A multi-axis acceleration time sequence feature road bump intelligent sensing and identification method

Abstract

The present application relates to the field of vehicle-mounted road condition intelligent sensing technology, and discloses a road bumping intelligent sensing and identification method based on multi-axis acceleration time sequence characteristics, which comprises the following steps: collecting vibration data stream of an acceleration sensor and synchronous vehicle speed, obtaining multi-channel motion signals through time alignment and adaptive band-stop denoising; estimating the gravity direction and heading by using quasi-static low-frequency components and horizontal acceleration principal components, and reconstructing a motion coordinate system to obtain attitude decoupling acceleration; performing spatial domain resampling driven by vehicle speed to obtain spatial domain multi-channel waveforms, separating the maneuvering context according to the wavelength, removing the vertical coupling thereof by using recursive least squares, and obtaining pure road excitation spatial residuals; inputting the spatial residuals into a lightweight time sequence network trained through speed conditioning modulation by vehicle speed and cross-speed consistency constraint, and outputting event types, bumping grades and confidence, and mapping the output into roughness values. The present application can realize cross-vehicle consistent road bumping identification and grading under the conditions of unknown installation attitude and time-varying vehicle speed.

Description

Technical Field

[0001] This invention relates to the field of vehicle-mounted road condition intelligent sensing technology, and more specifically, to a method for intelligent sensing and recognition of road bumps based on multi-axis acceleration temporal characteristics. Background Technology

[0002] With the increasing demand for road surface data from intelligent transportation and road maintenance, crowdsourced road surface data collection using low-cost accelerometer terminals mounted on private vehicles or inspection vehicles has become a practical approach to obtaining high-coverage, routine road surface defects and smoothness data. The collected results can be used for maintenance work order generation, repair priority scheduling, and vehicle suspension adjustment. These data collection terminals are mounted on different vehicle models at varying locations and angles, and the vehicle speed fluctuates continuously between congested crawling and normal traffic, while the journey also includes braking, acceleration, steering, and driving on ramps and bridges.

[0003] Most existing vibration-related road surface recognition solutions use single-axis acceleration thresholds for judgment or determine bumps in the time domain based on fixed thresholds, and simply eliminate vibrations caused by driving operations as interference. Some solutions use images or inertial inversion as auxiliary methods, and generally assume that the installation orientation of the sensing unit is known or has been calibrated.

[0004] However, the aforementioned existing technologies have the following technical problems: they assume the sensing orientation is known, ignoring the unknown installation angle and gravity leakage caused by time-varying road slopes, resulting in slopes and inclined installations being falsely reported as bumps; they judge based on fixed thresholds in the time domain, failing to account for the stretching or compression of waveforms of the same defect at different vehicle speeds, leading to inconsistent recognition results across vehicle speeds and across vehicles; they treat driving operation stimuli and road surface stimuli as easily separable, making it difficult to handle the superposition of braking, steering, and road impacts, thus causing misjudgments, and they cannot output quantitative grading that can be compared across vehicles, making them unsuitable for supporting the intelligent transportation system's need to collect standardized road surface data. Summary of the Invention

[0005] This invention provides an intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics, which solves the technical problem in related technologies that it is difficult to obtain consistent recognition and classification of the same road surface defects that are independent of vehicle speed and attitude and can be compared across vehicles under the conditions of unknown installation posture, time-varying vehicle speed and superposition of driving operation excitation and road surface excitation.

[0006] This invention provides an intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics, comprising the following steps: S1 collects vibration data streams from the accelerometer and synchronous vehicle speed, and obtains multi-channel motion signals and vehicle speed through time alignment and adaptive band-stop denoising. S2, based on multi-channel motion signals, uses quasi-static low-frequency components and horizontal acceleration principal components to estimate the gravity direction and heading and reconstruct the motion coordinate system, and projects to obtain the attitude decoupled acceleration in the three directions of road vertical, vehicle longitudinal and vehicle lateral. S3, the attitude decoupled acceleration is resampled in the spatial domain driven by vehicle speed, and converted into a spatial domain multi-channel waveform arranged at fixed intervals along the driving distance, and the driving distance and the passing speed of each spatial sample are obtained. S4. Separate the slowly varying maneuvering context from the longitudinal and transverse channels of the multi-channel waveform in the spatial domain according to the wavelength. After deducting the vertical gravity bias, remove the vertical coupling of the maneuvering context by recursive least squares to obtain the spatial residual of pure road excitation and the characteristics of the maneuvering context. S5 feeds the residuals of the pure road excitation space and the features of the motion context into a lightweight temporal network, so as to perform speed conditional modulation by vehicle speed and train through cross-speed consistency constraints, and output the road event type, bump level, confidence level, and the proportion of residual frequency band energy and dominant wavelength features of each event. S6, based on road event type, bump level, confidence level, residual frequency band energy ratio and dominant wavelength characteristics, combined with driving mileage, maps to roughness values ​​with the same dimensions as the International Roughness Index, and outputs time-series road condition datasets and road condition recognition records in the order of driving.

[0007] In a preferred embodiment, the time-aligned and adaptive band-stop denoising includes estimating the power spectrum distribution of the multi-channel motion signal online using a sliding window periodogram, identifying spectral peaks that shift with vehicle speed and have narrow bandwidth as tire noise-related components, and identifying spectral peaks with fixed positions as mechanical noise-related components. A bandstop is applied to the above spectral peaks while preserving the mid-to-low frequency band where the road surface excitation is located, and the depth and width of the bandstop are adaptively set according to the bandwidth of the spectral peaks.

[0008] In a preferred embodiment, the step of estimating the gravity direction and heading and reconstructing the motion coordinate system using quasi-static low-frequency components and horizontal acceleration principal components includes extracting quasi-static low-frequency components and updating the unit vector estimate of the gravity direction in a recursive manner with a forgetting factor, and assigning smaller confidence weights to samples whose magnitude of the composite force deviates more from the gravitational constant. The vehicle longitudinal direction corresponding to the heading is estimated from the principal component directions of the two acceleration components in the horizontal plane. The update of the vehicle longitudinal direction is frozen in the uniform speed segment and the pure turning segment, and the previous estimate is used. Using the opposite direction of gravity as the vertical axis of the road surface and the longitudinal direction of the vehicle as the longitudinal axis, the vehicle's lateral direction is obtained by cross product of the two after orthogonal normalization, thus forming a motion coordinate system.

[0009] In a preferred embodiment, the spatial domain resampling of attitude decoupling acceleration driven by vehicle speed includes using the cumulative mileage of wheel speed pulses as the driving mileage when the vehicle provides wheel speed pulses, and integrating the vehicle speed over time to obtain the driving mileage when there are no wheel speed pulses. Each attitude decoupled acceleration sample is re-indexed from the time index to the mileage index, and then resampled along the mileage axis at fixed spatial intervals to obtain a spatial domain multi-channel waveform arranged at fixed intervals along the travel distance. The spatial interval is determined according to the minimum defect spatial wavelength to be identified.

[0010] In a preferred embodiment, the interpolation resampling is a variable step size adaptive interpolation, which takes points sparsely in time in high-speed segments and densely in low-speed segments to maintain a fixed spatial interval, and uses piecewise cubic interpolation along the mileage axis. For the drift of driving mileage, when there are equally spaced artificial structures, the spatial waveform is autocorrelated and the cumulative mileage is proportionally corrected by the peak spacing of its regular peaks. When there are no equally spaced artificial structures, the local speed ratio is normalized within the sliding window. The smaller the confidence weight is assigned to spatial samples with instantaneous vehicle speeds less than a preset speed threshold, the less new spatial samples are generated when the vehicle stops and the mileage no longer increases.

[0011] In a preferred embodiment, separating the slowly changing maneuvering context from the longitudinal and transverse channels of the spatial domain multi-channel waveform by wavelength includes decomposing the longitudinal and transverse channels by spatial wavelength using zero-phase bidirectional filtering, taking the continuous component with a wavelength greater than the cutoff wavelength as the maneuvering context, and taking the component with a wavelength less than the cutoff wavelength as the fast-changing road surface excitation. The cutoff wavelength is determined based on the largest spatial scale of the disease to be identified, and a transition zone is set near the cutoff wavelength. Components falling into the transition zone are weighted according to their distance from both sides.

[0012] In a preferred embodiment, the step of removing the vertical coupling of the maneuvering context by recursive least squares after deducting the vertical gravity bias includes filtering out the quasi-static gravity bias of the ultra-long wavelength in the vertical channel with a zero-phase high-pass filter to obtain the debiased vertical component, and then estimating the decoupling coefficient online with recursive least squares with a forgetting factor, so that the energy in the debiased vertical component that is linearly related to the maneuvering context tends to be minimized to obtain the pure road surface excitation space residual. When estimating the decoupling coefficients, only the slowly varying maneuvering context is used as the regression quantity. Short-wavelength road impacts are not included in the estimation. An upper limit is set on the single-step update amplitude of the decoupling coefficients, and the update is paused when the maneuvering context energy is less than a preset threshold.

[0013] In a preferred embodiment, the lightweight temporal network takes the multi-channel sequence of the pure road excitation space residual as the main input and the motor context features as the auxiliary input. Before the pure road excitation space residual enters the network, it is centered in each sliding window by the median and the amplitude is normalized by the absolute deviation of the median. The network extracts multi-scale features layer by layer using depthwise separable one-dimensional convolutions. The depthwise separable one-dimensional convolutions consist of channel-wise convolutions that extract local waveforms along the spatial dimension and pointwise convolutions that are linearly combined between channels. Additive temporal attention is used to calculate scores for each spatial location and normalize them into attention weights before weighted convergence. When calculating the scores, the dynamic context features and the features of each spatial location are concatenated and then used. The velocity conditional modulation maps the passing speed of each spatial sample to a channel-wise scaling factor and offset factor through a multilayer perceptron. The convolutional features are first multiplied by the scaling factor and then the offset factor is added.

[0014] In a preferred embodiment, the training objective of the cross-speed consistency constraint training is composed of the cross-entropy of the event type, the ordered regression loss of the bump level and the weighted sum of the cross-speed consistency loss. The ordered regression loss transforms the determination of the bump level into the sum of the cumulative binary determinations of whether each level exceeds each boundary. The cross-speed consistency loss constrains the network intermediate features corresponding to different vehicle speeds on the same road segment to move closer to each other. The sample pairs that constitute the cross-speed consistency loss are obtained from the same road segment collected at different vehicle speeds and aligned by mileage, or from the same road segment time domain signal resampled at different vehicle speeds and then sequentially subjected to attitude decoupling, spatial domain resampling and removal of vertical coupling of the maneuver context.

[0015] In a preferred embodiment, the mapping to a roughness value with the same dimensions as the International Roughness Index includes mapping the roughness level and the residual frequency band energy ratio and dominant wavelength characteristics of each event, which are fitted on the marked road section, to a roughness value with the same dimensions as the International Roughness Index, and using the road event type as the semantic label of the roughness value. The duration of each event is obtained by multiplying its start and end span on the spatial index by the spatial interval. Records of multiple vehicles passing through the same road segment are first aligned by location and then finely aligned by cross-correlation of spatial waveforms. The time-series traffic data set is arranged by event and in ascending order of mileage. Recognition results with confidence levels less than a preset threshold are marked with quality labels.

[0016] The beneficial effects of this invention are as follows: By reconstructing the attitude reference system online and performing spatial domain resampling of the attitude decoupled acceleration driven by vehicle speed, the gravity leakage introduced by the unknown installation angle and road slope is separated. Even if the installation angle is unknown and the journey includes going up and down bridges and slopes, the acquisition terminal will not misreport the slope as bumps. Furthermore, the time domain waveform that expands and contracts with vehicle speed is normalized to a spatial waveform arranged along the travel distance. This allows the same physical defect to obtain a consistent spatial representation under different vehicle speeds, different installation postures, and different vehicles. As a result, the road condition data collected by crowdsourcing has a unified physical scale, which facilitates cross-vehicle aggregation. By separating the maneuver context by wavelength and removing its vertical coupling using recursive least squares, the model retains short-wavelength pothole impacts while deducting gradually changing operational couplings when braking, acceleration, steering, and road impacts coexist. The pure road excitation spatial residuals are extracted, and then a lightweight temporal network trained with vehicle speed conditional modulation and cross-speed consistency constraints is used to complete the identification and classification. This suppresses the velocity-related artifacts remaining after spatial domain resampling and outputs roughness values ​​with the same dimensions as the International Roughness Index that can be compared across vehicles. The identification results are accompanied by confidence and quality labels for easy weighted fusion in the cloud. The model is lightweight and easy to deploy online on embedded terminals, making it suitable for downstream applications such as road defect inspection and semi-active suspension preview. Attached Figure Description

[0017] Figure 1 This is a structural diagram of a road bump intelligent sensing and recognition method based on multi-axis acceleration time-series characteristics according to the present invention. Figure 2 This is a flowchart of an intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics, according to the present invention. Figure 3 This is a flowchart of a method for intelligent sensing and recognition of road bumps based on multi-axis acceleration temporal characteristics according to the present invention. Figure 1 ; Figure 4 This is a flowchart of a method for intelligent sensing and recognition of road bumps based on multi-axis acceleration temporal characteristics according to the present invention. Figure 2 . Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses an intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics, such as... Figure 1 The diagram shows the overall structure of the method of this invention. In the diagram, the vibration excitation source ⑦ of urban and national / provincial highways is collected by crowdsourcing vehicles ⑥ and sensed by the vehicle. A MEMS multi-axis motion sensor ③ provides three-axis or six-axis acceleration and angular velocity, a vehicle speed interface or wheel speed pulse ④ provides synchronous vehicle speed, and an optional satellite positioning module ⑤ provides positioning information. All of the above signals are sent to the vehicle-mounted embedded computing terminal ②. The terminal locally completes the online identification process in the following order: S1 acquisition alignment and adaptive bandstop, S2 attitude reconstruction and three-dimensional decoupling, S3 vehicle speed-driven spatial resampling, S4 vertical decoupling of the maneuver context, S5 temporal recognition and bump classification, and S6 roughness mapping and temporal output. The structured temporal road condition dataset, containing event type, bump level, roughness, confidence, location, and quality markers, is uploaded to the cloud-based road condition aggregation system ①, and then distributed to support downstream applications such as maintenance work order generation, maintenance priority scheduling, and semi-active suspension pre-aiming ⑧. Specific steps are as follows: Figures 2 to 4 As shown, it includes: S1 collects vibration data streams from the accelerometer and synchronous vehicle speed, and obtains multi-channel motion signals and vehicle speed through time alignment and adaptive band-stop denoising. Specifically, acceleration sensors from the microelectromechanical system (MEMS) installed on the vehicle are collected. A three-axis configuration collects acceleration in three orthogonal directions of vehicle motion, while a six-axis configuration adds three angular velocities as input for a preferred embodiment. Vehicle speed information is also acquired synchronously with the motion signals. The sources of this information are categorized into two types based on availability: if the vehicle can provide wheel speed pulses or a vehicle speed interface, this is used; otherwise, it is estimated and supplemented by longitudinal motion information in subsequent steps. In addition to speed, the terminal can optionally access a satellite positioning module. Its positioning information is used as an optional input and aligns with a unified time base for location binding in subsequent steps. Without satellite positioning, the location is calculated using mileage and the starting point of the journey. Since the motion channel sampling clock and the speed reporting clock are often inconsistent, time alignment is performed first, resampling data from different sources to the same time base, ensuring that each time sample simultaneously has three or six motion components and a speed value. In a six-axis configuration, the three angular velocities and accelerations share the same time base alignment and are also used for attitude estimation and complementary filtering. Narrowband interference from the angular velocity itself is suppressed during band-stop processing. Regarding the determination of the sampling frequency, the main frequency band range of the road surface excitation to be identified is first analyzed, and then a sampling frequency higher than twice the upper limit of the frequency band is selected according to the Nyquist sampling theorem to ensure that transient components such as pothole impacts are not undersampled. For ease of engineering implementation, for example, the default value of the acceleration sampling frequency in this embodiment is 100 Hz, and the default value of the velocity reporting frequency is 10 Hz. The two are unified to the acceleration time base after time alignment. The output of this step is a multi-channel original sequence aligned sample by sample. In one embodiment of the present invention, the original sequence contains three components: engine mechanical noise, tire noise, and road excitation. The mechanical noise and tire noise are often concentrated on several narrow bands that vary with engine speed and vehicle speed, while the road excitation energy is distributed in the low-to-mid frequency band related to vehicle speed and the size of the road surface. The multi-channel sequence is subjected to adaptive band-stop processing. First, the power spectrum distribution of the current signal is estimated online to locate the narrow band peaks related to engine speed and tire rotation. Then, band-stop suppression is applied to these narrow bands while retaining the low-to-mid frequency road surface excitation band. The center frequency of the band-stop is dynamically adjusted with engine speed and vehicle speed to avoid false suppression of the road surface excitation signal by the fixed filter when the operating conditions change. The online power spectrum estimation uses a sliding window periodogram. Discrete spectrum transformation is performed on the multi-channel signal within each window, and smoothing is performed between windows to obtain a spectrum distribution that is constantly updated with driving. The location of narrowband noise peaks does not depend on the engine speed interface, but rather on the sharpness of the peaks and the pattern of their translation with vehicle speed. Peaks that translate approximately linearly with vehicle speed and have narrow bandwidth are identified as tire noise-related components, while sharp peaks with relatively fixed positions are identified as mechanical noise-related components. Bandstops are applied to both, and the mid-to-low frequency road excitation band is preserved. The depth and width of the bandstop are adaptively set according to the sharpness of the peaks. The sharper the peak, the narrower the bandstop, so as not to cause false suppression of the excitation energy of the adjacent road surface. It should be noted that no full amplitude clipping is performed, nor are any road condition criteria set. Only time base unification and narrowband noise suppression are completed. The purpose is to provide undistorted quasi-static components and clean dynamic components for subsequent attitude estimation. The output of this step is a multi-channel motion signal with time axis alignment and suppression of mechanical and tire noise, along with sample-by-sample aligned velocity and optional positioning information.

[0020] S2, based on multi-channel motion signals, uses quasi-static low-frequency components and horizontal acceleration principal components to estimate the gravity direction and heading and reconstruct the motion coordinate system, and projects to obtain the attitude decoupled acceleration in the three directions of road vertical, vehicle longitudinal and vehicle lateral. Specifically, the multi-channel motion signal output from the previous step is received to eliminate gravity leakage caused by unknown installation angle and changes in road slope with travel distance. When the vehicle is traveling smoothly, the low-frequency quasi-static part of the motion signal mainly reflects the projection of gravity on the sensor body coordinate system, based on which the spatial orientation of the sensor relative to gravity can be inferred. First, the quasi-static component is extracted with a low-pass filter, and then the unit vector estimate of the gravity direction is updated online in a recursive manner with a forgetting factor, so that the estimate can track the slow change of slope without deviating due to single-frame noise. The cutoff frequency of the low-pass filter is selected below the lower limit of the main frequency band of the road excitation, so that the slowly changing gravity projection caused by slope and installation angle can pass through while the road excitation is blocked from the quasi-static component. If the cutoff frequency is too low, the gravity estimation is more stable but the tracking of slope change is slower. If it is too high, the tracking is faster but the road excitation is easily mixed in. Its value is determined by the spatial scale of common slope changes and the time scale converted from the driving speed. In the recursive update, a transient confidence weight is introduced. The further the magnitude of the synthesized specific force deviates from the gravitational constant, the more it indicates that the vehicle is in a transient state of acceleration / deceleration or crossing a pothole. The less valid the quasi-static assumption is, the lower the weight of that frame. The weight increases when the vehicle approaches a constant speed and travels smoothly. The initial value of the gravity direction estimation is determined in the first stationary or constant speed straight-line segment after the terminal is powered on. The stationary segment is determined by the three-axis synthesized specific force being close to the gravitational constant and the angular velocity being close to zero. At this time, the quasi-static component is the initial value of the gravity direction. Thereafter, the gradient change is tracked recursively. If the vehicle is in a driving state immediately after power-on, a coarse initial value is obtained by short-term low-pass filtering. Convergence is resumed when a reliable constant speed segment appears. When there is a situation where the quasi-static assumption is not valid for a long time, such as continuous acceleration or uphill driving, the gravity direction estimation is paused in the pure three-axis configuration and the most recent reliable estimate is used. In the six-axis configuration, the attitude continuity is maintained by angular velocity integration during this period. Convergence is resumed when a reliable constant speed segment is restored, so as not to lose the attitude reference due to long transients. Gravity direction alone can only determine one vertical axis; heading is also needed to construct a complete rotation matrix. This is addressed separately in pure three-axis and six-axis configurations. In the preferred six-axis embodiment, angular velocity integral gives short-term attitude change, which is fused with gravity direction given by acceleration through complementary filtering. Gravity direction is used in the low-frequency band, and angular velocity integral is used in the high-frequency band, directly obtaining the complete attitude including heading. In the pure three-axis embodiment, heading is estimated by the principal component directions of the two acceleration components in the horizontal plane. Specifically, a recursive covariance with a forgetting factor is maintained for the two components in the horizontal plane, and the eigenvector corresponding to its largest eigenvalue is taken as the longitudinal axis direction. Its physical meaning is the principal direction of horizontal acceleration during linear acceleration or braking, i.e., the vehicle's longitudinal direction. In the constant speed and pure turning segments, the principal direction of horizontal acceleration is unreliable. Therefore, the longitudinal axis is frozen and updated, and the previous reliable estimate is used. In the pure turning segment, the energy of the lateral continuous component is used for determination. Thus, even without a gyroscope, there is a definite heading solution for constant speed and turning conditions. The principal component analysis only provides the straight line containing the longitudinal axis without distinguishing between the positive and negative directions of the vehicle's front and rear. The disambiguation of the direction sign is based on the positive mileage increment when the vehicle is moving forward, i.e., assuming the positive direction of the longitudinal axis is consistent with the driving direction. In a six-axis configuration, the joint sign of angular velocity and acceleration is used for further verification. After the start of the journey, before the first reliable acceleration / deceleration event occurs, the longitudinal axis is coarsely determined based on the driving direction. The forward direction with positive velocity is taken as the coarse positive direction of the longitudinal axis. After the first reliable acceleration or deceleration event occurs, it is refined and fixed by the principal component direction. Thereafter, it is maintained according to the aforementioned freezing and updating logic. The time constant of the complementary filter determines the frequency division boundary between the gravity direction and the angular velocity integral. A longer time constant makes the gravity direction more reliable and resists angular velocity drift, while a shorter time constant tracks attitude changes faster. Its value is jointly determined by the sensor's zero bias stability and the rate of change of road slope. The low-pass extraction, recursion with forgetting factor, complementary filtering and principal component estimation involved in this step are all conventional methods in this field. Therefore, their data flow is described in words and conventional formulas are not listed. In one embodiment of the present invention, after obtaining the gravity direction and heading, a rotation matrix is ​​constructed from the sensor body coordinate system to the motion coordinate system. The motion coordinate system uses the opposite direction of gravity as the road vertical axis, the estimated heading as the vehicle longitudinal axis, and the lateral axis as the cross product of the vertical and longitudinal axes. The acceleration vector at each moment is multiplied by the rotation matrix, that is, the body reading mixed with installation angle and slope information is projected into three directional components with clear directional meanings. The angular velocity under the six-axis configuration is also projected accordingly to provide assistance for subsequent steering determination. When constructing the rotation matrix, the two directional vectors of the vertical and longitudinal axes are orthogonally normalized first, that is, the longitudinal axis is removed from its projection component on the vertical axis and then normalized. Then, the lateral axis is obtained by the cross product of the two axes. The three axes are unit orthogonal, thus ensuring that the rotation matrix is ​​an orthogonal matrix and avoiding scale distortion after projection. The attitude increment obtained by the integral of the angular velocity under the six-axis configuration is used to maintain the continuity of the rotation matrix between two quasi-static updates. After rotation, the quasi-static components caused by gravity and slope are incorporated into the gradually varying offset of the vertical axis. The rapidly varying fluctuations on the vertical axis correspond to the vertical excitation of the real road surface. The longitudinal and lateral axes respectively carry vehicle motion information such as braking acceleration and steering roll. It should be noted that only coordinate projection is performed here without deducting the gravity offset. The deduction of the gravity offset will be completed uniformly in the subsequent decoupling step to avoid introducing additional reference errors here. The output of this step is the decoupled acceleration of the road surface vertical, vehicle longitudinal, and vehicle lateral directions, and the sample-by-sample velocity and optional positioning information attached in the previous step are also transmitted. Regardless of how the sensors are installed or how the road is uneven, the signals entering the subsequent processing are three directions with a unified physical meaning.

[0021] S3, the attitude decoupled acceleration is resampled in the spatial domain driven by vehicle speed, and converted into a spatial domain multi-channel waveform arranged at fixed intervals along the driving distance, and the driving distance and the passing speed of each spatial sample are obtained. Specifically, based on the three-directional attitude decoupled acceleration and synchronization speed information output in the previous step, the problem of inconsistent time-domain waveforms of the same physical defect at different vehicle speeds is solved. This method treats road defects as spatial entities distributed along the road. The reason why the same pothole appears with different waveforms at different vehicle speeds due to equal-time sampling is that the travel distance corresponding to a unit of time changes with the vehicle speed. Therefore, it is necessary to reconstruct the mileage first. The mileage source is classified according to availability. When the vehicle provides wheel speed pulses, the cumulative mileage of the pulses is used as the standard, because the wheel speed pulse directly corresponds to the arc length rotated by the wheel and is not affected by acceleration integral drift. When there are no wheel speed pulses, it degenerates to integrating the vehicle speed along time to obtain the mileage. At this time, drift correction is performed as described later. The conversion from wheel speed pulses to mileage is obtained by multiplying the wheel angle corresponding to each pulse by the wheel rolling radius. The wheel rolling radius is determined by regression of known mileage road segments during the calibration stage and is slowly corrected with tire pressure. The mileage here is not directly given using satellite positioning. Firstly, the update rate of satellite positioning is too low to support centimeter-level spatial sampling. Secondly, satellite signals are prone to loss of lock between tunnels and tall buildings. Therefore, the mileage is primarily based on wheel speed pulses or vehicle speed, with satellite positioning only used for location binding in subsequent steps. After obtaining the mileage, each attitude decoupled acceleration sample is mapped from the time index to the mileage index, and then re-sampling is performed along the mileage axis at fixed spatial intervals to obtain a spatial domain sequence arranged at equal intervals according to the travel distance. Thus, a road surface defect of a given physical wavelength will occupy a consistent sample span and exhibit a similar waveform shape regardless of the vehicle's speed. The spatial sampling interval is determined by the minimum defect wavelength to be identified, and its value is chosen to allow spatial sampling to distinguish the contours of the smallest defect. For example, in this embodiment, the default spatial sampling interval is five centimeters. This value is first calculated from the minimum distinguishable defect wavelength using the sampling theorem, and then rounded using the terminal's computing power. The aforementioned mileage integration and interpolation resampling are routine operations, and their process is described in words. In one embodiment of the present invention, vehicle speed continuously changes during actual driving, and the time step corresponding to the resampling in the equally spaced spatial domain is not uniform. Mileage integrals also accumulate and drift at low speeds and idle speeds. Therefore, variable step-size adaptive interpolation is used: points are sparsely sampled in time when the vehicle speed is high and densely sampled when the vehicle speed is low, always maintaining equal spatial intervals. The density of sampling points is determined by the instantaneous vehicle speed and the mileage increment of adjacent samples. Interpolation along the mileage axis uses piecewise cubic interpolation, which maintains the waveform peak shape while suppressing the angle artifacts caused by linear interpolation. The interpolation nodes are the equally spaced positions on the mileage axis, and a piecewise method satisfying causal constraints is used instead of... Global frequency domain resampling is used to output segments of data without waiting for the entire data during streaming processing. Mileage drift correction follows two paths: when there are identifiable, equally spaced artificial structures along the route, such as expansion joints or manhole cover arrays, autocorrelation is first performed on the spatial waveform. The autocorrelation shows regular peaks at these equally spaced structures, and the peak spacing represents the spatial period of the reference object. Based on this, the cumulative mileage is proportionally corrected. When there are no such reference objects, the property that spatial domain resampling depends only on the local instantaneous velocity ratio and not on the absolute mileage is utilized to restrict the correction to a sliding window for local velocity ratio normalization, thus ensuring that long-range drift does not affect the consistency of the local spatial waveform. For low-speed samples, due to their high mileage integral noise, they are assigned low confidence weights and enter subsequent steps along with the signal. This allows these samples to be moderately weakened during identification rather than being directly discarded. The basis for determining low speed is that the instantaneous vehicle speed is lower than the speed threshold. When the vehicle is completely stopped, the speed is zero, the mileage no longer increases, and no new spatial samples are generated. Output continues only when the vehicle starts moving again and the mileage resumes growth, thus preventing vehicle vibration during idling from being mixed into the spatial sequence. The determination of low speed is given by the speed threshold. For example, in this embodiment, the default value of the low-speed threshold is taken as the common upper limit speed of congested crawling. This value is first determined by the relationship between mileage integral noise and vehicle speed, and then adapted in combination with the speed limit of the road section. The three directional channels are synchronously interpolated with the same mileage axis to ensure that the temporal correspondence between channels is not disrupted after resampling. The accumulated mileage is managed by trip. The mileage is reset to zero at the start of each trip and the starting point of the trip is taken as the spatial origin. The mileage increases monotonically within the trip to avoid the drift accumulation of cross trips being passed to the next trip. The trip boundary is determined by the long-term parking and engine-off states. It should be noted that the physical meaning of the signal is not changed; only the sampling independent variable is changed from time to distance. Therefore, the vertical, longitudinal, and lateral channels simultaneously complete spatial domain resampling, and the phase and amplitude relationship between channels is maintained. The output of this step is a spatial domain multi-channel waveform arranged at equal intervals according to the travel distance and does not drift with vehicle speed. This includes the confidence weight of each spatial sample, the passing speed of the sample, the mileage of the sample, and the optional positioning information. The passing speed refers to the instantaneous speed of the vehicle when it passes through the spatial location. The mileage and positioning information are continuously transmitted with the spatial sequence until the final output.

[0022] S4. Separate the slowly varying maneuvering context from the longitudinal and transverse channels of the multi-channel waveform in the spatial domain according to the wavelength. After deducting the vertical gravity bias, remove the vertical coupling of the maneuvering context by recursive least squares to obtain the spatial residual of pure road excitation and the characteristics of the maneuvering context. Specifically, the system receives the spatial domain multi-channel waveform and its confidence weights output from the previous step, and simultaneously receives the transmitted vehicle speed, mileage, and optional positioning information. The goal is to isolate the interference components coupled into the vertical channel by driving actions such as braking, acceleration, and steering. Braking and acceleration are mainly reflected in the continuous components of the longitudinal channel that change slowly with space, while steering and the resulting body roll are mainly reflected in the continuous components of the lateral channel. These two types of components differ in spatial form from short-wavelength road excitations such as potholes and impacts. Based on this, the longitudinal and lateral channels are decomposed according to spatial wavelength. Continuous components above the cutoff wavelength are classified as the driving context, and components below the cutoff wavelength are classified as rapidly changing road excitations. The cutoff wavelength is determined by the maximum spatial scale of the defect to be identified. The reason for separating by wavelength rather than by time is that after the previous spatial domain resampling step, the independent variable of the signal is now the driving distance. Driving actions exhibit a long span and gradual change in space, while road surface defects exhibit localized, energy-concentrated short-wavelength fluctuations. The two are more clearly distinguishable in the wavelength dimension. Sensors are generally not installed at the vehicle's center of gravity. During steering, the lever arm effect introduces centrifugal and tangential accelerations. These components gradually change with space during continuous steering and mainly fall within the lateral channel. Therefore, they are classified into the lateral maneuvering context and processed together in subsequent decoupling, and will not be mistakenly treated as road surface excitations. For ease of engineering implementation, for example, the default cutoff wavelength in this embodiment is taken as several times the width of a common speed bump. This value is first determined by the maximum spatial scale of the defect that needs to be retained, and then adapted in conjunction with the road segment type. The longitudinal and lateral continuous components are the maneuvering context, while the rapidly changing components, along with the vertical channel, are subsequently decoupled. Wavelength decomposition is achieved by zero-phase bidirectional filtering, which involves filtering the spatial sequence once in the forward direction and once in the reverse direction to prevent the filtering from introducing phase shifts, thus ensuring that the maneuvering context and the rapidly changing road surface excitation do not shift in spatial position. The cutoff wavelength is converted into the spatial frequency cutoff point through the spatial sampling interval, and the passband and stopband of the high-pass and low-pass filters are set accordingly. To avoid a sharp separation between the maneuvering context and the road surface excitation near the cutoff wavelength, a transition band is set between the high-pass and low-pass filters. The components in the transition band are weighted according to their distance from both sides. Components that fall within the transition band and change slowly with space are more likely to be assigned to the maneuvering context, while components that change rapidly with space are more likely to be assigned to the road surface excitation, thus making the separation at similar wavelengths smoother and preventing abrupt changes. In one embodiment of the present invention, after obtaining the maneuvering context, the quasi-static gravity bias of the vertical channel is first deducted, and then the component coupled into the vertical channel by the maneuvering context is deducted. The gravity component after attitude decoupling is manifested as a slowly varying DC component with an ultra-long wavelength in the vertical channel. Therefore, this ultra-long wavelength component is filtered out by a zero-phase high-pass filter consistent with the aforementioned wavelength separation, thus obtaining the debiased vertical component. Then, the maneuvering coupling is processed. In this way, the two steps are clear and do not interfere with each other. Since the degree of vertical coupling between vehicle pitch and roll changes with the vehicle state, the fixed coefficient cannot adapt to the changes in the strength of the superposition state. Therefore, recursive least squares (RLS) with a forgetting factor is used. The Squares method estimates decoupling coefficients online, aiming to minimize the energy linearly related to the maneuvering context in the vertical residuals after deduction, so that the residuals and maneuvering context tend to be orthogonal. When maneuvering is strong, the context energy is large, the coefficients are updated quickly, and more coupling is deducted. When maneuvering is weak, the update is slow and the coefficients are close to zero, so that the vertical signal is basically not modified, thus protecting the real road surface excitation during smooth driving. Recursive least squares is a conventional adaptive estimation method in this field. Its iterative process is described in words, that is, each time a new sample is received, the coefficients are updated using the current prediction error and the gain vector, and the covariance is updated recursively. Specifically, at each spatial sample, the vertical coupling component is first predicted using the current decoupling coefficients and the maneuvering context to obtain the prediction error of the sample. Then, the coefficients are corrected by multiplying the error by the gain vector. The gain vector is calculated from the covariance matrix of the previous step and the current maneuvering context vector. Then, the covariance matrix is ​​updated recursively according to the forgetting factor. The initial value of the covariance is a diagonal matrix with sufficiently large diagonal elements to make the initial update sensitive, and then converges with the accumulation of samples. The decoupled vertical residual is the pure road excitation spatial residual. At each spatial sample, its value is equal to the debiased vertical component minus the sum of the longitudinal kinematic context component multiplied by the longitudinal decoupling coefficient and the lateral kinematic context component multiplied by the lateral decoupling coefficient. The two decoupling coefficients are not preset constants, but are updated online at each spatial sample by the above recursive least squares. When updating the decoupling coefficients, the correlation between the residual energy and the maneuver context in the most recent samples is taken into account. The forgetting factor determines the memory length of historical samples. The closer the forgetting factor is to one, the smoother the coefficient changes and the better it can suppress occasional noise. The smaller the forgetting factor, the faster the coefficient tracks the changes in the maneuver state. This linear decoupling follows the conventional idea of ​​adaptive interference cancellation, takes the slowly changing maneuver context as the reference quantity, and takes the vertical component of the bias as the main input. Therefore, its operation process is described in words and not listed separately. To avoid subtracting real road surface excitations related to maneuvering incidents, recursive least squares only uses the slowly changing maneuvering context as the regression independent variable. Short-wavelength road impacts are not included in the coefficient estimation. Therefore, decoupling only applies to slowly changing coupling and does not affect rapidly changing road surface excitations. At the same time, an upper limit is set on the single-step update amplitude of the coefficients. When the maneuvering context energy is lower than the threshold, the update is paused to preserve the vertical signal of the stationary segment. This processing method can handle superimposed excitation states. When a vehicle drives over a pothole during braking, the vertical channel simultaneously contains braking coupling and pothole impact. Only the slowly varying coupling component related to the maneuvering context is deducted, while the short-wavelength pothole impact is retained, thus allowing for the extraction of clean road surface excitation even under superposition conditions. The output of this step is the pure road surface excitation spatial residual and the maneuvering context features characterizing the driving action. It also transmits the passing speed, mileage, and optional positioning information corresponding to each spatial sample. The residual is the main input, the maneuvering context features are auxiliary discrimination information, the vehicle speed is used for speed conditional compensation in the recognition process, and the mileage and positioning are further transmitted.

[0023] S5 feeds the residuals of the pure road excitation space and the features of the motion context into a lightweight temporal network, so as to perform speed conditional modulation by vehicle speed and train through cross-speed consistency constraints, and output the road event type, bump level, confidence level, and the proportion of residual frequency band energy and dominant wavelength features of each event. Specifically, based on the pure road excitation space residual and the motion context features output from the previous step, and taking into account the vehicle speed, mileage, and optional positioning information that are subsequently passed through, a lightweight temporal network that fuses deep separable one-dimensional convolution and additive temporal attention is used to complete the recognition and classification. The main input of the network is the multi-channel sequence of the pure road excitation space residual, and the auxiliary input is the motion context features. The vehicle speed is used for speed conditional modulation in the later text, and the mileage and positioning are passed through. Before the pure road excitation space residual enters the network, it is first robustly normalized in each sliding window. That is, it is centered by the median in the window and normalized by a robust dispersion scale such as the absolute deviation of the median. This absorbs the overall amplitude difference caused by the suspension stiffness of different vehicles, so that the network focuses on the waveform shape rather than the absolute amplitude. The motion context features are also normalized accordingly before being incorporated. The data first enters several layers of depthwise separable one-dimensional convolutions for multi-scale morphology extraction. This depthwise separable one-dimensional convolution breaks down the standard convolution into two steps: channel-wise convolution and pointwise convolution. Channel-wise convolution extracts local waveforms along the spatial dimension within each channel using a one-dimensional convolution kernel, with the output channel number being the same as the input. Pointwise convolution then uses a one-dimensional convolution kernel to linearly combine channels to adjust the channel number. This splitting reduces the number of parameters and computational load to a level acceptable to embedded terminals while maintaining the receptive field. Each convolutional layer is followed by batch normalization and non-linear activation, with downsampling layer by layer along the spatial dimension, and the channel number increasing layer by layer. This forms a multi-scale feature map ranging from fine-grained impact patterns to coarse-grained undulation patterns. The network depth and number of channels are set according to the terminal computing power. For example, in this embodiment, the default number of depth-separable convolutional layers is four, and the number of channels increases by doubling from shallow to deep. The default length of the one-dimensional convolutional kernel covers the sample span of the smallest disease in the spatial domain, so that the shallow receptive field is aligned with local impacts and the deep receptive field is aligned with continuous undulations. In order to cope with the situation that events such as potholes and manhole cover seams account for a small proportion in the samples, the cross-entropy of event types is weighted according to the inverse of the category frequency during training, so that events with a small proportion can obtain sufficient gradient contributions. Considering that the waveform still retains deformation related to vehicle speed after spatial domain resampling, the passing vehicle speed corresponding to each spatial sample is introduced into the network as a known conditional quantity to perform speed conditional modulation on the convolutional features. The data flow is as follows: the passing vehicle speeds of each spatial sample in the current sliding window are taken to form a vehicle speed sequence, which is mapped by a small multilayer perceptron to a set of channel-wise scaling factors and a set of channel-wise offset factors with the same number of channels as the modulated convolutional layer. Then, the convolutional features of this layer are multiplied by the scaling factors and then the offset factors are added to each channel. In this way, the network can make corresponding compensations for the features according to the current vehicle speed. The waveform smoothing caused by undersampling in the high-speed section is corrected, and the modulation is biased to restore its morphological details. The modulation effect is weaker in the low-speed section. The reason for using vehicle speed as a conditional quantity rather than simply embedding it in the signal is that the details of high-speed passage are lost during the sampling stage. Only by providing the known quantity of vehicle speed as an explicit input to the network can the network learn the compensation method that changes with vehicle speed. This speed-conditional subnetwork has very few parameters and does not change the lightweight characteristics of the backbone. Additive temporal attention is introduced on top of multi-scale features to focus on sparse and local road events in long sequences. The data flow of additive attention is as follows: First, the feature vector of each spatial location is concatenated with the maneuver context features aligned to that location to form a joint vector. Before concatenation, the maneuver context features are aligned with the convolutional feature map according to the spatial dimension and channel mapping is performed to unify the dimension. Then, the joint vector is processed by a linear mapping and hyperbolic tangent activation to obtain an intermediate representation. The intermediate representation is then multiplied by a learnable scoring vector to obtain the unnormalized score for that location. Subsequently, the scores of each location are processed along the spatial dimension by a normalized exponent to obtain the attention weights. Finally, the convolutional features at each location are weighted and summed. The convergence features are obtained; this scoring method takes into account both the road surface excitation pattern and the current driving action, making the weight allocation of vertical anomalies in highly maneuverable road sections more prudent; the convergence features are output in two paths, both of which share the aforementioned convolutional backbone and attention convergence, and are only divided into two fully connected heads at the end. One path gives the event type through a fully connected classification head, which includes event types such as potholes, washboards, speed bumps, manhole cover seams, and continuous undulations. The other path gives the bump level through a fully connected rating head. Sharing the backbone allows the two tasks to provide regularization for each other and save parameters; the aforementioned depthwise separable convolution, batch normalization, velocity conditional modulation, and additive attention are all described in words for their layer-by-layer data flow and are reproducible network components; In one embodiment of the present invention, the network training samples are obtained from the data collected by the inspection vehicle on the marked road sections. The markings are provided by the accompanying high-precision road cross-section measurement. The event type and bump level are determined by the actual defect category of the marked road section and the roughness grade measured by the cross-section. In order to cover the variable speed condition and suppress the model's dependence on the collected vehicle speed, sample pairs of different vehicle speeds on the same road section are constructed. The sample pairs have two sources. One is that the same road section is actually collected by the same vehicle multiple times or multiple vehicles at different speeds and aligned by mileage. The other is that the time domain signal of the marked road section after preprocessing is taken, and time domain resampling is performed using different vehicle speed curves to simulate the passage of different vehicle speeds. Then, the spatial residual is obtained by completely going through attitude decoupling, spatial domain resampling and motor decoupling. It should be noted that the second source must start from the time domain signal and completely execute the processing flow again, rather than directly transforming the obtained spatial residual. This is because only by starting from the time domain can the undersampling and interpolation distortion introduced by different vehicle speeds be truly reproduced, so as to obtain sample pairs with differences. The training objective consists of three parts: event type using cross-entropy, bump level using ordered regression loss to ensure that the cost of misclassifying adjacent levels is lower than that of misclassifying across levels, thus conforming to the ordered attribute of the levels; and cross-speed consistency loss to constrain the intermediate features of network samples from different vehicle speeds on the same road segment to converge. The cross-speed consistency loss is a constraint unique to this scenario, and its formula is as follows: in, For cross-velocity consistency loss, This is a set of sample pairs consisting of different vehicle speeds on the same road segment. For the number of sample pairs, For the index of one of the samples, This represents the intermediate feature mapping after network attention convergence. and These are the residual samples of the pure road surface excitation space obtained at different vehicle speeds on the same road segment. It is the vector norm 2; It should be noted that this loss borrows from the idea in representation learning to bring different views of the same object closer together in the feature space. However, here the view difference is specifically limited to vehicle speed difference, which is the speed-related artifact remaining after spatial domain resampling as mentioned earlier. Therefore, this loss forces the network to learn the intrinsic road surface representation that is independent of vehicle speed, rather than collecting vehicle-related pseudo-correlation. The three parts of the loss are weighted and summed according to the trade-off coefficients to form the joint training objective. That is, the total loss equals the event type cross-entropy loss, plus the ordered regression loss multiplied by its trade-off coefficient, plus the cross-speed consistency loss multiplied by its trade-off coefficient. When the trade-off coefficient of the ordered regression loss increases, the model emphasizes the correctness of the ranking. When the trade-off coefficient of the cross-speed consistency loss increases, the model emphasizes the independence of vehicle speed. The values ​​of the two trade-off coefficients are first searched and determined on the validation set based on the balance between cross-speed recognition consistency and ranking correctness. This weighted summation is a common notation, so it is described in words. Speed-conditional modulation compensates for speed-related deformations within the network, while cross-speed consistency loss constrains residual differences on the training objective; these two components work in tandem. The three losses are uniformly backpropagated to update network weights. After training, the network weights are quantized and channels are pruned to compress model size and inference latency while maintaining recognition consistency, thus meeting the requirements for embedded deployment. The ordered regression loss is implemented by transforming bump level judgments into several cumulative binary judgments, i.e., determining whether each judgment level is higher than each boundary. The sum of the cross-entropy of each binary judgment is the ordered regression loss. Thus, misjudgments between adjacent levels only affect the judgment of one boundary, which is less costly than misjudgments across multiple boundaries. Sample pairs are organized in pairs within each training batch; samples of different vehicle speeds on the same road segment are included in the same batch to calculate the consistency loss. Training stops when cross-speed consistency on the validation set no longer improves. During online inference, the network operates in a sliding window manner on the continuous bitstream of the spatial residual of pure road excitation. The window slides along the mileage axis at a fixed length and partially overlaps. Each window provides an event type and bump level determination. Adjacent overlapping windows vote to merge the determinations for the same spatial location, thereby determining the start and end spatial locations of events on the continuous stream without omitting cross-window events. The window length and sliding step size are set according to the spatial scale of the event to be identified. The window length covers the spatial span of long events such as continuous undulations, and the step size ensures that adjacent windows fully overlap. The confidence level is jointly given by the two outputs. It is obtained by taking the maximum class probability after the classification head is normalized and the concentration of the predicted distribution of the ranking head. The concentration of the ranking distribution is given by the consistency of the adjacent boundary judgment or the dispersion measure of the predicted distribution. The more concentrated the class probability and the narrower the ranking distribution, the higher the confidence level. In the preferred embodiment, the uncertainty measure can also be given by the prediction variance output by the ranking head in an evidentiary manner. In this step, while outputting the road event type, bump level and its confidence level, the frequency band energy ratio and dominant wavelength of the pure road excitation spatial residual of each identified event segment are statistically analyzed on-site as the incidental features of the event.

[0024] S6, based on road event type, bump level, confidence level, residual frequency band energy ratio and dominant wavelength characteristics, combined with driving mileage, maps to roughness values ​​with the same dimensions as the International Roughness Index, and outputs time-series road condition datasets and road condition recognition records in the order of driving. Specifically, based on the event type, bump level, confidence level, and the accompanying residual frequency band energy ratio and dominant wavelength characteristics of the event segment output from the previous step, and combined with mileage and location information, standardized road condition data is generated. To enable cross-vehicle comparison of results collected from different vehicles at different speeds, the bump level and the aforementioned residual frequency band energy ratio and dominant wavelength characteristics are mapped together to roughness values ​​with the same dimensions as the International Roughness Index. The mapping relationship is obtained by fitting the network output on the marked road segment with the International Roughness Index measured by high-precision cross-sections, so that the output roughness values ​​have a unified physical scale. The event type is also retained as a semantic label for the roughness value, used to distinguish different causes of defects such as potholes and washboards under the same roughness. The fitting here is a conventional regression. The process is described in words. The input features used for fitting are the bump level, residual frequency band energy ratio, and dominant wavelength. The output is a roughness value with the same dimensions as the International Roughness Index. The fitting is completed on the marked road segment with the goal of minimizing the roughness prediction error. The duration of each event is obtained by multiplying its start and end span on the spatial index by the spatial sampling interval. The location of the event is calculated by combining the mileage corresponding to the start and end spatial locations with the positioning information. The positioning information comes from satellite positioning or existing mileage markers. When there is no satellite positioning, the location is estimated by combining the mileage with the starting point of the road segment. Records of multiple vehicles passing through the same road segment are first coarsely aligned by positioning, and then finely aligned by cross-correlation of spatial waveforms, thereby superimposing multiple records describing the same defect. For identification results with low confidence or in the low-speed confidence weighting segment, a quality label is marked in the output. This label is used for cloud aggregation to fuse results from multiple vehicles passing through the same road segment multiple times according to confidence weight, further improving data consistency. The quality label is given based on the confidence threshold and the low-speed weighting state. Results below the confidence threshold or in the low-speed weighting segment are marked as pending fusion. The location information binds each identification result to a specific mileage marker or geographic coordinate, enabling multiple collections of the same road segment to be aligned and superimposed. In one embodiment of the present invention, the recognition results, including event type, bump level, roughness value, location, duration of space, and confidence level, are organized into a time-series road condition dataset according to the driving sequence and archived in a structured manner. Each record is based on an event, and the fields include the start and end points of the event on the mileage axis, the event type label, the bump level, the roughness value, the confidence level, and the quality marker. The records are arranged in ascending order of mileage, thus preserving the spatial distribution of road conditions along the road and facilitating playback by trip. The dataset can be transmitted back to the cloud in real time for generating road defect work orders and scheduling maintenance priorities, or it can be distributed to the vehicle's semi-active suspension controller. Used for anticipating and adjusting road conditions ahead; during anticipation, the vehicle retrieves the event type and bump level within a certain distance ahead from the aggregated road condition dataset based on its current position, and adjusts the suspension damping accordingly. Since the dataset is built on a cross-vehicle consistent roughness scale, records previously collected by other vehicles can also be used by this vehicle. Because all results are based on attitude decoupling, spatial domain resampling, and motor decoupling, records from different vehicles at different speeds and mounting attitudes on the same road segment can be aligned and compared, meeting the need for crowdsourced collection to form reliable road big data; the output of this step is a time-series road condition dataset and road condition recognition records.

[0025] In one embodiment of the present invention, the method is deployed in a crowdsourced vehicle fleet undertaking a survey of urban main roads; one of the vehicles in the fleet is equipped with a six-axis microelectromechanical sensing terminal, the installation position and angle of which are unknown relative to the vehicle body, in order to cover the typical condition of uncertain posture in a crowdsourced scenario; the road section being collected includes smooth asphalt, continuous washboard, several isolated potholes, a speed bump, and a section of uphill bridge, and the vehicle experiences two red light braking and one lane change during the journey, with the vehicle speed fluctuating between congested crawling and normal traffic; the terminal completes the identification locally according to S1 to S6, and uploads the structured results to the cloud for aggregation; Table 1 shows the initial readings in the sensor body coordinate system during the initial acquisition phase. Table 1. Initial readings in the initial stage The three-axis acceleration, three-way angular velocity and the synchronous vehicle speed given by the vehicle speed interface are aligned with a unified time reference. The tire noise peak that shifts with the vehicle speed and the relatively fixed mechanical noise peak are located by using a sliding window power spectrum. After applying adaptive band stop, the low- and mid-frequency road excitation is retained to obtain multi-channel motion signals and sample-by-sample vehicle speed. Gravity direction is estimated recursively using quasi-static low-frequency components. The six-axis angular velocity and gravity direction are filtered by complementary means to obtain the complete attitude including heading. The attitude is then projected into three directions: vertical road surface, longitudinal vehicle and lateral vehicle acceleration. The ramp causes the original vertical reading to deviate from the gravity constant for a long time. After attitude reconstruction, the slope component is incorporated into the vertical gradual offset to avoid being falsely reported as continuous bumps, thereby eliminating gravity leakage caused by installation angle and slope. The mileage is accumulated using wheel speed pulses. The three-directional signals are mapped from the time index to the mileage axis and interpolated at fixed spatial intervals to obtain a spatial domain multi-channel waveform that does not drift with vehicle speed and the vehicle speed of each sample. When the same pothole is passed by creeping and normal vehicle speed respectively, the width and peak shape of the time domain waveform are significantly different, while the peak position and span tend to be consistent in the spatial domain, providing a unified spatial scale for subsequent cross-vehicle speed identification. The longitudinal and lateral channels are separated into gradually changing maneuver contexts by wavelength. After deducting the vertical gravity bias, the maneuver coupling is removed by recursive least squares. The longitudinal gradually changing component of the red light braking segment is deducted, while the short-wavelength impact of potholes passed during braking is retained. The centrifugal and tangential accelerations of the lane change and turning segment fall into the lateral maneuver context. The longitudinal axis is updated and frozen and the previous reliable estimate is used. The heading does not deviate due to the turning. Thus, the pure road excitation space residual can still be extracted when the driving operation and road excitation are superimposed. The residuals of the pure road surface excitation space and the features of the motor context are fed into a lightweight temporal network to perform speed conditional modulation by vehicle speed. Under the condition of cooperating with the cross-speed consistency constraint in the training phase, inference is completed. The event types, bump levels, confidence levels, and residual frequency band energy proportions and dominant wavelength characteristics of each event are output through sliding window voting, such as potholes, washboards, speed bumps, manhole cover seams and continuous undulations. Based on this mapping, a roughness value with the same dimensions as the International Roughness Index is generated. Combined with mileage and optional positioning to bind the start and end positions of events and the duration of the space, quality markers are added to low-confidence or low-speed weighted segments. A time-series road condition dataset and road condition identification record are formed according to the driving sequence. After the dataset is uploaded to the cloud, it is used for the generation of road defect work orders and the prioritization of maintenance. It can also be sent to the semi-active suspension controller to retrieve the type of events ahead and the bump level for pre-aiming adjustment. This embodiment shows that even under conditions of unknown installation posture, continuous vehicle speed fluctuation, and superposition of braking, steering and road surface excitation, it can still provide time-series road condition results that can be compared across vehicles.

[0026] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent sensing and recognition of road bumps based on multi-axis acceleration temporal characteristics, characterized in that, Includes the following steps: S1 collects vibration data streams from the accelerometer and synchronous vehicle speed, and obtains multi-channel motion signals and vehicle speed through time alignment and adaptive band-stop denoising. S2, based on multi-channel motion signals, uses quasi-static low-frequency components and horizontal acceleration principal components to estimate the gravity direction and heading and reconstruct the motion coordinate system, and projects to obtain the attitude decoupled acceleration in the three directions of road vertical, vehicle longitudinal and vehicle lateral. S3, the attitude decoupled acceleration is resampled in the spatial domain driven by vehicle speed, and converted into a spatial domain multi-channel waveform arranged at fixed intervals along the driving distance, and the driving distance and the passing speed of each spatial sample are obtained. S4. Separate the slowly varying maneuvering context from the longitudinal and transverse channels of the multi-channel waveform in the spatial domain according to the wavelength. After deducting the vertical gravity bias, remove the vertical coupling of the maneuvering context by recursive least squares to obtain the spatial residual of pure road excitation and the characteristics of the maneuvering context. S5 feeds the residuals of the pure road excitation space and the features of the motion context into a lightweight temporal network, so as to perform speed conditional modulation by vehicle speed and train through cross-speed consistency constraints, and output the road event type, bump level, confidence level, and the proportion of residual frequency band energy and dominant wavelength features of each event. S6, based on road event type, bump level, confidence level, residual frequency band energy ratio and dominant wavelength characteristics, combined with driving mileage, maps to roughness values ​​with the same dimensions as the International Roughness Index, and outputs time-series road condition datasets and road condition recognition records in the order of driving.

2. The intelligent sensing and recognition method for road bumps based on multi-axis acceleration temporal characteristics according to claim 1, characterized in that, The time-aligned and adaptive band-stop denoising process includes online estimation of the power spectrum distribution of multi-channel motion signals using a sliding window periodogram, identifying spectral peaks that shift with vehicle speed and have narrow bandwidth as tire noise-related components, and identifying spectral peaks with fixed positions as mechanical noise-related components. A bandstop is applied to the above spectral peaks while preserving the mid-to-low frequency band where the road surface excitation is located, and the depth and width of the bandstop are adaptively set according to the bandwidth of the spectral peaks.

3. The intelligent sensing and recognition method for road bumps based on multi-axis acceleration temporal characteristics according to claim 1, characterized in that, The method of estimating the gravity direction and heading and reconstructing the motion coordinate system using quasi-static low-frequency components and horizontal acceleration principal components includes extracting quasi-static low-frequency components and updating the unit vector estimate of the gravity direction in a recursive manner with a forgetting factor, and assigning smaller confidence weights to samples whose magnitude of the composite force deviates more from the gravitational constant. The vehicle longitudinal direction corresponding to the heading is estimated from the principal component directions of the two acceleration components in the horizontal plane. The update of the vehicle longitudinal direction is frozen in the uniform speed segment and the pure turning segment, and the previous estimate is used. Using the opposite direction of gravity as the vertical axis of the road surface and the longitudinal direction of the vehicle as the longitudinal axis, the vehicle's lateral direction is obtained by cross product of the two after orthogonal normalization, thus forming a motion coordinate system.

4. The intelligent sensing and recognition method for road bumps based on multi-axis acceleration temporal characteristics according to claim 1, characterized in that, The spatial domain resampling of attitude decoupled acceleration driven by vehicle speed includes using the cumulative mileage of wheel speed pulses as the driving mileage when the vehicle provides wheel speed pulses, and integrating the vehicle speed over time to obtain the driving mileage when there are no wheel speed pulses. Each attitude decoupled acceleration sample is re-indexed from the time index to the mileage index, and then resampled along the mileage axis at fixed spatial intervals to obtain a spatial domain multi-channel waveform arranged at fixed intervals along the travel distance. The spatial interval is determined according to the minimum defect spatial wavelength to be identified.

5. The intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics according to claim 4, characterized in that, The interpolation resampling is a variable step size adaptive interpolation, which takes points sparsely in time in high-speed sections and densely in low-speed sections to maintain a fixed spatial interval, and uses piecewise cubic interpolation along the mileage axis. For the drift of driving mileage, when there are equally spaced artificial structures, the spatial waveform is autocorrelated and the cumulative mileage is proportionally corrected by the peak spacing of its regular peaks. When there are no equally spaced artificial structures, the local speed ratio is normalized within the sliding window. The smaller the confidence weight is assigned to spatial samples with instantaneous vehicle speeds less than a preset speed threshold, the less new spatial samples are generated when the vehicle stops and the mileage no longer increases.

6. The intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics according to claim 1, characterized in that, The method of separating the slowly changing maneuvering context from the longitudinal and transverse channels of the multi-channel waveform in the spatial domain according to wavelength includes decomposing the longitudinal and transverse channels according to spatial wavelength using zero-phase bidirectional filtering, taking the continuous component with wavelength greater than the cutoff wavelength as the maneuvering context, and taking the component with wavelength less than the cutoff wavelength as the fast-changing road surface excitation. The cutoff wavelength is determined based on the largest spatial scale of the disease to be identified, and a transition zone is set near the cutoff wavelength. Components falling into the transition zone are weighted according to their distance from both sides.

7. The intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics according to claim 1, characterized in that, The process of removing the vertical gravity bias and then removing the vertical coupling of the maneuvering context using recursive least squares includes filtering out the quasi-static gravity bias of the ultra-long wavelength in the vertical channel with a zero-phase high-pass filter to obtain the debiased vertical component, and then estimating the decoupling coefficient online using recursive least squares with a forgetting factor to minimize the energy in the debiased vertical component that is linearly related to the maneuvering context and thus obtain the pure road excitation space residual. When estimating the decoupling coefficients, only the slowly varying maneuvering context is used as the regression quantity. Short-wavelength road impacts are not included in the estimation. An upper limit is set on the single-step update amplitude of the decoupling coefficients, and the update is paused when the maneuvering context energy is less than a preset threshold.

8. The intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics according to claim 1, characterized in that, The lightweight temporal network takes the multi-channel sequence of the pure road excitation space residual as the main input and the motor context features as the auxiliary input. Before the pure road excitation space residual enters the network, it is centered by the median in each sliding window and the amplitude is normalized by the absolute deviation of the median. The network extracts multi-scale features layer by layer using depthwise separable one-dimensional convolutions. The depthwise separable one-dimensional convolutions consist of channel-wise convolutions that extract local waveforms along the spatial dimension and pointwise convolutions that are linearly combined between channels. Additive temporal attention is used to calculate scores for each spatial location and normalize them into attention weights before weighted convergence. When calculating the scores, the dynamic context features and the features of each spatial location are concatenated and then used. The velocity conditional modulation maps the passing speed of each spatial sample to a channel-wise scaling factor and offset factor through a multilayer perceptron. The convolutional features are first multiplied by the scaling factor and then the offset factor is added.

9. The intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics according to claim 1, characterized in that, The training objective of the cross-speed consistency constraint training consists of the weighted sum of the cross-entropy of the event type, the ordered regression loss of the bump level, and the cross-speed consistency loss. The ordered regression loss transforms the determination of the bump level into the sum of the cumulative binary determinations of whether each level exceeds each boundary. The cross-speed consistency loss constrains the intermediate network features corresponding to different vehicle speeds on the same road segment to move closer to each other. The sample pairs that constitute the cross-speed consistency loss are obtained from the same road segment collected at different vehicle speeds and aligned by mileage, or from the same road segment time domain signal resampled at different vehicle speeds and then sequentially subjected to attitude decoupling, spatial domain resampling and removal of vertical coupling of the maneuver context.

10. The intelligent sensing and recognition method for road bumps based on multi-axis acceleration time-series characteristics according to claim 1, characterized in that, The mapping is a roughness value with the same dimensions as the International Roughness Index. It includes mapping the roughness level and the residual frequency band energy ratio and dominant wavelength characteristics of each event to a roughness value with the same dimensions as the International Roughness Index by fitting the mapping relationship on the marked road section, and using the road event type as the semantic label of the roughness value. The duration of each event is obtained by multiplying its start and end span on the spatial index by the spatial interval. Records of multiple vehicles passing through the same road segment are first aligned by location and then finely aligned by cross-correlation of spatial waveforms. The time-series traffic data set is arranged by event and in ascending order of mileage. Recognition results with confidence levels less than a preset threshold are marked with quality labels.

Images