Vehicle offline detection method, device, equipment, storage medium and program product

Abstract

The invention provides a vehicle off-line detection method, device and equipment, a storage medium and a program product, and relates to the technical field of vehicle detection, and the method comprises the steps: carrying out the path planning according to the three-dimensional topographic data of a target road section and the detection requirements of a target vehicle, and obtaining a target driving path and a speed curve; controlling the target vehicle to pass through the target road section based on the target driving path and the speed curve, and acquiring vibration data and sound data of the target vehicle in the driving process; performing data preprocessing on the vibration data and the sound data to obtain a target feature vector; and inputting the target feature vector into a target detection model for analysis to obtain a vehicle detection result. According to the method, the vibration data and the sound data when the vehicle runs on the complex road section are collected, the data are processed and then input into the target detection model for deep analysis, the abnormal sound problem can be efficiently and accurately positioned, vehicle health assessment is carried out, and the quality and efficiency of vehicle offline detection are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle detection, in particular to a vehicle off-line detection method, device, equipment, storage medium and program product. BACKGROUND

[0002] In the process of automobile manufacturing, vehicle off-line detection is a key link to ensure product quality. Dynamic road test of complex road conditions such as winding road can effectively simulate the complex stress conditions in actual driving, and has significant meaning for evaluating the body structure strength and identifying vehicle abnormal sound.

[0003] At present, the vehicle off-line detection under complex road conditions in the industry still mainly relies on manual detection method. This method is easily affected by the subjective judgment and operation habit of the driver, which makes it difficult to accurately capture and locate the abnormal sound problem, and the consistency of the detection result is poor. At the same time, it is impossible to collect multi-source data in real time and fully excavate potential quality problems. Therefore, it is urgent to improve the accuracy and efficiency of vehicle off-line detection in complex road sections. SUMMARY

[0004] In view of the above shortcomings of the prior art, the present application provides a vehicle off-line detection method, device, equipment, storage medium and program product, which effectively solves the problems of low accuracy and efficiency of vehicle off-line detection in complex road sections.

[0005] In a first aspect, the present application provides a vehicle off-line detection method, which comprises: path planning according to three-dimensional terrain data of a target road section and detection requirements of a target vehicle, to obtain a target driving path and a speed curve; controlling the target vehicle to pass through the target road section based on the target driving path and the speed curve, and acquiring vibration data and sound data of the target vehicle during driving; data preprocessing of the vibration data and the sound data to obtain a target feature vector; inputting the target feature vector into a target detection model for analysis to obtain a vehicle detection result.

[0006] In an optional implementation, the method further comprises: acquiring environmental perception information and vehicle state information of the target vehicle during passing through the target road section, and dynamically adjusting the driving path of the target vehicle according to the environmental perception information and vehicle state information.

[0007] In an optional implementation, the path planning according to the three-dimensional terrain data of the target road section and the detection requirements of the target vehicle to obtain the target driving path and the speed curve comprises: construct a three-dimensional digital terrain model according to the three-dimensional terrain data, and set a constraint condition of the target vehicle according to the detection requirement; perform global path search based on the three-dimensional digital terrain model to obtain an initial driving path; perform path smoothing and verification according to the initial driving path and the constraint condition to obtain the target driving path; perform model predictive control based on the target driving path to generate the speed curve.

[0008] In an optional implementation, the data preprocessing of the vibration data and the sound data to obtain a target feature vector includes: performing multi-stage filtering, frequency band separation, and feature extraction on the vibration data to obtain a vibration feature vector; performing noise suppression, acoustic feature enhancement, and feature extraction on the sound data to obtain a sound feature vector; performing feature fusion and feature dimension reduction on the vibration feature vector and the sound feature vector to construct the target feature vector.

[0009] In an optional implementation, the target detection module at least includes a abnormal sound detection model and a structure health assessment model, and the input of the target feature vector into the target detection model for analysis to obtain a vehicle detection result includes: performing feature analysis on the target feature vector through the abnormal sound detection model to obtain an abnormal sound type and an abnormal sound position; performing feature analysis on the vibration feature vector through the structure health assessment model to obtain a comprehensive health state result of a vehicle structure, a component connection, and a three-electricity system; performing data fusion decision according to the abnormal sound type, the abnormal sound position, the comprehensive health state result, and a driving parameter of the target vehicle to obtain the detection result.

[0010] In an optional implementation, the dynamic adjustment of the driving path of the target vehicle according to the environment perception information and the vehicle state information includes: performing state coding and feature extraction on the environment perception information and the vehicle state information to obtain coded features; inputting the coded features into a reinforcement learning strategy network for inference calculation to obtain an initial target action; performing smoothness processing, physical constraint verification, and safety boundary checking on the initial target action to obtain a target action; converting the target action into a control instruction, and sending the control instruction to an execution mechanism of the target vehicle to adjust the driving path.

[0011] In a second aspect, the present application provides a vehicle off-line detection device, the device comprising: a path planning module, configured to perform path planning according to three-dimensional terrain data of a target road section and detection requirements of a target vehicle, to obtain a target driving path and a speed curve; a data acquisition module, configured to control the target vehicle to pass through the target road section based on the target driving path and the speed curve, and to acquire vibration data and sound data of the target vehicle during driving; a data processing module, configured to perform data preprocessing on the vibration data and the sound data, to obtain a target feature vector; a vehicle detection module, configured to input the target feature vector into a target detection model for analysis, to obtain a vehicle detection result.

[0012] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the vehicle off-line detection method according to the first aspect of the present application.

[0013] In a fourth aspect, the present application provides a computer-readable storage medium, having a computer program stored thereon, wherein the computer program is executable by a processor to implement the vehicle off-line detection method according to the first aspect of the present application.

[0014] In a fifth aspect, the present application provides a computer program product, comprising computer instructions stored in a computer-readable storage medium, wherein the computer instructions are read and executed by a processor of a computer device to implement the vehicle off-line detection method according to the first aspect of the present application.

[0015] The vehicle off-line detection method, device, equipment, storage medium and program product provided by the present application realize the automation and standardization of the vehicle off-line detection process through automatic path planning combined with automatic driving, eliminate the influence of manual driving operation differences on the detection result, and different batches and different drivers participating in the road test can follow a unified and accurate process, which greatly improves the consistency and reliability of the detection result. Through real-time, comprehensive and accurate acquisition of vibration data and sound data of the vehicle during driving on a complex road section, after data processing, the target detection model is input for deep analysis, which can efficiently and accurately locate the abnormal sound problem, and health assessment is performed on the vehicle structure and three-electric system, thereby effectively improving the quality and efficiency of vehicle off-line detection. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be considered as limiting the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0017] Figure 1 is a first schematic diagram of the vehicle offline detection method provided by the embodiments of the present application; Figure 2 is a second schematic diagram of the vehicle offline detection method provided by the embodiments of the present application; Figure 3 is a third schematic diagram of the vehicle offline detection method provided by the embodiments of the present application; Figure 4 is a fourth schematic diagram of the vehicle offline detection method provided by the embodiments of the present application; Figure 5 is a fifth schematic diagram of the vehicle offline detection method provided by the embodiments of the present application; Figure 6 is a schematic diagram of the structure of the vehicle offline detection device provided by the embodiments of the present application; Figure 7 is a schematic diagram of the structure of an electronic device provided by the embodiments of the present application.

[0018] Main element symbol explanation: 200, vehicle offline detection device; 210, path planning module; 220, data acquisition module; 230, data processing module; 240, vehicle detection module; 300, electronic device; 310, processor; 320, communication interface; 330, memory; 340, communication bus. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the present application more clear, the following will combine the drawings in the embodiments of the present application to make the technical solutions of the present application further clear and complete. It should be noted that the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0020] In addition, the terms "first", "second", "third", etc. are used only for descriptive purposes and are not to be construed as indicating or implying relative importance or an ordered ranking of the indicated technical features. Thus, features defined with "first", "second" or "third" can include, explicitly or implicitly, one or more of such features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise expressly and specifically defined.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0022] In vehicle off-line detection, complex road section test as a core component of dynamic road test can effectively simulate the stress state of the vehicle under complex road conditions, and has important value for evaluating the body structure strength, detecting the connection reliability of parts and identifying potential abnormal sound problems. However, the current industry generally adopts manual detection method, which has significant limitations: first, due to the dependence on the subjective judgment and operation habit of the driver, this method is difficult to realize the accurate positioning of abnormal sound problems, resulting in poor repeatability of the detection results. Secondly, the traditional manual detection cannot obtain real-time multi-source dynamic data of the vehicle passing through the complex road section, and also lacks systematic data processing and analysis capability, so it is difficult to comprehensively identify potential quality defects.

[0023] The embodiments of the present application provide a vehicle off-line detection method, which effectively solves the problems of low accuracy and efficiency of vehicle off-line detection in complex road sections. Figure 1 is a first schematic diagram of the vehicle off-line detection method provided by the embodiments of the present application, as shown in the figure, the method comprises the following steps: Figure 1 S100, path planning is performed according to three-dimensional terrain data of a target road section and detection requirements of a target vehicle, and a target driving path and a speed curve are obtained.

[0024] In the embodiments of the present application, the target road section includes but is not limited to complex road sections such as twisty road, concave-convex road, cobblestone road, Belgium road, steel cable road and gravel road. The three-dimensional terrain data of the target road section is obtained by fusing Beidou satellite positioning, inertial navigation system and high-precision map data, and the centimeter-level positioning accuracy is realized. Figure 2 is a second schematic diagram of the vehicle off-line detection method provided by the embodiments of the present application, as shown in the figure, the path planning specifically comprises the following steps: Figure 2 S110, a three-dimensional digital terrain model is constructed according to the three-dimensional terrain data, and a constraint condition of the target vehicle is set according to the detection requirements.

[0025] ​​Optionally, a three-dimensional digital terrain model containing road surface elevation, curvature, slope and obstacle position is constructed according to the three-dimensional terrain data of the target section. At the same time, the constraint conditions of the target vehicle are set according to the detection requirements. Exemplarily, the vehicle dynamics constraint conditions are set as follows: minimum turning radius 5.5 meters, maximum lateral acceleration 0.3g and maximum longitudinal acceleration 0.25g; the detection requirement constraint conditions are set as follows: maintaining a constant detection speed of 10 km / h and ensuring that all detection sensors have effective coverage area.

[0026] S120, global path search is performed based on the three-dimensional digital terrain model to obtain an initial driving path.

[0027] As an optional implementation of the embodiments of the present application, an improved A-star search algorithm can be used for global path search in the three-dimensional digital terrain model, including the following steps: First, state space discretization is performed, and the three-dimensional digital terrain model is discretized into 0.1 meter x 0.1 meter grid cells, each cell containing attributes such as elevation and road surface type.

[0028] Then, a cost function is designed to integrate multiple optimization objectives. The cost function includes but is not limited to path smoothness cost, safety cost, detection effect cost and energy consumption cost. The path smoothness cost evaluates the path curvature change rate to make the vehicle turn more smoothly; the safety cost calculates the minimum distance between the path and the obstacle to ensure driving safety; the detection effect cost evaluates the coverage integrity of the path for abnormal sound detection to improve detection quality; and the energy consumption cost considers the impact of slope change on power consumption to achieve energy-saving planning.

[0029] Finally, the heuristic search process is performed, starting from the starting point and expanding nodes. Each expansion fully considers multiple directions that the vehicle can reach. For each candidate node, the comprehensive cost value is calculated according to the cost function, and then the node with the smallest cost value is selected for further expansion. This cycle continues until the target point is reached, completing efficient and accurate global path search.

[0030] S130, path smoothing and verification are performed according to the initial driving path and the constraint conditions to obtain a target driving path.

[0031] Exemplarily, path smoothing and verification can include B-spline curve fitting, curvature continuity check, dynamics feasibility verification and safety boundary confirmation. Among them, B-spline curve fitting is a smoothing process for discrete path points to ensure second-order continuity of the path; curvature continuity check verifies that the curvature change of the initial driving path is within the vehicle turning ability range; dynamics feasibility verification ensures that the initial driving path is executable within the vehicle dynamics constraints; and safety boundary confirmation checks that the initial driving path maintains a sufficient safety distance from the obstacle.

[0032] S140, performing model predictive control based on the target driving path to generate a speed curve.

[0033] As an optional implementation of the embodiment of the application, the model predictive control method can be used to generate an optimal speed curve based on the target driving path, and the specific implementation is as follows: First, a dynamic model is established, and the vehicle model comprehensively covers longitudinal dynamics and fully considers the influence of slope, including but not limited to the influence of motor torque characteristics, transmission efficiency, rolling resistance, and slope resistance on vehicle acceleration. The output characteristics of the motor torque characteristics directly affect the vehicle power, and the transmission efficiency is also considered. The loss of energy in the transmission process will affect the final power output. At the same time, the rolling resistance as the inherent resistance of the vehicle during driving is also included in the model. In addition, the influence of slope resistance on vehicle acceleration is accurately quantified, and the vehicle acceleration or deceleration is different under different slopes. This vehicle model can accurately simulate this dynamic process.

[0034] Then, the constraint conditions are set, for example, the speed is limited in the range of 5-15 km / h to ensure that the detection is effectively performed, the acceleration is limited to ±0.25g to avoid discomfort caused by rapid acceleration or deceleration, the jerk is limited to ±0.5g / s to make the control process smoother.

[0035] Finally, the optimal speed curve is obtained through multi-objective optimization, for example, the main target is to keep the detection speed stable at 10 km / h, while minimizing energy consumption to improve economy, and ensuring that all safety and comfort requirements can be met.

[0036] The embodiment of the application generates an optimal target driving path and speed curve that meets the detection requirements and guarantees safety and comfort through path planning, and provides standardized test conditions for subsequent detection. The complex path planning problem is decomposed into multiple manageable sub-problems, and the comprehensive optimality of the final path is ensured through sequential optimization. The parameters of each step can be adjusted according to the specific vehicle model and detection requirements.

[0037] S200, controlling the target vehicle to pass through the target road section based on the target driving path and the speed curve, and obtaining vibration data and sound data of the target vehicle in the driving process.

[0038] In the embodiment of the application, acceleration sensors, displacement sensors, and strain gauges are distributedly installed at key positions of the body frame, the three-electricity system, and the suspension system of the target vehicle. The key positions include but are not limited to A-pillar, B-pillar, C-pillar, roof, floor, power battery, drive motor, electronic control system, shock absorber, spring, swing arm, and the like. These sensors collect vibration data of the vehicle in real time at a high frequency of 1000 Hz or more when the vehicle drives on the target road section, covering vibration information in X, Y, and Z directions.

[0039] Meanwhile, high-sensitivity microphone arrays are arranged in areas such as the driver's cabin, passenger cabin, and trunk of the vehicle, and active noise reduction technology is combined to collect sound data during the driving process of the target vehicle. Through hardware noise reduction and software filtering algorithms, environmental noise and high-frequency noise interference of the motor of the target vehicle are removed to obtain pure abnormal sound data.

[0040] For example, hardware noise reduction first optimizes the arrangement of the microphone array. One omnidirectional microphone is arranged on the central top of the front and rear seats in the driver's cabin of the target vehicle, one directional microphone is arranged on the inner side of the B-pillar on the left and right sides of the passenger cabin, one omnidirectional microphone is arranged on the top center of the trunk, and two anti-interference microphones are arranged on the inner side of the firewall.

[0041] Then, active noise reduction hardware is set, including arranging reference microphones and configuring secondary sound sources. Two reference microphones can be arranged near the motor compartment to collect high-frequency noise of the motor, two reference microphones can be arranged on the vehicle chassis to collect road noise, and one reference microphone can be arranged on the air outlet of the air conditioner to collect wind noise signals. The secondary sound source configuration can include arranging four noise reduction speakers on the ceiling of the vehicle, and arranging eight small vibration speakers in the door trim to form a closed-loop control system.

[0042] Finally, analog signal preprocessing is performed through a signal conditioning circuit and an adaptive noise reduction circuit. The signal conditioning circuit uses an instrument amplifier for preamplification, designs an anti-aliasing filter with a cutoff frequency of 20 kHz, adjusts the signal dynamic range, and the adaptive noise reduction circuit uses an LMS adaptive filter hardware to realize real-time calculation of the motor noise reverse sound wave, which is emitted through the noise reduction speaker to cancel the noise.

[0043] For example, the software filtering algorithm includes the following steps: first, digital signal preprocessing is performed, the signal is digitized, appropriate sampling frequency and quantization accuracy are set, synchronous sampling technology is used to align the time of each channel, pre-emphasis processing is performed, and a first-order high-pass filter is used to compensate for the high-frequency attenuation of sound propagation.

[0044] Then, environmental noise suppression is performed. The input signal is subjected to fast Fourier transform, the power spectral density is calculated to identify the noise frequency band, a noise spectrum template library is established, and spectral subtraction is used for noise reduction. The noise power spectrum is estimated and updated in the silent section, the noise power spectrum is subtracted, and over-subtraction technology and spectral flooring are used to avoid music noise.

[0045] Next, motor noise elimination is performed. The fundamental frequency and harmonic frequency are calculated based on the motor speed signal, a harmonic model is established, the speed is tracked in real time to adjust the parameters, the normalized least mean square algorithm is used to update the filter coefficients, and the step size parameter is adaptively adjusted.

[0046] Then, wind noise and road noise are processed, the non-stationary characteristics of wind noise are identified, the transient components are analyzed by wavelet transform, the wind noise is removed based on threshold shrinkage of wavelet coefficients, the road excitation frequency is identified by combining the data of the vibration sensor, a transfer function model is established, and the road noise is removed by multi-channel adaptive filtering.

[0047] Finally, the abnormal sound signal is enhanced, short-time Fourier analysis is adopted, mel frequency cepstral coefficients are calculated, a time-frequency mask is constructed to enhance the abnormal sound components, the signal is enhanced based on an improved Wiener filter, the naturalness is maintained by a phase reconstruction algorithm, and pure abnormal sound data is output. Meanwhile, quality monitoring, parameter adaptive adjustment, fault detection and fault tolerance mechanism can be set to ensure the stability of the software system.

[0048] In the embodiments of the present application, a high-speed data acquisition card can be used to synchronously acquire vibration data of the vibration sensor and abnormal sound data output by the microphone array, and perform preprocessing operations such as AD conversion, denoising filtering and data compression. The preprocessed vibration data and sound data are transmitted in real time to the data processing module through the vehicle-mounted Ethernet.

[0049] For example, AD conversion mainly includes two key steps of signal conditioning and buffering and quantization and coding. The signal conditioning and buffering processes different sensor signals. The vibration sensor signal is amplified by 100 times by an instrument amplifier to effectively acquire weak signals, and the microphone signal is adjusted in level by a preamplifier to match the ADC input range. At the same time, an anti-aliasing filter with a cutoff frequency of 40% of the sampling frequency is used for signal filtering. All channels are synchronously sampled using the same clock source to ensure that the time alignment accuracy is less than 1 microsecond.

[0050] The quantization and coding use uniform quantization to quantize the level number, set a reasonable quantization interval to avoid signal clipping, and dynamically adjust the reference voltage to adapt to different amplitude signals. Then, data format conversion is performed to convert the quantized digital signal to a binary complement format, add microsecond-level precision timestamp information, and attach sensor ID and state information.

[0051] For example, the denoising filtering process can include time domain filtering, frequency domain filtering, and outlier processing steps. The time domain filtering process is aimed at the direct current component, a first-order high-pass filter is used to eliminate the influence of sensor zero drift and temperature drift, and the signal baseline is kept stable. For power frequency interference, an adaptive notch filter is used, which can automatically track the power frequency change and effectively eliminate the interference introduced by the power line.

[0052] The frequency domain filtering process is aimed at the vibration signal, a reasonable band-pass filtering range is set, a Butterworth filter is used to ensure that the passband is flat and the stopband attenuation is greater than the set range, a reasonable band-pass filtering range is set for the sound signal, and a linear phase FIR filter is used for processing to avoid phase distortion, and the passband ripple is less than the set range.

[0053] The outlier processing is first based on amplitude threshold to detect burst interference, uses median filtering to eliminate impulse noise, detects normal data points through a sliding window, labels the detected outliers, and uses linear interpolation to repair missing data, thereby maintaining data continuity and integrity.

[0054] For example, the data compression processing can include the steps of lossless compression preprocessing, lossy compression optimization, and compression quality evaluation. The lossless compression preprocessing first groups the data, processes it by sensor type, and each group contains 256 sampling points. The group header information containing the sensor ID, timestamp, and data length is added, and then differential encoding is carried out. The first-order difference is made to the continuous sampling value, thereby reducing the data dynamic range and improving the subsequent compression efficiency.

[0055] The lossy compression optimization uses a wavelet transform compression algorithm for vibration data. The db8 wavelet basis is used for multi-layer decomposition, and the coefficients with an energy ratio of 95% are retained through threshold processing. The compression ratio is controlled at 4:1 to 8:1. For sound data, an improved MPEG audio compression algorithm is used. Based on the psychoacoustic model, the inaudible components are removed, and the main abnormal sound characteristic frequency components are retained. The compression ratio is controlled at 6:1 to 10:1.

[0056] The compression quality evaluation first performs distortion detection, calculates the signal-to-noise ratio before and after compression, analyzes the frequency characteristic retention degree, and verifies the abnormal sound characteristic integrity. Then, the compression parameters are dynamically adjusted according to the signal characteristics, the compression ratio is optimized under the premise of ensuring the quality, and the compression effect is monitored in real time.

[0057] Through hardware noise reduction, software filtering algorithm, and data acquisition preprocessing, the embodiments of the present application can ensure that the collected vibration data and sound data have the characteristics of high quality, low noise, and appropriate data volume, thereby providing a reliable data basis for subsequent abnormal sound identification and structure health analysis.

[0058] As a further embodiment of the present application, the driving path of the target vehicle can also be dynamically adjusted according to the environmental perception information and vehicle state information of the target vehicle during the process of passing through the target road section.

[0059] Optionally, a multi-line laser radar, a millimeter wave radar, a surround camera, and an ultrasonic sensor are deployed on the target vehicle. The laser radar is used to construct a three-dimensional point cloud map of the twisted road in real time, detect the road surface undulation and obstacles; the millimeter wave radar monitors the distance, speed and angle of the objects around the vehicle in bad weather; the surround camera identifies the road boundary and traffic signs, and the ultrasonic sensor assists in detecting close-range obstacles, thereby providing comprehensive environmental information for automatic driving decision-making.

[0060] Based on the environmental perception information and the preset detection strategy, a reinforcement learning algorithm is used to dynamically generate automatic driving decision instructions, and through a by-wire steering, by-wire braking and by-wire driving system, the steering angle, speed and power output of the target vehicle are accurately controlled to realize stable and safe automatic driving of the vehicle on the target road section. Figure 3 is a third schematic diagram of a vehicle off-line detection method provided by the embodiment of the application, as shown in Figure 3 The reinforcement learning algorithm is used to dynamically generate automatic driving decision instructions to realize dynamic adjustment of the path, which specifically includes the following steps: S210, state encoding and feature extraction are performed on the environmental perception information and vehicle state information to obtain encoded features.

[0061] In the embodiment of the application, the environmental perception information includes but is not limited to laser radar point cloud data, visual perception data, millimeter wave radar data and ultrasonic radar data. The vehicle state information includes but is not limited to the current position coordinates, the heading angle, the current speed, the acceleration, the steering angle, the vehicle attitude and the tire and road surface contact force estimation value and the like.

[0062] First, the environmental perception information from different sensors such as cameras, radars and laser radars is integrated to ensure that the environmental perception information is accurately aligned in the time dimension and to avoid information errors caused by time differences. Then, feature engineering technology is used to extract key state information from the environmental perception information and the vehicle state information, such as vehicle speed, acceleration, distance from surrounding objects and relative speed. Subsequently, the extracted key state information is normalized and standardized to unify the data of different dimensions to the same range to obtain the encoded features. For example, the speed data is normalized to the [0, 1] interval to eliminate the magnitude difference between the data, so that the subsequent policy network can more efficiently process and analyze the data.

[0063] S220, the encoded features are input into a reinforcement learning policy network for inference calculation to obtain an initial target action.

[0064] In the embodiment of the application, a reinforcement learning algorithm framework needs to be constructed, which specifically includes the following steps: First, define the state space, which includes historical environmental perception information, historical vehicle state data and detection task state, and the detection task state includes but is not limited to the proportion of the completed detection road section, the current detection quality score and the remaining detection time budget.

[0065] Then define the action space, which includes continuous values of vehicle control instructions, such as acceleration instruction range and target speed setting for longitudinal control actions; steering angle instruction range and steering angle velocity range for lateral control actions; specific road section deceleration detection instructions, key area repeated detection instructions and abnormal situation emergency handling instructions for special detection actions.

[0066] Finally, a reward function is designed, which includes a weighted combination of multiple optimization objectives, which can include a safety reward term, a detection effect reward term, and an efficiency reward term. Among them, the safety reward term can be to calculate the collision avoidance reward based on the minimum distance from the obstacle, to calculate the stability reward based on the lateral acceleration and yaw rate of the vehicle, and to calculate the boundary keeping reward based on the distance between the vehicle and the road boundary.

[0067] The detection effect reward term can be to calculate the speed consistency reward based on the deviation of the actual speed from the standard detection speed, to calculate the path tracking reward based on the deviation of the actual path from the standard detection path, and to calculate the detection integrity reward based on the proportion of the covered detection area.

[0068] The efficiency reward term can be to calculate the time efficiency reward based on the time to complete the detection task, to calculate the energy efficiency reward based on the energy consumption, and to calculate the smoothness reward based on the change rate of the control instruction.

[0069] The reinforcement learning algorithm framework is trained to obtain a reinforcement learning strategy network, and the training includes an offline training phase and an online learning phase.

[0070] The offline training phase includes simulation environment construction, model pre-training, and strategy verification and optimization. The simulation environment construction includes: building a high-fidelity target road section three-dimensional simulation environment, accurately simulating the details of the real road conditions. At the same time, simulate various weather and light conditions, such as heavy rain, heavy fog, strong light, etc., and light changes at different times, so that the agent can adapt to complex environments. In addition, set up various fault scenarios, such as sensor failure and vehicle component failure, etc., to improve the ability of the reinforcement learning model to deal with unexpected situations.

[0071] Model pre-training uses the proximal policy optimization algorithm to start initial training, which can effectively balance exploration and utilization. Through curriculum learning, start from simple scenes and gradually increase difficulty to complex scenes, so that the model can steadily improve its ability. Use experience replay mechanism to store past experience and randomly sample learning to break data correlation and improve training efficiency.

[0072] Strategy verification and optimization test the trained strategy in the simulation environment, analyze its performance, adjust the weight of the reward function according to the test results, and guide the model to learn better behavior. At the same time, optimize the neural network structure and hyperparameters, such as the number of layers and learning rate, to improve the performance of the reinforcement learning model.

[0073] The online learning stage includes real-time policy updating and adaptive adjustment. The real-time policy updating continuously optimizes the policy based on actual operation data, making the model adapt to real scenarios. Meanwhile, transfer learning is used to apply learned knowledge to different vehicle models and road conditions, accelerating the learning process. Integrated learning is used to fuse decisions from multiple models, improving decision robustness. Adaptive adjustment adjusts perception weights according to sensor performance changes to ensure data accuracy, adjusts control parameters based on vehicle state changes to achieve precise control, and optimizes decision logic based on detection results to improve overall decision quality.

[0074] The encoded features are input into a pre-trained reinforcement learning policy network, which has the ability to make reasonable decisions based on different states through learning from a large amount of driving scene data. Inside the network, the input encoded features are calculated layer by layer through a forward propagation algorithm, and finally the action probability distribution is output. Based on this action probability distribution, a sampling method can be used to randomly select an action, or the optimal action with the highest probability can be selected to determine the target vehicle's next action direction and obtain the initial target action.

[0075] S230, smoothing processing, physical constraint verification and safety boundary check are performed on the initial target action to obtain the target action.

[0076] In the embodiments of the present application, the initial target action is subjected to smoothing processing to avoid sudden changes in action that can cause the vehicle to drive unstably, for example, when accelerating or turning, making the action change more continuous and natural. Then, physical constraint verification is performed to ensure that the selected action meets the physical characteristics of the vehicle, such as not exceeding the acceleration, braking and turning limits of the vehicle. Finally, safety boundary check is performed to ensure that the action is executed within a safe range, avoiding dangerous situations such as collisions with other objects.

[0077] S240, the target action is converted into control instructions, and the control instructions are sent to the actuators of the target vehicle to adjust the driving path.

[0078] In the embodiments of the present application, the processed target action is converted into specific control instructions, such as throttle opening, brake pressure and steering angle, etc. These control instructions are sent to the actuators of the vehicle, such as the engine, brake system and steering system, through the drive-by-wire system. At the same time, the execution effect of the control instructions is monitored in real time, and the instructions are adjusted in a timely manner according to the actual situation to ensure that the vehicle drives safely as expected.

[0079] The embodiments of the present application can dynamically generate optimal driving decision instructions based on real-time environmental perception information and vehicle state information through reinforcement learning algorithms, ensuring that the target vehicle safely, stably and efficiently completes the detection task on the target road segment.

[0080] S300, data pre-processing is performed on the vibration data and the sound data to obtain a target feature vector.

[0081] In the embodiments of the present application, the vibration data and the sound data obtained are subjected to secondary filtering and feature extraction, and the vibration data and the sound data are converted into a target feature vector that can be recognized by a target detection model. Figure 4 is a fourth schematic diagram of a vehicle off-line detection method provided by the embodiments of the present application, as shown in the figure, the data pre-processing specifically includes the following steps: Figure 4 S310, multi-stage filtering, frequency band separation and feature extraction are performed on the vibration data to obtain a vibration feature vector.

[0082] As an optional implementation manner of the embodiments of the present application, the multi-stage filtering includes adaptive Kalman filtering, wavelet packet noise reduction and morphological filtering. Among them, the adaptive Kalman filtering establishes a vehicle vibration state space model containing position, velocity and acceleration state variables, dynamically adjusts the process noise covariance matrix according to the real-time vibration characteristics, and then eliminates the sensor measurement noise through the prediction and correction cycle. The wavelet packet noise reduction selects db8 wavelet basis function for 5-layer wavelet packet decomposition, adopts Stein unbiased risk estimation threshold to perform soft threshold processing on the detail coefficients, and finally reconstructs the signal to retain the effective vibration frequency band of 0.5-2000Hz. The morphological filtering uses a circular structural element for open-close combined operation to eliminate the pulse interference and baseline drift in the signal.

[0083] The frequency band separation processing divides the vibration signal into a low frequency band of 0.5-20Hz, a medium frequency band of 20-200Hz and a high frequency band of 200-2000Hz according to the frequency. In the low frequency band, the Butterworth low-pass filter is used to retain the rigid body vibration characteristics of the vehicle body, in the medium frequency band, the Chebyshev band-pass filter is used to extract the suspension system vibration characteristics, and in the high frequency band, the elliptical filter is used to capture the component resonance characteristics.

[0084] The feature extraction includes time domain feature extraction, frequency domain feature extraction and time-frequency domain feature extraction, wherein the time domain features include statistical features, envelope features and impact features. The time domain features are calculated by the mean, variance, skewness, kurtosis, waveform factor and pulse factor of the signal, which can reflect the concentration tendency, dispersion degree and distribution form of the vibration signal. The envelope feature uses Hilbert transform to extract the envelope line of the signal and calculates the envelope root mean square value, which can capture the amplitude modulation information of the vibration signal. The impact feature counts the peak value and calculates the impact index, which is used to identify the impact component in the vibration signal.

[0085] ​Frequency domain characteristics are analyzed by calculating the power spectral density using Fast Fourier Transform (FFT) to examine the frequency components of the signal. The signal is divided into 1 / 3 octave bands, and the relative energy ratio of each band is calculated to obtain the band energy distribution and understand the signal energy distribution across different frequency bands. Spectral features such as spectral centroid, spectral variance, spectral skewness, and spectral kurtosis are calculated to describe the spectral distribution characteristics. Simultaneously, the first 10 natural frequencies and damping ratios are extracted to identify the system's resonance characteristics.

[0086] The time-frequency domain features include wavelet packet energy entropy, empirical mode decomposition (EMD), and Wigner quasi-probability distribution. Wavelet packet energy entropy is calculated through a five-level wavelet packet decomposition, determining the relative energy entropy of each node to reflect the signal's energy distribution in the time-frequency domain. EMD extracts the energy features of the first eight intrinsic mode functions (EMFs) to analyze the signal's nonlinear and non-stationary characteristics. Wigner quasi-probability distribution calculates the singular value features of the time-frequency matrix, obtaining the joint time-frequency information of the signal.

[0087] S320. Perform noise suppression, acoustic feature enhancement, and feature extraction on the sound data to obtain the sound feature vector.

[0088] In this embodiment, noise suppression includes using spectral subtraction to suppress steady-state background noise and using adaptive filtering to suppress high-frequency motor noise.

[0089] To suppress steady-state background noise using spectral subtraction, the noise power spectrum is first estimated while the target vehicle is not running, and this estimate is used as a baseline. Then, the noise power spectrum is subtracted from the noisy speech spectrum. To prevent over-subtraction from causing musical noise, an over-subtraction factor is introduced, and spectral flooring is set to ensure that the processed spectrum does not contain abnormally low values.

[0090] To suppress high-frequency noise from a motor using adaptive filtering, the mapping relationship between motor speed and noise frequency is first established to clarify the characteristics of high-frequency noise generated by the motor at different speeds. Then, the least mean square algorithm is used to dynamically adjust the filter coefficients based on the real-time input signal and the desired signal, effectively filtering out the high-frequency noise generated by the motor while retaining sound information in the 200-8000Hz frequency range that is sensitive to the human ear.

[0091] Acoustic feature enhancement begins with pre-emphasis processing, using a first-order high-pass filter to boost the high-frequency components of the audio signal, increasing the amplitude of the high-frequency signal and enhancing the clarity of the audio signal. Next, frame-by-frame windowing is applied using a 25ms Hamming window with a 10ms frame shift, dividing the continuous audio signal into multiple short frames for easier subsequent analysis. Finally, cepstral smoothing is performed using homomorphic filtering to smooth the cepstral spectrum of the audio signal, eliminating the influence of duct excitation effects and highlighting the acoustic features of the audio data.

[0092] The sound feature vector extraction is divided into three aspects of traditional acoustic features, advanced acoustic features and auditory perception features, wherein the traditional acoustic features include mel-frequency cepstral coefficients, linear prediction coefficients and pitch frequency. The mel-frequency cepstral coefficients are obtained by pre-processing the sound signal, performing frame division and windowing, performing Fourier transform, obtaining energy through a mel filter bank, taking logarithm, performing discrete cosine transform, extracting 13-dimensional mel-frequency cepstral coefficients, and calculating first-order and second-order differences. The linear prediction coefficients are obtained by using an autoregressive model and performing linear prediction analysis to obtain 12-order LPC coefficients and calculate residual energy. The pitch frequency is obtained by using an autocorrelation function to find the most periodic component in the signal and extracting the fundamental frequency and harmonic structure.

[0093] The advanced acoustic features include voiceprint features, timbre features and pitch features. The voiceprint features are obtained by calculating spectral centroid, roll-off point, flux and decay time, and are used to depict the sound spectrum features. The timbre features are obtained by extracting Brightness, Spread and Skewness spectral shape parameters, and reflect the sound timbre. The pitch features are obtained by analyzing the instantaneous frequency and extracting the frequency modulation features.

[0094] The auditory perception features include loudness features, sharpness and roughness. The loudness features are obtained by calculating the instantaneous loudness of the sound signal according to the ISO 532B standard. The sharpness is obtained by calculating the spectral sharpness based on the Zwicker model. The roughness is obtained by extracting the fluctuation intensity of the modulation frequency in the range of 15-300 Hz.

[0095] S330, the vibration feature vector and the sound feature vector are fused and dimensionally reduced to construct a target feature vector.

[0096] In the embodiments of the present application, the feature fusion first splices the extracted vibration feature vector and sound feature vector at the feature layer, so that the features of different modalities converge together. Then, the maximum and minimum value normalization method is used to unify the dimensions of the feature vectors, so as to avoid the influence of the dimension difference on the subsequent analysis. Finally, the feature selection algorithm is used to remove redundant features, reduce the data dimension, and improve the calculation efficiency.

[0097] The feature dimension reduction uses principal component analysis to retain the principal components with a cumulative contribution rate of more than 95%, so as to reduce the dimension and retain the main information of the data as much as possible. Linear discriminant analysis is used to maximize the ratio of inter-class scatter to intra-class scatter, so as to enhance the classification ability of the features. Local linear embedding can also be used to maintain the local topological structure of the feature space, so that the feature vector after dimension reduction can still reflect the local relationship of the original data.

[0098] Finally, a fixed-dimension target feature vector is constructed, which integrates multi-dimensional information such as time domain, frequency domain and time-frequency domain, and has rotation invariance and scale invariance, thereby providing high-quality feature input for subsequent model training and classification recognition.

[0099] S400, inputting the target feature vector into a target detection model for analysis to obtain a vehicle detection result.

[0100] As a preferred embodiment in the embodiments of the present application, the target detection module includes an abnormal sound detection model and a structure health assessment model. Figure 5 is a fifth schematic diagram of a vehicle offline detection method provided by the embodiments of the present application, as shown in the figure, the vehicle detection specifically includes the following steps: Figure 5 S410, performing feature analysis on the target feature vector by the abnormal sound detection model to obtain an abnormal sound type and an abnormal sound position.

[0101] In the embodiments of the present application, the abnormal sound detection model is constructed based on a deep learning framework, integrates a convolutional neural network and a long short-term memory network, and learns abnormal sound feature patterns through a large amount of vehicle abnormal sound data as training samples, wherein the vehicle abnormal sound data includes vibration data and sound data under normal working conditions and fault working conditions. The abnormal sound detection model can analyze the target feature vector in real time, accurately identify the abnormal sound type, and locate the abnormal sound source position.

[0102] Optionally, the abnormal sound detection model includes an input layer, an improved convolutional neural network layer, an improved long short-term memory network layer, a feature fusion layer and an output layer.

[0103] The input layer adopts a multi-modal data fusion input strategy, the vibration feature vector is input in a dimension of 3 channels x 256 time steps x 20 sensors, the sound feature vector is input in a dimension of 1 channel x 1024 time steps x 8 microphones, and 16-dimensional vehicle state parameters are introduced as auxiliary input.

[0104] The improved convolutional neural network layer includes parallel convolution paths, depth separable convolution and attention enhancement mechanism. Through the parallel convolution paths, 7x7 large kernel convolution is used to capture macro vibration patterns, 3x3 medium kernel convolution is used to extract medium scale features, and 1x1 small kernel convolution is used to focus on local details. The depth separable convolution decomposes the standard convolution into depth convolution and point-by-point convolution, thereby reducing the parameter quantity while maintaining the feature extraction capability. The attention enhancement mechanism includes three parts: the channel attention module adaptively adjusts the feature channel weight, the spatial attention module focuses on the abnormal sound related area, and the time domain attention module highlights the key time point features, thereby forming a dynamic feature weighting mechanism.

[0105] ​The improved long short-term memory network layer adopts a bidirectional LSTM structure: the forward layer encodes historical information dependencies, the backward layer captures future information correlations, and the bidirectional outputs are weighted and fused through an attention mechanism. In terms of gated mechanism optimization, the input gate introduces an adaptive forgetting factor to balance new and old information, the output gate dynamically adjusts the weight in combination with the current input, and the cell state update adopts gradient clipping to prevent gradient explosion. Multi-layer LSTM stacking realizes hierarchical extraction of time series features: the bottom layer captures local time dependencies, the middle layer models medium time scale patterns, and the top layer captures long-term time series relationships, forming complete time series representations from transient to steady state.

[0106] The feature fusion layer is realized through cross-modal interaction. The vibration feature vector and the acoustic feature vector use a cross-attention mechanism, and the vehicle state parameters affect the feature weight through conditional embedding. Finally, the joint feature representation is formed through weighted fusion. In terms of spatio-temporal feature alignment, 3D convolution is used to capture joint spatio-temporal features, a feature pyramid network is constructed in the time dimension to realize multi-scale time series analysis, and spatial structure information is preserved through feature map concatenation.

[0107] The output layer is designed with a dual-task architecture: supporting multi-label identification for heterogeneous sound classification, and outputting class confidence and Monte Carlo Dropout uncertainty estimation. Heterogeneous sound location positioning generates a heat map through a deconvolution network, and combines bounding box regression and spatial probability distribution to realize three-dimensional source positioning.

[0108] Based on the constructed heterogeneous sound detection model, model training is performed, and the specific steps of model training are as follows: First, prepare the training data. To improve data diversity, use enhancement strategies: time domain enhancement simulates different collection conditions through time warping and speed disturbance, frequency domain enhancement enhances the robustness of spectral features using frequency masking and frequency band enhancement; spatial enhancement simulates device failure scenarios by randomly discarding sensor data; environmental enhancement adds background noise of different signal-to-noise ratios to simulate real environment interference. In terms of sample weight adjustment, a difficult mining strategy is implemented for classification error samples, and their loss weights are dynamically increased to force the model to focus; for minority class samples, use class balancing techniques to alleviate the class imbalance problem through oversampling or loss weighting; at the same time, based on the labeling quality, the samples are weighted, and the high confidence labeling data is used preferentially.

[0109] Then a hierarchical progressive training process is adopted. The first stage is self-supervised pre-training, which uses large-scale unlabeled data to construct robust low-level feature representation through contrastive learning, avoiding excessive dependence on labeled data. The second stage is supervised fine-tuning, which uses labeled data to optimize model parameters end-to-end, adopts a hierarchical learning rate strategy and gradual unfreezing technology, the lower convolutional layers use a smaller learning rate to preserve the pre-trained features, and the top fully connected layers use a larger learning rate to accelerate convergence, starting from the output layer and gradually unfreezing the network layers to avoid gradient shock in the initial stage. The third stage is adversarial training, which generates adversarial samples through projected gradient descent and alternately trains on clean samples and adversarial samples, significantly improving the model's robustness to noisy inputs.

[0110] Optionally, the design of the loss function adopts multi-task collaborative optimization, and a multi-task loss function is constructed. The multi-task loss function can include a classification loss, a positioning loss, and a regularization loss. The classification loss includes a focal loss, label smoothing, and knowledge distillation. The focal loss solves the class imbalance problem, the label smoothing prevents overfitting to the training labels, and the knowledge distillation uses the soft labels provided by the teacher model. The positioning loss includes a heatmap loss, a bounding box loss, and a distance loss. The heatmap loss supervises the probability distribution of the heterogeneous source through mean square error, the bounding box loss optimizes the spatial coordinate accuracy through intersection over union, and the distance loss is used to constrain the spatial distance between the predicted position and the true position. The regularization loss includes weight decay, feature distribution alignment, and consistency regularization. The weight decay prevents overfitting through L2 regularization, the feature distribution alignment reduces the distribution difference between the training set and the test set through domain adaptive technology, and the consistency regularization forces the model output to be stable through input perturbation.

[0111] Optionally, the training parameters are configured as follows: the optimizer can be AdamW, which combines weight decay to avoid the conflict between L2 regularization and adaptive learning rate; the initial learning rate is set to 0.001 and cosine annealing scheduling is used to achieve smooth decay; the gradient clipping threshold is set to 1 to prevent gradient explosion in the adversarial training phase; the batch size is set to 32 to balance memory usage and gradient estimation stability, and the gradient accumulation technique is used to simulate the effect of large batch training; mixed precision training accelerates calculation using FP16, which avoids numerical underflow by dynamically scaling the loss value, significantly improving training efficiency.

[0112] Based on the trained heterogeneous sound detection model, model performance evaluation and optimization are performed. For the classification task, four-dimensional indicators are used: accuracy measures overall classification correctness, precision focuses on the accuracy of heterogeneous sound detection, recall evaluates the detection ability of heterogeneous sound samples, and F1 score balances the accuracy and recall. For the positioning task, three aspects are quantified: positioning accuracy measures the spatial accuracy of the heterogeneous sound source position error distance, intersection over union calculates the overlap between the detected region and the true region, and recall calculates the proportion of correctly positioned heterogeneous sound sources.

[0113] The model optimization strategy includes two paths: neural network architecture search automatically explores the optimal network structure using reinforcement learning or evolutionary algorithms, and balances accuracy and computational efficiency through multi-objective optimization; knowledge distillation adopts a teacher-student framework to transfer the generalization ability of a large teacher model to a lightweight student model, combining intermediate layer feature alignment feature distillation and soft label supervised output distillation to significantly reduce computational overhead while maintaining performance.

[0114] The heterophonic detection model provided in the embodiments of the present application has high heterophonic classification accuracy, small positioning error, and multi-label recognition capability to cope with complex scenarios; at the same time, the single inference time is short, the model parameter amount is small, and the generalization capability is significantly enhanced, which can quickly adapt to different complex road conditions and has strong robustness to background noise interference.

[0115] S420, performing feature analysis on the vibration feature vector through the structure health assessment model to obtain a comprehensive health state result of the vehicle structure, component connection and three-electric system.

[0116] In the embodiments of the present application, a structure health assessment model based on vibration data is established in combination with a vehicle dynamics model and finite element analysis results. The vibration frequency, amplitude and phase change characteristics of key parts of the vehicle body are analyzed through the structure health assessment model to evaluate the vehicle body structure strength, component connection state and mechanical stability of the three-electric system, and to predict potential failure risks. The structure health assessment model includes the following steps: First, a vehicle dynamics model including the vehicle body, suspension, steering and powertrain is constructed in a multi-body dynamics simulation platform. The model parameters are determined through multi-source data fusion: the mass matrix is obtained by calculating the mass properties based on the geometric parameters extracted from the CAD models of each component, the stiffness matrix is obtained by analyzing the stiffness characteristics of the key connection points through finite element analysis, and the damping matrix is parameterized using an equivalent viscous damping model based on the damping ratio data measured by experimental modal analysis. The vehicle dynamics model can simulate the dynamic response of the vehicle under target road conditions, providing a dynamic basis for subsequent vibration feature analysis.

[0117] Then, finite element analysis of key structures is performed. For high-risk heterophonic source components such as the vehicle body, battery pack shell and motor support, fine finite element modeling is carried out. First, element meshing is performed, and then boundary conditions and loads under typical working conditions are applied, including: static strength analysis to simulate the stress distribution and deformation under extreme conditions such as full load and emergency braking, modal analysis to extract the first 20 natural frequencies and modes to identify structural resonance risks; frequency response analysis to calculate the vibration transfer function under road excitation and establish a component-level vibration characteristic map.

[0118] Finally, a health state benchmark database is constructed. The vibration data of 50 qualified vehicles covering mainstream models such as SUVs and sedans are collected on a target road section in a standard test field. The data acquisition system synchronously records three-axis acceleration signals. The measuring points are arranged at 12 key positions such as the suspension tower top and the motor support. The characteristic parameters of each measuring point in the frequency range of 0-200 Hz are extracted, including the first 10 natural frequencies, modal damping ratios, characteristic frequency point vibration amplitudes, and phase difference matrices. Finally, a benchmark database containing vehicle models, measuring point positions, and characteristic parameters is constructed to provide a health state reference standard for subsequent vehicle detection.

[0119] The structural health assessment model performs multi-dimensional vibration characteristic analysis on the vibration characteristic vector, realizing real-time state assessment of the vehicle structure, component connection, and three-electricity system.

[0120] For example, the vehicle structure state assessment includes frequency deviation analysis, mode correlation analysis, and vibration energy distribution analysis. Among them, the frequency deviation analysis collects vibration data of each measuring point in real time, extracts the first 10 natural frequencies, calculates the relative deviation from the benchmark value, and triggers an abnormal vehicle structure warning when the relative deviation is greater than the deviation threshold. The mode correlation analysis quantifies the matching degree of the measured mode and the benchmark mode using the modal confidence index. When the modal confidence index is less than the confidence threshold, it is determined that the vehicle structure has a deformation risk. The vibration energy distribution analysis calculates the root mean square value of each measuring point in the characteristic frequency band, compares it with the benchmark value, and locates the vibration abnormal amplification area.

[0121] For example, the component connection state assessment includes transmission path abnormality identification, connection stiffness evaluation, and stress level monitoring. Among them, the transmission path abnormality identification traces the vibration energy transmission characteristics through transmission path analysis to identify abnormal paths. The connection stiffness evaluation calculates the coherence coefficient of the input and output signals at the connection point. When the coherence coefficient significantly decreases at the characteristic frequency, it is determined that the connection stiffness is insufficient. The stress level monitoring evaluates whether the dynamic stress of the connection part exceeds the material allowable value in combination with real-time strain gauge data.

[0122] For example, the three-electricity system mechanical stability assessment includes power battery system evaluation, drive motor system evaluation, and electric control system evaluation. The power battery system evaluation monitors the vibration of the battery pack mounting point in the frequency band of 200-500 Hz, and identifies the relative motion risk between modules through phase analysis. The drive motor system analyzes the vibration spectrum of the suspension point, monitors the rotation frequency and its harmonic amplitude, and identifies the electromagnetic force induced vibration using order analysis. The electric control system evaluation assesses the vibration response of the controller support and detects the contact reliability of the connector plug under vibration conditions.

[0123] Based on the real-time state evaluation results of the vehicle structure, component connection and three-electric system, a health degree evaluation system including multiple evaluation indexes is established, and the evaluation indexes include, but are not limited to, frequency deviation, mode change, coherence coefficient, transmission path change and characteristic frequency vibration level. The analytic hierarchy process is used to determine the weight of each index, and finally the comprehensive health index is calculated. The comprehensive health index is divided into four levels of excellent, good, medium and poor.

[0124] In S430, data fusion decision is made according to the abnormal sound type, abnormal sound position, comprehensive health state result and driving parameter of the target vehicle, and a detection result is obtained.

[0125] As a preferred embodiment of the present application, D-S evidence theory and Bayesian network algorithm are used to fuse the output results of the abnormal sound detection model and the structure health evaluation model, and the driving parameter of the target vehicle is combined to make data fusion decision, so as to comprehensively evaluate the overall quality of the target vehicle, generate detailed detection conclusion and repair suggestion, and obtain the detection result. The data fusion decision includes establishing a multi-source evidence system, D-S evidence theory fusion, Bayesian network reasoning and comprehensive evaluation and decision output.

[0126] The multi-source evidence system is established by first defining the identification framework and determining the complete set of vehicle quality state, including: abnormal sound state of no abnormal sound, slight abnormal sound and severe abnormal sound; structure state of intact structure, local loosening and structure damage; three-electric state of stable, slight abnormal and severe abnormal; and comprehensive quality level of excellent, good, qualified and unqualified.

[0127] Then, the evidence sources are configured. The abnormal sound detection model output is used as the first evidence source to provide the abnormal sound type identification result and its confidence, abnormal sound position positioning accuracy and abnormal sound severity score. The structure health evaluation model output is used as the second evidence source to provide the structure health index, key connection point stability evaluation and three-electric system mechanical stability score.

[0128] Finally, the driving parameters of the target vehicle are normalized, for example, the actual vehicle speed is mapped to the deviation coefficient of the standard detection speed to obtain the vehicle speed parameter, the deviation degree of the steering angle from the standard value is calculated to obtain the steering angle parameter, and the battery state parameters including state of charge, temperature and voltage consistency are obtained.

[0129] D-S evidence theory fusion firstly constructs basic probability assignment function. For abnormal sound evidence, basic probability is assigned according to confidence degree: m1(no abnormal sound) = confidence degree of no abnormal sound output by abnormal sound detection model; m1(mild abnormal sound) = confidence degree of mild abnormal sound recognition x position positioning accuracy; m1(severe abnormal sound) = confidence degree of severe abnormal sound recognition x position positioning accuracy; the remaining probability is assigned to uncertain set. For structure evidence, basic probability is assigned based on comprehensive health index: m2(structure intact) = comprehensive health index / 100 x connection stability coefficient; m2(local loosening) = (1-comprehensive health index / 100) x specific frequency deviation weight; m2(structure damage) = vibration energy anomaly coefficient x modal confidence degree deviation; the remaining probability is assigned to uncertain set.

[0130] Then, multiple evidence sources are combined by using Dempster combination rule. Firstly, conflict factor K is calculated to measure the consistency between evidences. When conflict factor K is less than 0.5, Dempster standard combination formula is used, and when K is greater than or equal to 0.5, weighted conflict redistribution strategy is used to reassign the conflict probability to each hypothesis according to evidence credibility, so as to avoid distortion of the combination result.

[0131] Finally, decision rules are formulated. Support degree decision selects the hypothesis with the maximum basic probability assignment, confidence interval decision calculates the confidence interval and likelihood interval of the hypothesis, and threshold-based decision sets the minimum support threshold of the quality level.

[0132] Bayesian network reasoning firstly defines network structure. Root nodes include abnormal sound state, structure state and three-electricity state, intermediate nodes include chassis, vehicle body and three-electricity system and other subsystem quality states, and leaf nodes are overall quality level.

[0133] Then, conditional probability table is constructed. Based on historical detection data, conditional probability between states is calculated. For example, P(overall quality = excellent | abnormal sound state = no abnormal sound, structure state = structure intact) = 0.95; P(overall quality = good | abnormal sound state = mild abnormal sound, structure state = local loosening) = 0.85; P(overall quality = unqualified | abnormal sound state = severe abnormal sound, structure state = structure damage) = 0.98.

[0134] Then, evidence propagation and updating are performed. D-S fusion result is input as prior probability into Bayesian network, and conditional probability is adjusted in combination with real-time vehicle driving parameters. For example, when vehicle speed deviation is greater than 10%, confidence degree weight of structure state is reduced; when steering angle deviation is greater than 15%, prior probability of abnormal sound state is adjusted; when battery temperature is abnormal, uncertainty of three-electricity state is increased.

[0135] Finally, the probability inference calculation is performed, the joint tree algorithm is used for accurate inference, and the posterior probability distribution is calculated, such as P (overall quality grade | all detection evidence) and P (specific fault reason | quality grade unqualified).

[0136] The comprehensive evaluation and decision output first calculates a comprehensive quality score, and the calculation formula is as follows: comprehensive quality score = D-S support degree x first weight + Bayesian posterior probability x second weight, wherein the first weight and the second weight are dynamically adjusted according to the evidence conflict degree.

[0137] Then, fault positioning and maintenance suggestions are generated, the most possible fault reason is determined based on the maximum posterior probability, the accurate maintenance position is generated in combination with the abnormal sound positioning result and the structure abnormal position, and the maintenance priority is recommended according to the fault severity.

[0138] Finally, a detection report is generated, including overall quality grade evaluation, detailed evaluation results of each subsystem, specific fault diagnosis conclusion, maintenance suggestions and process improvement suggestions and other detection results.

[0139] For example, the abnormal sound detection model analyzes and finds that the target vehicle left rear door area exists abnormal vibration and friction sound with a frequency of 500Hz-800H, judges that the abnormal sound is caused by loose door hinge, and locates the abnormal sound source at the left rear door and the body connection. The structure health assessment model analyzes the body vibration data and does not find other structure abnormalities. After the two model results are integrated by the data fusion decision, the detection result is generated, which clearly points out the problem of loose left rear door hinge, and suggests tightening the hinge screw.

[0140] As a further embodiment of the present application, the detection result and the original detection data can also be uploaded to the cloud platform. A Hadoop distributed file system HDFS and a NoSQL database are used to build a mass detection data storage cluster. The original detection data, model analysis results and detection results are stored according to the vehicle VIN code, detection time, vehicle type and other dimensions, supporting efficient storage and fast retrieval of PB-level data.

[0141] Optionally, the historical detection data is deeply mined by using a Spark big data processing framework. Through clustering analysis and association rule mining algorithms, the vehicle abnormal sound occurrence rule and quality defect mode are summarized. At the same time, based on the model analysis result, the parameters of the abnormal sound detection model and the structure health assessment model are automatically optimized, and the model detection accuracy and generalization ability are improved.

[0142] Optionally, the managers and engineers can remotely log in the cloud platform through the web or mobile terminal to monitor the target vehicle detection process and results in real time. When serious abnormal noise or potential failure is detected, the system automatically pushes the alarm information and provides the failure diagnosis report. Multi-department collaborative decision-making is supported, and the production department can optimize the assembly process according to the detection data, and the R&D department can obtain vehicle performance feedback to improve the design scheme. For example, if the production manager checks the detection results through the remote monitoring system and finds that the vehicle has abnormal noise problem, the system automatically generates a repair work order and pushes it to the repair workshop. The repair personnel tighten the left rear door hinge according to the report indication. After the repair is completed, the target vehicle enters the target road section detection area again for re-inspection, and the re-inspection result shows that the abnormal noise problem of the target vehicle is eliminated, the detection is qualified, and the vehicle is allowed to enter the sales link.

[0143] The cloud platform is used to realize centralized management and deep analysis of detection data, big data mining and model optimization are used to provide data support for production process improvement and product design optimization, and the vehicle quality is continuously improved. At the same time, the remote monitoring and collaborative decision-making system supports multi-department real-time sharing of detection information, realizes rapid response and collaborative processing of failure, shortens the problem solving cycle, and improves the production management efficiency.

[0144] Based on the same technical concept, the embodiment of the present application also provides a vehicle off-line detection device. Figure 6 is a structural schematic diagram of the vehicle off-line detection device provided by the embodiment of the present application, as Figure 6 shown, the vehicle off-line detection device 200 comprises: The path planning module 210 is configured to perform path planning according to the three-dimensional terrain data of the target road section and the detection requirements of the target vehicle, and obtain a target driving path and a speed curve.

[0145] The data acquisition module 220 is configured to control the target vehicle to pass through the target road section based on the target driving path and the speed curve, and acquire vibration data and sound data of the target vehicle in the driving process.

[0146] The data processing module 230 is configured to perform data preprocessing on the vibration data and the sound data, and obtain a target feature vector.

[0147] The vehicle detection module 240 is configured to input the target feature vector into a target detection model for analysis, and obtain a vehicle detection result.

[0148] The vehicle off-line detection device provided by the embodiment of the present application can efficiently and accurately locate the abnormal noise problem by collecting vibration data and sound data of the vehicle in real time, comprehensively and accurately when driving on a complex road section, and inputting the data after data processing into a target detection model for deep analysis. The health of the vehicle structure and the three-electric system is evaluated, and the quality and efficiency of the vehicle off-line detection are effectively improved.

[0149] It can be understood that the implementation of the vehicle offline detection method in the above embodiments is also applicable to the embodiments of the present application and can achieve the same technical effects, and therefore will not be described again.

[0150] Based on the same concept, the embodiments of the present application also provide an electronic device, Figure 7 is a structural schematic diagram of an electronic device provided by the embodiments of the present application, as Figure 7 shown, the electronic device 300 can include a processor 310, a communication interface 320, a memory 330 and a communication bus 340, wherein the processor 310, the communication interface 320 and the memory 330 complete communication with each other through the communication bus 340. The processor 310 can call the logical instructions in the memory 330 to execute the steps of the vehicle offline detection method in the above embodiments. For example, it includes: S100, path planning according to the three-dimensional terrain data of the target section and the detection requirements of the target vehicle, obtaining the target driving path and the speed curve; S200, controlling the target vehicle to pass through the target section based on the target driving path and the speed curve, and obtaining the vibration data and the sound data of the target vehicle in the driving process; S300, data preprocessing of the vibration data and the sound data, obtaining the target feature vector; S400, inputting the target feature vector into the target detection model for analysis, obtaining the vehicle detection result.

[0151] The processor 310 can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Chips, or a combination of the above various types of chips.

[0152] Further, the logic instructions in the memory 330 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art or parts of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0153] The memory 330 can include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required by a function; the data storage area can store data created by the processor and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory disposed remotely with respect to the processor, and these remote memories can be connected to the processor through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0154] Based on the same idea, the embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the computer program includes at least one code, which can be executed by a host device to control the host device to implement the steps of the vehicle offline detection method in the above-mentioned embodiments. For example, it includes: S100, path planning is performed according to three-dimensional terrain data of a target road section and detection requirements of a target vehicle, to obtain a target driving path and a speed curve; S200, the target vehicle is controlled to pass through the target road section based on the target driving path and the speed curve, and vibration data and sound data of the target vehicle in a driving process are acquired; S300, data preprocessing is performed on the vibration data and the sound data, to obtain a target feature vector; S400, the target feature vector is input into a target detection model for analysis, to obtain a vehicle detection result.

[0155] Based on the same technical concept, the embodiments of the present application further provide a computer program for implementing the above method embodiments when executed by a host device. The computer program can be stored in whole or in part on a computer readable storage medium packaged together with the processor, or stored in part or in whole on a memory not packaged together with the processor.

[0156] Based on the same technical concept, the embodiments of the present application further provide a processor for implementing the above method embodiments. The processor can be a chip.

[0157] Based on the same technical concept, the embodiments of the present application further provide a computer program product including computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor reads and executes the computer instructions from the computer readable storage medium to implement all or part of the steps of the methods shown in the above embodiments of the present application.

[0158] In summary, the vehicle off-line detection method, device, equipment, storage medium and program product provided by the present application realize the automation and standardization of the vehicle off-line detection process through automatic path planning combined with automatic driving, eliminate the influence of manual driving operation differences on the detection results, and different batches and different drivers participating in the road test can follow a unified and accurate process, greatly improving the consistency and reliability of the detection results. Through real-time, comprehensive and accurate collection of vibration data and sound data of the vehicle when driving on complex road sections, after data processing, the target detection model is input for deep analysis, which can efficiently and accurately locate the abnormal sound problem, and the health of the vehicle structure and the three-electric system is evaluated, effectively improving the quality and efficiency of the vehicle off-line detection.

[0159] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. As will be apparent to those of ordinary skill in the art, embodiments described herein can be combined with other embodiments.

[0160] The above-described embodiments are merely representative of several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the patent of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

[0161] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the same; although the present application has been described in detail with reference to the foregoing examples, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for vehicle off-line inspection, characterized in that, The method includes: Based on the three-dimensional terrain data of the target road segment and the detection requirements of the target vehicle, path planning is performed to obtain the target driving path and speed curve; Based on the target driving path and the speed curve, the target vehicle is controlled to pass through the target road segment, and vibration and sound data of the target vehicle are acquired during the driving process; The vibration data and the sound data are preprocessed to obtain the target feature vector; The target feature vector is input into the target detection model for analysis to obtain the vehicle detection result.

2. The vehicle off-line inspection method according to claim 1, characterized in that, The method further includes: The system acquires environmental perception information and vehicle status information of the target vehicle during its passage through the target road segment, and dynamically adjusts the driving path of the target vehicle based on the environmental perception information and vehicle status information.

3. The vehicle off-line inspection method according to claim 1, characterized in that, The process of path planning based on the three-dimensional terrain data of the target road segment and the detection requirements of the target vehicle to obtain the target driving path and speed curve includes: A three-dimensional digital terrain model is constructed based on the three-dimensional terrain data, and constraints on the target vehicle are set according to the detection requirements. A global path search is performed based on the aforementioned three-dimensional digital terrain model to obtain the initial driving path; Based on the initial driving path and the constraints, path smoothing and verification are performed to obtain the target driving path; Based on the target driving path, model predictive control is performed to generate the speed curve.

4. The vehicle off-line inspection method according to claim 1, characterized in that, The step of preprocessing the vibration data and the sound data to obtain the target feature vector includes: The vibration data is subjected to multi-level filtering, frequency band separation, and feature extraction to obtain a vibration feature vector; The sound data is subjected to noise suppression, acoustic feature enhancement, and feature extraction to obtain a sound feature vector; The vibration feature vector and the sound feature vector are fused and their dimensions reduced to construct the target feature vector.

5. The vehicle off-line inspection method according to claim 4, characterized in that, The target detection module includes at least an abnormal noise detection model and a structural health assessment model. The step of inputting the target feature vector into the target detection model for analysis to obtain vehicle detection results includes: The abnormal noise detection model is used to perform feature analysis on the target feature vector to obtain the type and location of the abnormal noise. The vibration feature vector is analyzed by the structural health assessment model to obtain the comprehensive health status results of the vehicle structure, component connections and the three-electric system. The detection result is obtained by performing data fusion decision based on the type of abnormal noise, the location of the abnormal noise, the comprehensive health status result, and the driving parameters of the target vehicle.

6. The vehicle off-line inspection method according to claim 2, characterized in that, The step of dynamically adjusting the driving path of the target vehicle based on the environmental perception information and vehicle status information includes: The environmental perception information and the vehicle state information are subjected to state encoding and feature extraction to obtain encoded features; The encoded features are input into a reinforcement learning policy network for inference calculation to obtain the initial target action; The initial target action is smoothed, its physical constraints are verified, and its safety boundaries are checked to obtain the target action. The target action is converted into a control command, and the control command is sent to the actuator of the target vehicle to adjust the driving path.

7. A vehicle off-line inspection device, characterized in that, The device includes: The path planning module is used to plan the path based on the three-dimensional terrain data of the target road segment and the detection requirements of the target vehicle, and obtain the target driving path and speed curve. The data acquisition module is used to control the target vehicle to pass through the target road segment based on the target driving path and the speed curve, and to acquire the vibration data and sound data of the target vehicle during the driving process; The data processing module is used to preprocess the vibration data and the sound data to obtain the target feature vector; The vehicle detection module is used to input the target feature vector into the target detection model for analysis and obtain vehicle detection results.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the vehicle off-line inspection method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the vehicle off-line detection method as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium, which are read and executed by a processor of a computer device to implement the vehicle off-line inspection method as described in any one of claims 1-6.

Images