response layer and delayed response layer; establishing a reverse anchor point according to the immediate response layer; using the reverse anchor point to compensate for the delayed response layer to determine the perception pre-time; and using the perception pre-time to construct a time anchor sequence.

[0058] The sensor delay spectrum is used to distinguish between immediate response layers and delayed response layers. The sensor delay spectrum is analyzed in layers, and the sensor response is divided into different time layers according to the length of the delay time. The delay threshold is set to 20 milliseconds, and when the sensor delay time is less than 20 milliseconds, it is classified as an immediate response layer, and when the delay time is greater than or equal to 20 milliseconds, it is classified as a delayed response layer. The data distribution of the sensor delay spectrum is analyzed to identify fast response sensor groups and slow response sensor groups. The immediate response layer mainly includes ultrasonic sensors (delay 5-15 milliseconds) and infrared sensors (delay 8-18 milliseconds) and other devices with faster response speeds. The delayed response layer mainly includes visual cameras (delay 30-80 milliseconds) and laser radars (delay 25-60 milliseconds) and other devices that require complex data processing. A sensor layer database is established to record the hierarchical attribution, delay characteristics and response capacity of each sensor. The division threshold of the immediate response layer and the delayed response layer is dynamically adjusted according to the system running state to adapt to the delay changes under different working conditions. The delay distribution characteristics of the sensors in each layer are analyzed, and the average delay value, maximum delay value and delay stability index are calculated.

[0059] The immediate response layer is used to establish a reverse anchor point. The identified sensor characteristics of the immediate response layer are used to determine a high-stability and strong-reference-value reverse time anchor point on the time axis. Through statistical analysis of the sensor delay in the immediate response layer, the sensor with the smallest delay fluctuation and the most stable response is identified as an anchor point candidate. The quality of the anchor point candidate sensor is evaluated, and the quality evaluation considers three dimensions: delay stability, response reliability and data accuracy. The sensor time point with the highest quality score in the immediate response layer is selected as the main reverse anchor point, and the sensor time point with the second highest score is selected as the auxiliary reverse anchor point. For example, in a UAV obstacle avoidance system, if the front ultrasonic sensor has a delay stability of about 10 milliseconds and a variation amplitude of less than ±2 milliseconds, it is selected as the main reverse anchor point. An anchor point quality evaluation mechanism is established to ensure the continuous effectiveness of the anchor point by continuously monitoring the delay characteristics of the anchor point sensor. When the quality of the existing anchor point decreases, a sensor with better quality is selected from the immediate response layer as a new anchor point. The anchor point switching condition is set, and when the main anchor point delay fluctuation exceeds ±5 milliseconds, the anchor point switching process is started.

[0060] The reverse anchor point is used to compensate the delay response layer and determine the perception pre-time. Based on the established reverse anchor point, the sensor delay in the delay response layer is compensated by time derivation calculation. The basic principle of compensation derivation is to use the anchor point time as the reference, subtract the estimated delay time of the delay response layer sensor, and get the perception pre-time that the sensor needs to start in advance. For example, if the main reverse anchor point time is 100 milliseconds, and the visual camera delay of the delay response layer is 50 milliseconds, the perception pre-time of the camera is 50 milliseconds, that is, it needs to start at 50 milliseconds to complete the processing at 100 milliseconds. Establish a delay estimation mechanism to predict the delay time of each sensor through statistical analysis of historical delay data. The moving average of the delay is calculated by the sliding window method, and the window length is set to 10 samples to improve the accuracy of delay estimation. A hierarchical compensation strategy is adopted for sensors with different delay levels in the delay response layer: short delay sensors (20-40 milliseconds) use simple linear compensation, medium delay sensors (40-70 milliseconds) use weighted average compensation, and long delay sensors (> 70 milliseconds) use trend prediction compensation. A compensation error control mechanism is established to continuously optimize the compensation parameters and strategies by comparing the actual delay with the predicted delay. Set the compensation accuracy target to control the calculation error of the perception pre-time within ±3 milliseconds.

[0061] The perception pre-time is used to construct the time anchor sequence. The calculated perception pre-time is arranged and organized in time sequence to form a complete time reference sequence. The sparse perception pre-time is processed by time interpolation, and the linear interpolation method is used to generate a dense time anchor sequence with an interpolation interval of 5 milliseconds. A sequence smoothing mechanism is established to eliminate abnormal values and mutation points in the perception pre-time using a three-point median filter method to ensure the continuity and stability of the time anchor sequence. The quality of the time anchor sequence is controlled through consistency check (difference between adjacent times less than 10 milliseconds), continuity check (no time jump), and reasonableness check (time increasing) to ensure the quality of the sequence. The storage format of the time anchor sequence is established, and the time sequence is stored in an array structure to support fast query and real-time access. The time anchor sequence is divided into three time periods: start phase, stable phase, and adjustment phase, according to the time characteristics. Set the sequence validity verification to verify the accuracy and practicality of the time anchor sequence by comparing with the actual running time.

[0062] The window activation matrix is generated based on the time-anchored sequence determination of each sensor activation window. The generated time-anchored sequence is used as a time reference to design personalized activation time windows and activation control strategies for each sensor. The time interval and distribution density of each anchor point in the time-anchored sequence are analyzed to determine the basic time parameters of the sensor activation window. The window length is determined according to the sensor type and task requirements: the window length of fast response sensors is 50 milliseconds, the window length of medium response sensors is 100 milliseconds, and the window length of slow response sensors is 200 milliseconds. Considering the coordination requirements between sensors, the window overlap strategy is set: the window overlap degree of complementary sensors (such as radar and camera) is 30%, and the window overlap degree of competitive sensors (such as multiple sensors of the same type) is 10%. For example, in the unmanned aerial vehicle landing scene, the downward-looking camera and height sensor are complementary sensors, and their activation windows overlap for 30 milliseconds to ensure time synchronization for data fusion. According to the characteristics of the time-anchored sequence, different sensors are assigned corresponding activation priorities: obstacle avoidance related sensors have a priority of 1 (highest), navigation related sensors have a priority of 2, and monitoring related sensors have a priority of 3. The window activation matrix is constructed, and the matrix dimension is N x T (N is the number of sensors, and T is the number of time segments), and the matrix elements represent the activation state of the sensor in the corresponding time segment using 0 and 1.

[0063] In some embodiments, the time interleaving of the channel encoding matrix according to the window activation matrix generates a multi-source perception interleaving graph, including: identifying the time coupling constraints between sensors through the window activation matrix; implementing interleaving arrangement processing using the time coupling constraints and the channel encoding matrix to form a spiral scheduling graph; determining the weight delivery rhythm according to the spiral scheduling graph; and forming a multi-source perception interleaving graph through the weight delivery rhythm.

[0064] The time coupling constraints between the sensors are identified through the window activation matrix. Based on the generated window activation matrix, the time overlapping patterns and mutual constraints of the sensor activation windows in the matrix are analyzed. The time coupling strength between the sensors is calculated, and the coupling degree is quantified by the ratio of the overlapping duration of the two activation windows to the length of the larger window. Through the row and column analysis of the window activation matrix, the sensor combinations that need to be activated synchronously and the sensor combinations that need to be activated staggered are identified. The synchronous activation combination usually contains sensors with strong complementarity, such as the combination of forward-looking cameras and forward-looking radars; the staggered activation combination contains sensors with resource competition, such as the combination of multiple vision processing modules. A time constraint rule library is established, including three types of constraints: mandatory synchronization constraint (coupling strength > 0.8), recommended synchronization constraint (coupling strength 0.5-0.8), and independent operation constraint (coupling strength < 0.5). The periodic patterns in the window activation matrix are analyzed to identify the periodic rules and non-periodic burst demands of sensor activation. For example, obstacle avoidance sensors exhibit high-frequency periodic activation (every 100 milliseconds), while state monitoring sensors exhibit low-frequency periodic activation (every 1 second). The constraint conflicts in the window activation matrix are identified, such as the simultaneous activation of multiple high-priority sensors at the same time. A conflict resolution mechanism is established to solve the constraint conflict problem through priority sorting and time fine-tuning.

[0065] The time coupling constraints and channel coding matrix are used to implement interleaving arrangement processing to form a spiral scheduling diagram. The identified time coupling constraints are used as arrangement rules to reorganize and arrange the channel coding matrix in the time dimension. The interleaving arrangement processing guides the time sequence arrangement of channel coding through time constraints, achieving optimal allocation and time distribution of coding resources. An interleaving arrangement strategy is designed to arrange the channel coding in a spiral pattern according to the time coupling constraint strength. The spiral arrangement strategy can evenly distribute coding resources to avoid resource conflicts and processing bottlenecks caused by time concentration. Spiral arrangement rules are established: the coding channels of strongly coupled sensors are arranged closely, and the coding channels of weakly coupled sensors are arranged dispersedly. For example, the forward-looking camera and the forward-looking radar are arranged adjacent to each other in the spiral diagram due to their strong coupling relationship, while the state monitoring sensors are dispersed in different positions of the spiral due to their strong independence. The spiral parameters are set through the hierarchical setting of time coupling constraints: the spiral tightness is 0.8 for strong constraints, 0.5 for medium constraints, and 0.2 for weak constraints. A multi-layer spiral structure is established to arrange coding channels of different priorities to different spiral levels, with high-priority channels located in the inner spiral and low-priority channels located in the outer spiral.

[0066] The weight release rhythm is determined according to the spiral scheduling diagram. Based on the formed spiral scheduling diagram, the distribution density and connection mode of the coded channel in the diagram are analyzed to determine the release time and release strength of the weight resource. The weight release rhythm controls the timing characteristics of the weight distribution, and realizes the optimal configuration and dynamic adjustment of the weight resource. The weight release rhythm is calculated by comprehensively considering the spiral density and channel priority. The high-density area corresponds to fast rhythm release, and the low-density area corresponds to slow rhythm release. The high-priority area increases the release frequency, and the low-priority area reduces the release frequency. The weight release is divided into three rhythm levels: fast rhythm release (release interval of 20 milliseconds, used for high-priority and strong-coupling channels), medium rhythm release (release interval of 50 milliseconds, used for medium-priority channels), and slow rhythm release (release interval of 100 milliseconds, used for low-priority and weak-coupling channels). A rhythm adaptive adjustment mechanism is established to dynamically adjust the release rhythm according to the system load and task demand. A time quantization method is used to convert the continuous spiral structure into a discrete time release sequence, and the quantization precision is set to 5 milliseconds. A coordination mechanism for the release rhythm is established to ensure the coordination between different rhythm levels and avoid rhythm conflicts.

[0067] A multi-source perception interweaving diagram is formed through the weight release rhythm. The determined weight release rhythm is used as a time reference, and the perception information of multiple sensors is interwoven and fused according to the rhythm pattern. The multi-source perception interweaving diagram represents the topology and timing relationship of data fusion between sensors, and provides a structured basis for the optimal value extraction. The network structure of the interweaving diagram is constructed, with sensors as nodes and connection relationships defined by the weight release rhythm as edges to form the network topology. The network topology provides the basic framework of the interweaving diagram, supports the representation and analysis of complex interweaving relationships. The interweaving strength is calculated, which quantifies the closeness of information fusion between sensors through the product of rhythm synchronization degree and weight amplitude. The rhythm synchronization degree reflects the degree of time synchronization between sensors, and the weight amplitude reflects the importance of fused data. A layered structure of the interweaving diagram is established, including sensor-level interweaving (data fusion within a single sensor), sensor group-level interweaving (fusion between functionally related sensors), and system-level interweaving (comprehensive fusion of the whole system). The layered structure provides different granularity views of interweaving, supporting multi-level fusion analysis and optimization. For example, in an unmanned aerial vehicle obstacle avoidance system, the forward-looking camera and the forward-looking radar form a sensor group-level interweaving, and multiple obstacle avoidance sensor groups and navigation sensor groups form a system-level interweaving.

[0068] The modulation weight values of each sensor are extracted according to the multi-source perception interweaving graph. The modulation weight values reflect the relative importance and contribution degree of the sensors in the overall system, and provide a quantitative basis for sensor resource allocation. The node importance analysis method is used to calculate the connection degree, influence range and contribution value of each sensor node in the multi-source perception interweaving graph. The connection degree reflects the association degree of the sensor with other sensors, the influence range reflects the propagation breadth of the sensor data, and the contribution value reflects the contribution size of the sensor to the system performance. Through the connection strength analysis of the interweaving graph, the influence and contribution of each sensor in the collaborative network are quantified, and the connection strength comprehensively considers factors such as the number of connections, connection weight and connection stability. The weight values are calculated and normalized to ensure that the sum of all sensor modulation weight values is 1, satisfying the basic constraint condition of weight allocation. Considering the constraints of system power consumption and computing resources, the extracted modulation weight values are re-optimized under resource constraints. The modulation weight values are divided into three levels according to importance: core weight (weight > 0.3, including key obstacle avoidance sensors), important weight (weight 0.1-0.3, including navigation auxiliary sensors) and auxiliary weight (weight < 0.1, including state monitoring sensors), realizing adaptive weight allocation of multi-sensor data.

[0069] In order to perform a kind of multi-sensor data adaptive weight allocation method in dynamic environment corresponding to the above-mentioned method embodiment, to realize corresponding function and technical effect. Referring to Figure 2 , Figure 2 The structure block diagram of a kind of multi-sensor data adaptive weight allocation system 200 in dynamic environment provided by the embodiment of the application is shown. For ease of illustration, only the part related to the present embodiment is shown, and the multi-sensor data adaptive weight allocation system 200 in dynamic environment provided by the embodiment of the application includes: The data acquisition module 201 is used to acquire multi-modal sensing data in the environmental perception process, and the multi-modal sensing data includes visual features and point cloud information. The multi-modal sensing data is analyzed and processed to construct a sensing degradation compensation spectrum. The path analysis module 202 is used to identify signal distortion propagation paths based on the sensing degradation compensation spectrum, perform reverse tracking analysis on the signal distortion propagation paths to generate excitation anchor point positions, establish a weight fluctuation suppression mechanism at the excitation anchor point positions to generate fluctuation control signals, and use the fluctuation control signals to perform mutual exclusion distribution analysis to determine independent adjustment domains. The weight evaluation module 203 is used to perform contribution degree reverse evaluation on the independent adjustment domains to generate reverse indexes, identify dominant sensors and subordinate sensors from the independent adjustment domains based on the reverse indexes, perform weight cross modulation on the dominant sensors and the subordinate sensors to generate modulated weight configurations, and re-encode perception channels based on the modulated weight configurations to generate channel encoding matrices. The policy adjustment module 204 is configured to perform fusion policy cross analysis on the channel coding matrix to construct a cross-influence node, perform adaptive parameter weight adjustment on the cross-influence node to generate a step adjustment sequence, set different adjustment delays for the step adjustment sequence to generate a sensor delay spectrum, and perform perception time reverse analysis on the sensor delay spectrum to generate a time anchor sequence. The execution fusion module 205 is configured to determine a window activation matrix based on the time anchor sequence, perform time interleaving on the channel coding matrix based on the window activation matrix to generate a multi-source perception interleaving graph, and extract a sensor modulation weight value based on the multi-source perception interleaving graph.

[0070] The above-described dynamic environment multi-sensor data adaptive weight distribution system 200 can implement a dynamic environment multi-sensor data adaptive weight distribution method. The optional items in the above-described method embodiments are also applicable to this embodiment, and will not be described in detail herein. The remaining content of the present embodiment can be referred to the content of the above-described method embodiments, and will not be described in detail herein.

[0071] The above only describes some or preferred embodiments of the present application, and neither the text nor the drawings can limit the scope of protection of the present application. Any equivalent structural transformation based on the content of the present application and the drawings, or direct / indirect application in other related technical fields is included in the scope of protection of the present application.

Claims

1. A method for adaptive weight distribution of multi-sensor data in dynamic environment, characterized in that, The method comprises the following steps: acquiring multi-modal sensing data in an environmental perception process, the multi-modal sensing data comprising visual features and point cloud information, and performing feature analysis and processing on the multi-modal sensing data to construct a sensing degradation compensation spectrum; identifying a signal distortion propagation path based on the sensing degradation compensation spectrum, performing reverse sequence tracking analysis on the signal distortion propagation path to generate an excitation anchor point position, establishing a weight fluctuation suppression mechanism at the excitation anchor point position to generate a fluctuation control signal, and performing mutual exclusion distribution analysis using the fluctuation control signal to determine an independent adjustment domain; performing reverse evaluation of contribution degree on the independent adjustment domain to generate a reverse index, identifying a dominant sensor and a subordinate sensor from the independent adjustment domain based on the reverse index, performing weight cross modulation on the dominant sensor and the subordinate sensor to generate a modulated weight configuration, and re-encoding a perception channel based on the modulated weight configuration to generate a channel encoding matrix; performing fusion strategy cross analysis on the channel encoding matrix to construct a cross-influence node, implementing adaptive parameter weight adjustment at the cross-influence node to generate a step adjustment sequence, and setting different adjustment delays for the step adjustment sequence to generate a sensor delay spectrum; performing perception time reverse analysis on the sensor delay spectrum to generate a time anchoring sequence, determining a sensor activation window based on the time anchoring sequence to generate a window activation matrix, and performing time interleaving on the channel encoding matrix according to the window activation matrix to generate a multi-source perception interleaving graph, and extracting a sensor modulation weight value according to the multi-source perception interleaving graph.

2. The method of claim 1, wherein, The method of performing feature analysis and processing on the multi-modal sensing data to construct a sensing degradation compensation spectrum comprises: generating a sensor mutual interference feature based on the visual feature; performing singular point capture on the sensor mutual interference feature to identify a sensing blind area boundary; constructing a noise immunity curve based on the sensing blind area boundary and the point cloud information; establishing a sensing degradation compensation spectrum based on the noise immunity curve.

3. The method of claim 1, wherein, The method of performing reverse sequence tracking analysis on the signal distortion propagation path to generate an excitation anchor point position comprises: acquiring an abnormal aggregation point of the signal distortion propagation path; deducing a cross-sensor distortion feedback loop based on the abnormal aggregation point; performing reverse sequence tracking through the feedback loop to generate a compensation preferred point; determining an excitation anchor point position based on the compensation preferred point.

4. The method of claim 1, wherein, The method of performing mutual exclusion distribution analysis using the fluctuation control signal to determine an independent adjustment domain comprises: resolving the fluctuation control signal into a dynamic weight sequence; extracting a time sequence offset and an amplitude drift rate of the dynamic weight sequence; constructing a mutual exclusion node distribution graph between sensors according to the time sequence offset; using the amplitude drift rate and the mutual exclusion node distribution graph to delineate an independent adjustment domain.

5. The method of claim 1, wherein, The method of performing weight cross modulation on the dominant sensor and the subordinate sensor to generate a modulated weight configuration comprises: extracting a weight fluctuation feature of the dominant sensor to obtain a weight evolution trend; mining potential contribution ability reserves from the subordinate sensor; performing coupling analysis on the weight evolution trend and the contribution ability reserves to obtain a modulation benefit index; performing weight cross modulation according to the modulation benefit index to generate a modulated weight configuration.

6. The method of claim 1, wherein, The adaptive parameter weight adjustment generating step adjustment sequence in the cross-influence node comprises: Detecting an inter-sensor delay accumulation mode in the cross-influence node; Based on the delay accumulation mode, an elastic window configuration is used to obtain an adaptive parameter; According to the adaptive parameter, an exponential decay adjustment interval is generated; The exponential decay adjustment interval is used to construct a step adjustment sequence.

7. The method of claim 1, wherein, The time anchoring sequence generated by the time reverse analysis of the sensor delay spectrum comprises: The sensor delay spectrum is used to distinguish an instant response layer and a delay response layer; According to the instant response layer, a reverse anchor point is established; The reverse anchor point is used to compensate and deduce the delay response layer to determine a pre-perception time; The pre-perception time is used to construct a time anchoring sequence.

8. The method of claim 1, wherein, The multi-source perception interleaving graph generated by the time interleaving of the channel coding matrix according to the window activation matrix comprises: The window activation matrix is used to identify the time coupling constraint between sensors; The time coupling constraint and the channel coding matrix are used to implement interleaving arrangement processing to form a spiral scheduling graph; According to the spiral scheduling graph, a weight delivery rhythm is determined; The weight delivery rhythm is used to form a multi-source perception interleaving graph.

9. The method of claim 5, wherein, The modulation benefit index obtained by coupling analysis of the weight evolution trend and the contribution ability reserve comprises: The alignment accuracy is determined by coupling strength analysis of the weight evolution trend, and the coupling strength includes change rate, jitter amplitude and reserve saturation; According to the alignment accuracy, a weight conversion parameter is set; The weight conversion parameter is used to redistribute the contribution ability reserve to generate a modulation benefit index.

10. A multi-sensor data adaptive weight distribution system in a dynamic environment, characterized by, Comprise: A data acquisition module is configured to acquire multi-modal sensing data in an environmental perception process, the multi-modal sensing data comprising visual features and point cloud information, and perform feature analysis processing on the multi-modal sensing data to construct a sensing degradation compensation spectrum; A path analysis module is configured to identify a signal distortion propagation path based on the sensing degradation compensation spectrum, perform reverse tracking analysis on the signal distortion propagation path to generate an excitation anchor point position, establish a weight fluctuation suppression mechanism at the excitation anchor point position to generate a fluctuation control signal, and perform mutual exclusion distribution analysis on the fluctuation control signal to determine an independent adjustment domain; A weight evaluation module is configured to perform contribution degree reverse evaluation on the independent adjustment domain to generate a reverse index, identify a dominant sensor and a subordinate sensor from the independent adjustment domain based on the reverse index, perform weight cross-modulation on the dominant sensor and the subordinate sensor to generate a modulated weight configuration, and re-encode a perception channel based on the modulated weight configuration to generate a channel coding matrix; A strategy adjustment module is configured to perform fusion strategy cross-analysis on the channel coding matrix to construct a cross-influence node, implement adaptive parameter weight adjustment in the cross-influence node to generate a step adjustment sequence, and set different adjustment delays for the step adjustment sequence to generate a sensor delay spectrum. The execution fusion module is used for performing time reverse analysis on the sensor delay spectrum to generate a time anchor sequence, determining each sensor activation window based on the time anchor sequence to generate a window activation matrix, performing time interleaving on the channel coding matrix according to the window activation matrix to generate a multi-source perception interleaving graph, and extracting each sensor modulation weight value according to the multi-source perception interleaving graph.

Images