A method and system for generating a road load spectrum by fusing multi-sensor data

Abstract

The application discloses a kind of road surface load spectrum generation method and system of fusion multi-sensor data, it is related to vehicle durability test technical field, method includes: time synchronization and noise suppression are carried out to multi-source sensing data, generate standardization multi-source data stream;Extract transient load event therein, construct operation feature vector;Operation feature vector is extracted and fused with space damage feature and time sequence load feature;Dynamic weighting calibration is carried out, and fine load spectrum segment is generated;Complete time series load spectrum is constructed, and authenticity is verified;Extract load cycle characteristics, generate target load spectrum for durability test;Dynamic adjustment sensor sampling strategy and data fusion weight, form adaptive optimization cycle.The application improves the generation precision and engineering applicability of road surface load spectrum by multi-source data fusion, knowledge graph guided feature extraction, multi-task deep learning evaluation, bayesian calibration, sequence generation verification and reinforcement learning optimization.

Description

Technical Field

[0001] This invention relates to the field of vehicle durability testing technology, and in particular to a method and system for generating road load spectra by fusing multi-sensor data. Background Technology

[0002] In the field of vehicle durability testing and structural fatigue analysis, road load spectrum, as a key data point reflecting the dynamic loads experienced by vehicles during actual driving, directly impacts the reliability of the overall vehicle structural design and the effectiveness of experimental verification through its generation accuracy. Traditional load spectrum generation methods primarily rely on data acquisition from a single type of sensor and load spectrum reconstruction through statistical processing. However, with the increasing complexity of test scenarios and the development of multi-source sensing technologies, existing methods have gradually revealed limitations in terms of data processing accuracy, feature extraction capabilities, and closed-loop optimization mechanisms.

[0003] Currently, in the process of generating road load spectra, due to the reliance on single sensor acquisition and traditional statistical processing methods, it is impossible to deeply fuse and time-synchronize multi-source heterogeneous sensor data such as wheel center six-component force, vehicle body vibration, vehicle speed and position, resulting in time sequence deviation and noise interference in the load signal. At the same time, existing technologies lack knowledge guidance and adaptive modeling capabilities based on road event characteristics, making it difficult to accurately extract transient load events from complex signals and construct multi-dimensional features. Furthermore, it is impossible to combine historical data and expert experience to dynamically calibrate the evaluation results and quantify the uncertainty.

[0004] Therefore, a method and system for generating road load spectra by fusing multi-sensor data is proposed to solve the above problems. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for generating road load spectra by fusing multi-sensor data, in order to solve the problems mentioned in the background above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for generating a road load spectrum by fusing multi-sensor data, the method comprising the following steps: S1. Use the Kalman filter algorithm to perform time synchronization and noise suppression on multi-source sensor data collected by multi-type sensor arrays to generate a standardized multi-source data stream. S2. Based on the road surface feature knowledge graph, transient load events in the standardized multi-source data stream are extracted using the wavelet packet decomposition algorithm, and an operational feature vector containing event type, intensity and timestamp is constructed. S3. Utilize a multi-task deep learning model to extract and fuse spatial damage features and temporal load features from the operational feature vector, and output the original load spectrum segmented evaluation results. S4. Combining historical load database and expert calibration rules, the original load spectrum segmentation evaluation results are dynamically weighted and calibrated using Bayesian inference algorithm to generate refined load spectrum segments containing confidence intervals. S5. Based on refined load spectrum segments, a complete time series load spectrum is constructed through a sequence generation model, and a generative adversarial network is used for authenticity verification to remove abnormal segments. S6. Based on the time series load spectrum, the load cycle characteristics are extracted by the rainflow counting method, and the target load spectrum for durability testing is generated by mapping the finite element model. S7. Based on the error distribution between the target load spectrum and the real road sampling data, the sensor sampling strategy and data fusion weights are dynamically adjusted through reinforcement learning algorithms to form an adaptive optimization loop based on load spectrum accuracy feedback.

[0007] Preferably, generating a standardized multi-source data stream in step S1 includes the following steps: S11. The triaxial force and triaxial torque at the connection between the wheel and the wheel center are synchronously collected at a preset sampling frequency by the wheel center six-component force sensor to form the original force load time series. S12. Use an accelerometer to collect the three-axis vibration time-domain signal of key points of the vehicle, and use the global positioning system to collect the vehicle's driving trajectory, instantaneous speed and latitude and longitude coordinates. S13. The Kalman filter algorithm is used to align the original force load time series, triaxial vibration time domain signal and vehicle driving trajectory in time, and to filter the inherent noise of the sensor to generate a standardized multi-source data stream.

[0008] Preferably, the construction of the operational feature vector in S2 includes the following steps: S21. Construct a road surface feature knowledge graph, where nodes represent road surface event types, including impact craters, joints, and speed bumps, and edges represent spatial relationships and typical load transfer patterns between events. S22. Wavelet packet decomposition algorithm is used to decompose the standardized multi-source data stream at multiple resolutions, extract the energy mutation features corresponding to road events, and identify transient load events. S23. Based on the intensity and timestamp of the transient load event, and combined with the road surface feature knowledge graph, perform event type matching to construct an operational feature vector containing event type, intensity, and timestamp.

[0009] Preferably, the output of the original load spectrum segmentation evaluation results in S3 includes the following steps: S31. Design a multi-task deep learning model, in which the convolutional neural network branch is used to extract the damage accumulation features of the load signal in the spatial dimension, and the long short-term memory network branch is used to learn the change pattern of the load sequence in the time dimension. S32. Input the operational feature vector into the multi-task deep learning model, and output the load spectrum segmentation and the corresponding damage contribution for different road events respectively. S33. Spatial damage features and temporal load features are spliced ​​together through a feature fusion layer to generate a multidimensional load feature map, and the original load spectrum segmented evaluation results are output using a fully connected layer.

[0010] Preferably, the generation of the refined load spectrum segment in S4 includes the following steps: S41. Establish an expert calibration rule base to store standard load spectrum fragments and corresponding confidence scores calibrated by domain experts under different road surface types and vehicle speed conditions. S42. Search the historical load database, extract historical load spectrum data similar to the current test scenario and their deviation trend from the true value, and use them as dynamic weighting factors. S43. The original load spectrum segmentation evaluation results are probabilistically fused with the expert calibration rule base using the Bayesian inference algorithm, and the posterior probability is calculated by combining the dynamic weighting factor to generate a refined load spectrum segment containing the confidence interval.

[0011] Preferably, the step S43, which uses a Bayesian inference algorithm to perform probability fusion and posterior probability calculation, includes the following steps: S431. Construct a load spectrum uncertainty quantification model and estimate the uncertainty distribution of the multi-task deep learning model output using the Monte Carlo dropout method; S432. Use a fuzzy logic system to process the qualitative rules in the expert calibration rule base, transform the qualitative evaluation into quantitative weights, and remove abnormal calibration records through a consistency check algorithm. S433. In the Bayesian inference process, the statistical prior of the historical payload database is introduced, the posterior probability is calculated based on the similarity of the current payload spectrum segment, the fusion weight is dynamically adjusted, and a refined payload spectrum segment is generated.

[0012] Preferably, the construction of the time-series load spectrum in S5 includes the following steps: S51. Based on the refined load spectrum segments, the gated recurrent unit network in the sequence generation model is used to learn the temporal dependencies between load segments and generate a complete time series load spectrum. S52. Construct a generative adversarial network, using the time series load spectrum as the generator output, and use a discriminator to distinguish between the generated load spectrum and the real load spectrum, and calculate the authenticity score. S53. When the authenticity score is lower than the preset threshold, the parameters of the sequence generation model are adjusted through the backpropagation algorithm, and abnormal segments in the generated time series load spectrum that do not conform to physical laws are removed.

[0013] Preferably, generating the target load spectrum for durability testing in step S6 includes the following steps: S61. Cyclic counting of the time series load spectrum is performed using the rainflow counting method to extract the cyclic matrix of load amplitude and mean. S62. Based on the cyclic matrix and the transfer function of the vehicle suspension system, calculate the equivalent load transmitted to the critical fatigue location of the vehicle body. S63. Based on the equivalent load, load spectrum blocks with preset damage levels are generated by mapping through the finite element model, which serve as the target load spectrum for durability testing.

[0014] Preferably, the adaptive optimization loop based on load spectrum accuracy feedback in step S7 includes the following steps: S71. Calculate the error distribution between the target load spectrum and the actual road sampling data in the time and frequency domains, and determine the key deviation interval for load spectrum reconstruction. S72. Construct a reinforcement learning agent, using the sampling frequency and fusion weights of a multi-type sensor array as the action space, and the negative value of the error distribution as the reward function. Iteratively optimize the sampling strategy and fusion weights through a deep Q-network. S73. The optimized sampling strategy and fusion weights are sent to the sensor array controller, and the updated configuration is executed in the next round of acquisition, forming an adaptive optimization loop based on load spectrum accuracy feedback.

[0015] A road load spectrum generation system that integrates multi-sensor data, the system comprising: The data acquisition and preprocessing end collects multi-source sensor data through multi-type sensor arrays and uses Kalman filtering algorithm for time synchronization and noise suppression to generate standardized multi-source data streams; The analysis and evaluation end is based on the road feature knowledge graph. It extracts transient load events and constructs operational feature vectors through wavelet packet decomposition algorithm. It uses a multi-task deep learning model to extract spatial damage features and temporal load features, and outputs the original load spectrum segmented evaluation results. Combining historical load database and expert calibration rules, it performs dynamic weighted calibration through Bayesian inference algorithm to generate refined load spectrum segments. The feedback and update end constructs a time-series load spectrum and uses a generative adversarial network for authenticity verification. Based on the time-series load spectrum, load cycle features are extracted using the rainflow counting method and combined with a finite element model to generate a target load spectrum. Based on the error distribution between the target load spectrum and real road sampling data, the sensor sampling strategy and data fusion weights are dynamically adjusted through a reinforcement learning algorithm to form an adaptive optimization cycle.

[0016] The present invention has the following beneficial effects: 1. In this invention, the Kalman filter algorithm is used at the data acquisition and preprocessing end to perform time synchronization and noise suppression on multi-source sensor data collected by multi-type sensor arrays, generating a standardized multi-source data stream. This solves the problem of low data fusion accuracy caused by different sampling frequencies, timing deviations and inherent noise interference of multi-source sensors, and improves the time consistency and signal quality of multi-source data.

[0017] 2. In this invention, transient load events are extracted by wavelet packet decomposition algorithm guided by road feature knowledge graph at the analysis and evaluation end. Spatial damage features and temporal load features are extracted by multi-task deep learning model. Combined with historical load database and expert calibration rules, dynamic weighted calibration is performed by Bayesian inference algorithm. This solves the problems of existing technology in accurately identifying transient load events from complex signals, lacking multi-dimensional feature fusion and uncertainty quantification capabilities, and improves the accuracy and confidence of load spectrum segmentation evaluation.

[0018] 3. In this invention, a time-series load spectrum is constructed using a sequence generation model through feedback and update, and its authenticity is verified by combining it with a generative adversarial network. A target load spectrum for durability testing is generated based on the rainflow counting method and the finite element model. Based on the error distribution between the target load spectrum and the real road sampling data, the sensor sampling strategy and data fusion weights are dynamically adjusted through a reinforcement learning algorithm. This solves the problem that traditional methods lack a load spectrum authenticity verification and sensor configuration closed-loop optimization mechanism, and improves the engineering applicability and adaptive optimization capability of the target load spectrum. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for generating a road load spectrum by fusing data from multiple sensors according to the present invention. Figure 2 This is a schematic diagram of the architecture of a road load spectrum generation system that integrates multi-sensor data according to the present invention. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] Example 1, please refer to Figures 1 to 2 As shown: A method for generating road load spectrum by fusing multi-sensor data, comprising the following steps: S1. The Kalman filter algorithm is used to perform time synchronization and noise suppression on the multi-source sensing data collected by multi-type sensor arrays to generate a standardized multi-source data stream.

[0022] S2. Based on the road surface feature knowledge graph, transient load events in the standardized multi-source data stream are extracted using the wavelet packet decomposition algorithm, and an operational feature vector containing event type, intensity, and timestamp is constructed.

[0023] S3. Utilize a multi-task deep learning model to extract and fuse spatial damage features and temporal load features from the operational feature vector, and output the original load spectrum segmented evaluation results.

[0024] S4. Combining historical load databases and expert calibration rules, the original load spectrum segmentation evaluation results are dynamically weighted and calibrated using a Bayesian inference algorithm to generate refined load spectrum segments containing confidence intervals.

[0025] S5. Based on refined load spectrum fragments, a complete time series load spectrum is constructed through a sequence generation model, and generative adversarial networks are used for authenticity verification to remove abnormal fragments.

[0026] S6. Based on the time series load spectrum, extract the load cycle characteristics using the rainflow counting method, and combine them with the finite element model to generate the target load spectrum for durability testing.

[0027] S7. Based on the error distribution between the target load spectrum and the real road sampling data, the sensor sampling strategy and data fusion weights are dynamically adjusted through reinforcement learning algorithms to form an adaptive optimization loop based on load spectrum accuracy feedback.

[0028] Generating a standardized multi-source data stream in S1 includes the following steps: S11. The triaxial force and triaxial torque at the connection between the wheel and the wheel center are synchronously collected at a preset sampling frequency using a six-component force sensor at the wheel center to form the original force load time series. In specific implementation: Wheel center six-component force sensors are installed at the connection points between each wheel and the wheel center of the vehicle, with a sampling frequency of 1000Hz, to simultaneously collect the forces acting on the wheels in three directions during driving. and torque in three directions The collected data is arranged in chronological order to form the original force load time series, which is represented as follows: , where t is the sampling time and T represents the transpose.

[0029] S12. Use an accelerometer to collect triaxial vibration time-domain signals of key points on the vehicle, and use a global positioning system to collect the vehicle's driving trajectory, instantaneous speed, and latitude and longitude coordinates. Specifically: Three-axis accelerometers are installed at key points on the vehicle frame to collect vibration acceleration signals along three axes during vehicle operation at the same sampling frequency as the wheel center six-component force sensor. , , Simultaneously, the vehicle's driving trajectory and instantaneous speed are collected in real time through the onboard GPS receiver. In addition to latitude and longitude coordinates, the above data is stored in association with timestamps.

[0030] S13. The Kalman filter algorithm is used to align the original force load time series, triaxial vibration time-domain signal, and vehicle trajectory on the time axis, and to filter the inherent noise of the sensors to generate a standardized multi-source data stream. Specifically: First, construct the state vector of the Kalman filter. ,in For the force load data of the kth sampling point, For vibration acceleration data, Instantaneous vehicle speed Let T denote the position coordinates; set the state transition matrix A and the observation matrix H, and update them through the time update equation. Predict the state and update the equation by measurement. The predicted values ​​are corrected, where For Kalman gain, Based on actual observations, the data from the three sensors are unified onto the same time axis through iterative calculations. At the same time, the inherent noise of the sensors is filtered out, and a standardized multi-source data stream with time alignment and noise suppression is generated.

[0031] Example 2: The construction of the operational feature vector in S2 includes the following steps: S21. Construct a road surface feature knowledge graph, where nodes represent road event types, including potholes, joints, and speed bumps, and edges represent spatial relationships and typical load transfer patterns between events. In specific implementation: Constructing a knowledge graph of road surface features The node set V contains road event types: craters, joints, speed bumps, manhole covers, and expansion joints. Each node... Related attribute vector ,in The mean and standard deviation of the load amplitude, The mean and standard deviation of the dominant frequency band, The mean and standard deviation of the event durations are obtained statistically from measured data over a distance of at least 2000 kilometers; the edge set E represents the spatial correlation and load transfer patterns between events, with each edge... Probability of association ,in Let i be the total number of times event i occurs. The number of times event j occurs immediately after event i within a specified distance range (e.g., 10-50 meters) is defined. The knowledge graph is stored in the Neo4j graph database, and spatial indexes are built for nodes to support fast proximity retrieval.

[0032] S22. Wavelet packet decomposition algorithm is used to perform multi-resolution decomposition on the standardized multi-source data stream, extracting energy mutation features corresponding to road events and identifying transient load events. Specifically, in implementation: The Daubechies5 wavelet (db5) is used as the basis function to normalize the vertical force load signal in the multi-source data stream. Perform 5-level wavelet packet decomposition. The recursive relationship of wavelet packet decomposition is as follows: ; ; Where: j is the number of decomposition levels m is the frequency band index , The coefficient of the k-th wavelet packet in the m-th frequency band of the j-th layer; These are the coefficients of the low-pass filter. These are the coefficients of the high-pass filter, and both are uniquely determined by the db5 wavelet basis function; Calculate the instantaneous energy of the 32 frequency bands obtained from the 5th layer decomposition. ,in Calculate the moving average of the energy mean for each frequency band within a sliding window centered at t (window length 0.1 seconds). with standard deviation When a frequency band exists satisfy When the event occurs, it is determined to be a transient load event, and the event timestamp is recorded. (Moment when energy first exceeds threshold), event intensity Dominant frequency band index and duration (Duration of time the energy exceeds the threshold).

[0033] S23. Based on the intensity and timestamp of the transient load event, and combined with the road feature knowledge graph, perform event type matching to construct an operational feature vector containing event type, intensity, and timestamp. In specific implementation: Event timestamp Align with GPS track, take The average GPS coordinates within the window are used as event space coordinates. And record the average vehicle speed within that window. ;by Search the knowledge graph with a center and a radius of 20 meters to obtain a set of candidate nodes. ; For each candidate node Calculate the feature matching score: ; in For the magnitude attribute of node v, For frequency band attributes ( (bandwidth) Duration attribute; spatial matching score ,in Let v be the spatial location of node v in the knowledge graph. This represents the standard deviation of GPS positioning error. Overall confidence level Among them, vehicle speed correction factor ,when ; ,when ; ,when Select the node with the largest C as the matching result. These are then marked as unrecognized events; finally, an operational feature vector is constructed: , where type is the event type code (unrecognized events are coded as 0).

[0034] The output of the original load spectrum segmentation evaluation results in S3 includes the following steps: S31. Design a multi-task deep learning model, where a convolutional neural network branch is used to extract the damage accumulation features of the load signal in the spatial dimension, and a long short-term memory network branch is used to learn the variation pattern of the load sequence in the temporal dimension. Specifically, construct a multi-task deep learning model containing two parallel branches and a multi-task output head. The convolutional neural network branch takes the generated operation feature vector sequence as input. ,in The vector is 8-dimensional, and K is the number of events within the sliding window (window length is 2 seconds, corresponding to approximately 10-20 events); three layers of 1-dimensional convolution are used, with kernel sizes of 3, 5, and 3, and channel numbers of 64, 128, and 64 respectively. The padding method is set to samepadding to maintain the sequence length. The l-th layer convolution operation is represented as: ; in For convolution kernel weights, For bias, The operation involves convolution, with ReLU as the activation function. Global average pooling follows the convolution, outputting a spatially impaired feature vector. ; Long Short-Term Memory (LSTM) branch: Used to learn the variation of the load sequence over time. The input is the original vertical force load time-series signal. The segment (1 second in length, corresponding to 1000 sampling points) uses a two-layer LSTM with a hidden layer dimension of 128 and a time step of 1000. The LSTM unit state update follows a standard gating mechanism, and the hidden state of the last time step is taken as the temporal feature vector. .

[0035] S32. Input the operational feature vector into the multi-task deep learning model, and output the load spectrum segments and corresponding damage contributions for different road events. In specific implementation: spatial damage characteristics With time series characteristics By concatenating the features, a fused feature vector is obtained. ; The model sets up N task output heads (N is the number of road event types, N=8 in this embodiment), each task head contains two fully connected layers (the first layer is 128-dimensional, and the second layer is determined according to the output dimension), and the i-th task head outputs: Load spectrum segment sequence ,in The length of a typical load spectrum segment for the i-th type of event (determined by statistical analysis of historical data; 200 points for impact crater type with T=0.2 seconds, and 500 points for speed bump type with T=0.5 seconds), each... Normalized load amplitude; Damage contribution This represents the percentage of this type of event's contribution to the overall fatigue damage. The damage contribution is normalized using the Softmax function. in The original output score of the i-th task head is given by... It is obtained through mapping via a fully connected layer.

[0036] S33. Spatial damage features and temporal load features are concatenated through a feature fusion layer to generate a multi-dimensional load feature map. The original load spectrum is then segmented and evaluated using a fully connected layer. Specifically, the model employs multi-task joint training, and the loss function is defined as: ; Where: MSE is the mean square error of the reconstructed load spectrum sequence, and KL is the KL divergence of the damage contribution distribution. , The weights are assigned to the task. Training data originates from historical load spectrum data collected in S1, which is then processed by S2 to form a labeled sample set containing at least 50,000 labeled event samples. This set is divided into training, validation, and test sets in an 8:1:1 ratio. The model uses the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64. The model converges after 50 training epochs. After training, the model outputs the original load spectrum segmented evaluation results. ; Each of them This is the load spectrum segment sequence corresponding to the i-th type of road surface event. Its contribution to damage.

[0037] The generation of refined load spectrum fragments in S4 includes the following steps: S41. Establish an expert calibration rule base to store standard load spectrum fragments and corresponding confidence scores calibrated by domain experts under different road surface types and vehicle speed conditions. In practice: Road surface types are coded as follows: asphalt road = 1, cement road = 2, gravel road = 3, pothole road = 4; vehicle speeds are divided into 10 intervals (20-120 km / h) at 10 km / h intervals, and the median of each interval is used as the vehicle speed identifier. For each combination of (road surface type, vehicle speed), at least three field experts are invited to each select a typical load spectrum segment and provide a confidence score. The confidence score of expert i for sample j is defined as: ,in The expert level coefficient is 0.9 for senior experts and 0.7 for intermediate experts. The sample typicality is determined by normalizing the inverse Mahalanobis distance between the sample and the mean of similar samples; for the same scenario, the arithmetic mean of all expert ratings is used as the final confidence level. The rule base is stored in key-value pairs: the key is... The value is a standard load spectrum fragment. With confidence level tuples.

[0038] S42. Retrieve historical load databases, extract historical load spectrum data similar to the current test scenario and their deviation trends from the true values, and use this as a dynamic weighting factor. Specifically, in implementation: Based on the road surface type of the current scene and vehicle speed Calculation and historical records Similarity distance: ; Where denominator 3 represents the maximum difference in road type codes, and denominator 100 represents the maximum value of the vehicle speed range, used for normalization. The top K=10 historical records with the smallest similarity distance are selected. For the i-th record, the model output is calculated. Compared with the actual collected values Deviation: ; The dynamic weighting factor is the mean deviation weighted by similarity: ; in Let be the similarity distance between the i-th record and the current scene. As a similarity weight, the smaller the distance, the greater the weight; S43. Using a Bayesian inference algorithm, the original load spectrum segmentation evaluation results are probabilistically fused with the expert calibration rule base, and the posterior probability is calculated by combining dynamic weighting factors to generate refined load spectrum segments containing confidence intervals, including the following steps: S431. Construct a load spectrum uncertainty quantification model, and estimate the uncertainty distribution of the multi-task deep learning model output using the Monte Carlo dropout method. Specifically: In the multi-task deep learning model built with S31, a dropout layer is added to each convolutional layer and fully connected layer with a dropout rate of 0.3. During the inference phase, the dropout activation is maintained, and the same input is forward-propagated for T=50 times to obtain the output set. Calculate the sample mean and variance: ; in The center estimate of the model output, As a measure of uncertainty, the likelihood function is set to a Gaussian distribution: ; S432. A fuzzy logic system is used to process the qualitative rules in the expert calibration rule base, transforming the qualitative evaluation into quantitative weights, and anomaly calibration records are removed through a consistency check algorithm. In specific implementation: Define fuzzy sets for vehicle speeds: low speed [0, 40] km / h, medium speed [30, 80] km / h, high speed [60, 120] km / h, and use a trapezoidal membership function. ; ; ; Fuzzy rule: Weighting coefficient when vehicle speed is low At medium speed At high speed The expert weights are: ; Consistency check: Calculate the mean expert scores for the same scenario. with standard deviation Eliminate those that meet the requirements Exception records; S433. In the Bayesian inference process, statistical priors from the historical payload database are introduced. Based on the similarity of the current payload spectrum fragments, the posterior probability is calculated, the fusion weights are dynamically adjusted, and refined payload spectrum fragments are generated. In specific implementation: Extract similar scenes from the historical database (similarity distance) to the current scene. Given M records, calculate the prior distribution: ; in Let be the similarity distance between the i-th record and the current scene. For scale parameters, The indicator function (takes 1 when the load spectrum S equals the i-th record), expert prior. Take the standard load spectrum fragment of the corresponding scenario from the S41 rule base, and use the confidence level. As weight; The comprehensive prior is a weighted fusion of statistical priors and expert priors: ; in That is, when there is sufficient historical data, the statistical prior weight is increased, and when there is insufficient historical data, the expert prior is used as the main factor. Bayesian posterior probability: in Given the likelihood function of S431, the posterior maximization is solved by enumerating candidate load spectrum segments (taking the top 50 segments with the highest similarity distance in the historical database and standard segments in the expert rule base): ; Using the loading amplitudes corresponding to the 2.5% and 97.5% quantiles of the posterior cumulative probability distribution as the 95% confidence interval, output... Its confidence interval is used as a refined load spectrum segment.

[0039] Example 3: The construction of the time series load spectrum in S5 includes the following steps: S51. Based on refined load spectrum segments, a gated recurrent unit network in the sequence generation model is used to learn the temporal dependencies between load segments, generating a complete time-series load spectrum. Specifically, the implementation involves: The refined load spectrum fragment generated by S43 is denoted as Each of them The payload amplitude vector has 100 sampling points (corresponding to 0.1 seconds), and L is the number of segments. A two-layer cascaded gated recurrent unit network is constructed as the generator, with 256 hidden units in each layer. The input at step t is the current segment. The output is the predicted value for the next segment. The output mapping relationship is as follows: ,in The hidden state output by GRU (dimension 256). To output the weight matrix, For bias; The model training employs a teacher-mandated strategy, with the loss function being the mean squared error. ; After training is complete, inference is performed using the last real fragment. Starting with the previous output, autoregression generates subsequent segments: each time, the previous output is used as the next input, and so on. Step 1 to obtain the complete sequence , where T is the expected total number of segments (in this embodiment, T=500).

[0040] S52. Construct a generative adversarial network (GAN), using the time-series load spectrum as the generator output, and employing a discriminator to distinguish between the generated load spectrum and the real load spectrum, calculating the authenticity score. In specific implementation: The discriminator takes a payload sequence of fixed length N=5000 sampling points (corresponding to 5 seconds) as input and employs a one-dimensional convolutional structure: three convolutional layers (kernel size 3 / 5 / 3, number of channels 64 / 128 / 256, stride 2), each followed by LeakyReLU activation, global average pooling, and then a fully connected layer to output a truth score. A higher score indicates a closer approximation to the true load spectrum. The S51 GRU model is used as the generator G, and adversarially trained against the discriminator D. The generator loss incorporates a reconstruction term to maintain its relevance to the input. ; The discriminator loss is: ; in To reconstruct the loss weights, For the mean squared error loss in S51, an alternating optimization strategy is adopted, updating the discriminator twice for each generator update, with learning rates set to... and Training continues until the discriminator loss converges.

[0041] S53. When the authenticity score is lower than a preset threshold, the parameters of the sequence generation model are adjusted using the backpropagation algorithm, and abnormal segments in the generated time series load spectrum that do not conform to physical laws are removed. Specifically: For the generated complete load spectrum The trained discriminator is used to calculate the authenticity score. ,like If the sample is deemed low-quality, the generation process of S51-S52 is repeated. To verify the physical laws, three constraints are defined: Amplitude constraints: Take 95% of the sensor's range; Rate of change constraint: Rate of change of load between adjacent points The determination is made by vehicle dynamics simulation; Energy continuity constraint: Energy within a sliding window (window length 100 points) coefficient of variation ; For any anomalous segment (more than 50 consecutive sampling points) that violates any constraint, the segment is filled by a weighted average of the corresponding positions of adjacent normal segments: ; in The waveforms on both sides of the abnormal segment are normal. The distance from the point to be filled to both sides is the value of the time series load spectrum. If the proportion of abnormal segments exceeds 10% of the total length, the complete sequence is regenerated, and the final output is the complete time series load spectrum that passes all tests.

[0042] Generating the target load spectrum for durability testing in S6 includes the following steps: S61. Perform cyclic counting on the time-series load spectrum using the rainflow counting method, and extract the cyclic matrix of load amplitude and mean. Specifically: The complete time-series load spectrum generated by S53 As input, the three-point rainflow counting method is used to extract load cycles: first, peak and valley values ​​are detected in the load sequence, and all extreme points are retained to form a simplified sequence. Take three consecutive points in sequence. Calculate the difference between adjacent amplitudes ;;when At that time, extract a complete loop, and its amplitude mean Then delete the point. and Then reconnect the remaining sequence; otherwise, move to the next set of three points and continue judging; after traversing all extreme points, divide the amplitude and mean into nA=32 levels and nM=16 levels at logarithmic intervals, and construct a cyclic matrix. ,element This indicates the number of cycles in which the amplitude falls at level i and the mean falls at level j.

[0043] S62. Based on the cyclic matrix and the transfer function of the vehicle suspension system, calculate the equivalent load transmitted to the critical fatigue locations of the vehicle body. In practice: Obtain the frequency domain transfer function H(f) of the vehicle suspension system. This function characterizes the amplitude attenuation and phase change from the wheel center load to the load at a critical fatigue location on the vehicle body (obtained through bench testing or vehicle dynamics simulation). For each load cycle extracted by S61, take its actual time-domain waveform segment. (Instead of a sinusoidal approximation), perform a Fast Fourier Transform to obtain the frequency domain load. The frequency domain load at the vehicle body position is calculated using the transfer function: ; The time-domain equivalent load is obtained through inverse Fourier transform. Extract its equivalent amplitude and mean ; Based on the material's SN curve (using the Basquin formula) Determine the equivalent load amplitude Fatigue life under action Where m is the material fatigue index (m=4 in this embodiment), C is the material constant, and the total damage of the (i, j)th cycle is calculated using the Miner linear cumulative damage criterion: ; Total damage is .

[0044] S63. Based on the equivalent load, a load spectrum block with a preset damage level is generated through finite element model mapping, which serves as the target load spectrum for durability testing. In specific implementation: A detailed finite element model of the critical fatigue locations of the vehicle body was established (mesh size 2-5mm, focusing on welds and stress concentration areas). Representative cycles with the largest amplitude (K=20) from the equivalent load cycles were selected and applied to the model for static analysis to obtain the stress generated at the critical locations during each cycle. For cycles that are not directly analyzed, stress values ​​are obtained through linear interpolation, and fatigue life is calculated based on the SN curve. Set target damage level (Corresponding to bench test half-life verification), calculate the number of cycles required to achieve the target damage: ; in The original load spectrum has a total number of cycles. All cycle types are sorted in descending order of equivalent amplitude, and their damage contributions are accumulated sequentially. When the accumulated damage contribution reaches 90% of the total damage, the previous value is truncated. This process involves a loop to generate the target payload spectral block. ; in For equivalent amplitude, For the equivalent mean, To determine the number of test cycles, the load spectrum block is output as a standard format file for bench durability testing.

[0045] The steps involved in forming an adaptive optimization loop based on load spectrum accuracy feedback in S7 are as follows: S71. Calculate the error distribution between the target load spectrum and the actual road sampling data in the time and frequency domains, and determine the key deviation interval for load spectrum reconstruction. In specific implementation: The target load spectrum generated by S63 Compared with real road data collected by S1 For comparison, the point-by-point absolute error is calculated in the time domain. Root mean square error Where N is the sequence length, and the power spectral density is calculated in the frequency domain. To avoid division by zero, the frequency domain relative error is defined as follows: The time period with a cumulative time domain error exceeding 80% and the frequency band with a cumulative frequency domain error exceeding 80% were selected as the critical deviation intervals. .

[0046] S72. Construct a reinforcement learning agent, using the sampling frequency and fusion weights of a multi-type sensor array as the action space, and the negative value of the error distribution as the reward function. Iteratively optimize the sampling strategy and fusion weights through a deep Q-network. Specifically: Construct a deep Q-network agent to optimize sensor configuration; the state space is defined as an 8-dimensional vector. The components are as follows: root mean square error in the time domain, mean maximum error in the frequency domain, percentage of deviation interval in the time domain, percentage of deviation interval in the frequency domain, peak error frequency, peak error amplitude, error change trend in the last 5 rounds, and current iteration round; The action space is a discrete combination of configuration parameters: six-component force sampling frequency. Accelerometer sampling frequency GPS sampling frequency fusion weight Take the satisfaction And each component The combination of 189 movements; The reward function takes into account both error and resource consumption. ; in This represents the average error across all frequency bands. This is the resource penalty coefficient; The Deep Q-Network employs a dual-network structure: both the main network and the target network have an 8-dimensional input layer, two fully connected layers each with 128 neurons, and an output layer with 189 dimensions. It uses an empirical replay pool (capacity 10,000), and the target network synchronizes every 100 steps. The optimizer is Adam, and the learning rate is... Discount factor ;use Strategy, Initial The reward decreases by 0.01 every 100 rounds to a minimum of 0.1, and the training continues for 500 rounds until the average reward converges.

[0047] S73. The optimized sampling strategy and fusion weights are sent to the sensor array controller, and the updated configuration is executed in the next round of acquisition, forming an adaptive optimization loop based on load spectrum accuracy feedback.

[0048] A road load spectrum generation system that integrates multi-sensor data includes: The data acquisition and preprocessing end collects multi-source sensor data through multi-type sensor arrays and uses Kalman filtering algorithm for time synchronization and noise suppression to generate a standardized multi-source data stream.

[0049] The analysis and evaluation end is based on the road feature knowledge graph. It extracts transient load events and constructs operational feature vectors through wavelet packet decomposition algorithm. It uses a multi-task deep learning model to extract spatial damage features and temporal load features, and outputs the original load spectrum segment evaluation results. Combining historical load database and expert calibration rules, it performs dynamic weighted calibration through Bayesian inference algorithm to generate refined load spectrum segments.

[0050] The feedback and update end constructs a time-series load spectrum and uses a generative adversarial network for authenticity verification. Based on the time-series load spectrum, load cycle features are extracted using the rainflow counting method and combined with a finite element model to generate a target load spectrum. Based on the error distribution between the target load spectrum and real road sampling data, the sensor sampling strategy and data fusion weights are dynamically adjusted through a reinforcement learning algorithm to form an adaptive optimization cycle.

[0051] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for generating a road load spectrum by fusing multi-sensor data, characterized in that, The method includes the following steps: S1. Use the Kalman filter algorithm to perform time synchronization and noise suppression on multi-source sensor data collected by multi-type sensor arrays to generate a standardized multi-source data stream. S2. Based on the road surface feature knowledge graph, transient load events in the standardized multi-source data stream are extracted using the wavelet packet decomposition algorithm, and an operational feature vector containing event type, intensity and timestamp is constructed. S3. Utilize a multi-task deep learning model to extract and fuse spatial damage features and temporal load features from the operational feature vector, and output the original load spectrum segmented evaluation results. S4. Combining historical load database and expert calibration rules, the original load spectrum segmentation evaluation results are dynamically weighted and calibrated using Bayesian inference algorithm to generate refined load spectrum segments containing confidence intervals. S5. Based on refined load spectrum segments, a complete time series load spectrum is constructed through a sequence generation model, and a generative adversarial network is used for authenticity verification to remove abnormal segments. S6. Based on the time series load spectrum, the load cycle characteristics are extracted by the rainflow counting method, and the target load spectrum for durability testing is generated by mapping the finite element model. S7. Based on the error distribution between the target load spectrum and the real road sampling data, the sensor sampling strategy and data fusion weights are dynamically adjusted through reinforcement learning algorithms to form an adaptive optimization loop based on load spectrum accuracy feedback.

2. The method for generating a road load spectrum by fusing multi-sensor data according to claim 1, characterized in that, The process of generating a standardized multi-source data stream in S1 includes the following steps: S11. The triaxial force and triaxial torque at the connection between the wheel and the wheel center are synchronously collected at a preset sampling frequency by the wheel center six-component force sensor to form the original force load time series. S12. Use an accelerometer to collect the three-axis vibration time-domain signal of key points of the vehicle, and use the global positioning system to collect the vehicle's driving trajectory, instantaneous speed and latitude and longitude coordinates. S13. The Kalman filter algorithm is used to align the original force load time series, triaxial vibration time domain signal and vehicle driving trajectory in time, and to filter the inherent noise of the sensor to generate a standardized multi-source data stream.

3. The method for generating a road load spectrum by fusing multi-sensor data according to claim 1, characterized in that, The construction of the feature vector in S2 includes the following steps: S21. Construct a road surface feature knowledge graph, where nodes represent road surface event types, including impact craters, joints, and speed bumps, and edges represent spatial relationships and typical load transfer patterns between events. S22. Wavelet packet decomposition algorithm is used to decompose the standardized multi-source data stream at multiple resolutions, extract the energy mutation features corresponding to road events, and identify transient load events. S23. Based on the intensity and timestamp of the transient load event, and combined with the road surface feature knowledge graph, perform event type matching to construct an operational feature vector containing event type, intensity, and timestamp.

4. The method for generating a road load spectrum by fusing multi-sensor data according to claim 1, characterized in that, The output of the original load spectrum segmentation evaluation results in S3 includes the following steps: S31. Design a multi-task deep learning model, in which the convolutional neural network branch is used to extract the damage accumulation features of the load signal in the spatial dimension, and the long short-term memory network branch is used to learn the change pattern of the load sequence in the time dimension. S32. Input the operational feature vector into the multi-task deep learning model, and output the load spectrum segmentation and the corresponding damage contribution for different road events respectively. S33. Spatial damage features and temporal load features are spliced ​​together through a feature fusion layer to generate a multidimensional load feature map, and the original load spectrum segmented evaluation results are output using a fully connected layer.

5. The method for generating a road load spectrum by fusing multi-sensor data according to claim 1, characterized in that, The generation of the refined load spectrum fragment in S4 includes the following steps: S41. Establish an expert calibration rule base to store standard load spectrum fragments and corresponding confidence scores calibrated by domain experts under different road surface types and vehicle speed conditions. S42. Search the historical load database, extract historical load spectrum data similar to the current test scenario and their deviation trend from the true value, and use them as dynamic weighting factors. S43. The original load spectrum segmentation evaluation results are probabilistically fused with the expert calibration rule base using the Bayesian inference algorithm, and the posterior probability is calculated by combining the dynamic weighting factor to generate a refined load spectrum segment containing the confidence interval.

6. The method for generating a road load spectrum by fusing multi-sensor data according to claim 5, characterized in that, The S43 step of using a Bayesian inference algorithm for probability fusion and posterior probability calculation includes the following steps: S431. Construct a load spectrum uncertainty quantification model and estimate the uncertainty distribution of the multi-task deep learning model output using the Monte Carlo dropout method; S432. Use a fuzzy logic system to process the qualitative rules in the expert calibration rule base, transform the qualitative evaluation into quantitative weights, and remove abnormal calibration records through a consistency check algorithm. S433. In the Bayesian inference process, the statistical prior of the historical payload database is introduced, the posterior probability is calculated based on the similarity of the current payload spectrum segment, the fusion weight is dynamically adjusted, and a refined payload spectrum segment is generated.

7. The method for generating a road load spectrum by fusing multi-sensor data according to claim 1, characterized in that, The construction of the time-series load spectrum in S5 includes the following steps: S51. Based on the refined load spectrum segments, the gated recurrent unit network in the sequence generation model is used to learn the temporal dependencies between load segments and generate a complete time series load spectrum. S52. Construct a generative adversarial network, using the time series load spectrum as the generator output, and use a discriminator to distinguish between the generated load spectrum and the real load spectrum, and calculate the authenticity score. S53. When the authenticity score is lower than the preset threshold, the parameters of the sequence generation model are adjusted through the backpropagation algorithm, and abnormal segments in the generated time series load spectrum that do not conform to physical laws are removed.

8. The method for generating a road load spectrum by fusing multi-sensor data according to claim 1, characterized in that, The steps in S6 to generate the target load spectrum for durability testing include: S61. Cyclic counting of the time series load spectrum is performed using the rainflow counting method to extract the cyclic matrix of load amplitude and mean. S62. Based on the cyclic matrix and the transfer function of the vehicle suspension system, calculate the equivalent load transmitted to the critical fatigue location of the vehicle body. S63. Based on the equivalent load, load spectrum blocks with preset damage levels are generated by mapping through the finite element model, which serve as the target load spectrum for durability testing.

9. The method for generating a road load spectrum by fusing multi-sensor data according to claim 1, characterized in that, The adaptive optimization loop based on load spectrum accuracy feedback in S7 includes the following steps: S71. Calculate the error distribution between the target load spectrum and the actual road sampling data in the time and frequency domains, and determine the key deviation interval for load spectrum reconstruction. S72. Construct a reinforcement learning agent, using the sampling frequency and fusion weights of a multi-type sensor array as the action space, and the negative value of the error distribution as the reward function. Iteratively optimize the sampling strategy and fusion weights through a deep Q-network. S73. The optimized sampling strategy and fusion weights are sent to the sensor array controller, and the updated configuration is executed in the next round of acquisition, forming an adaptive optimization loop based on load spectrum accuracy feedback.

10. A pavement load spectrum generation system that integrates multi-sensor data, comprising the pavement load spectrum generation method that integrates multi-sensor data according to any one of claims 1-9, characterized in that, The system includes: The data acquisition and preprocessing end collects multi-source sensor data through multi-type sensor arrays and uses Kalman filtering algorithm for time synchronization and noise suppression to generate standardized multi-source data streams; The analysis and evaluation end is based on the road feature knowledge graph. It extracts transient load events and constructs operational feature vectors through wavelet packet decomposition algorithm. It uses a multi-task deep learning model to extract spatial damage features and temporal load features, and outputs the original load spectrum segmented evaluation results. Combining historical load database and expert calibration rules, it performs dynamic weighted calibration through Bayesian inference algorithm to generate refined load spectrum segments. The feedback and update end constructs a time-series load spectrum and uses a generative adversarial network for authenticity verification. Based on the time-series load spectrum, load cycle features are extracted using the rainflow counting method and combined with a finite element model to generate a target load spectrum. Based on the error distribution between the target load spectrum and real road sampling data, the sensor sampling strategy and data fusion weights are dynamically adjusted through a reinforcement learning algorithm to form an adaptive optimization cycle.

Images