Road surface mechanical response signal adaptive de-noising filtering method based on signal-to-noise ratio

Abstract

The present application relates to the technical field of signal denoising, and particularly relates to a road surface mechanics response signal adaptive denoising filtering method based on signal-to-noise ratio, comprising collecting original mechanics response time domain signals of a road surface structure under the action of vehicle load; performing segmented processing on the original mechanics response time domain signals to obtain a plurality of continuous signal segments; for each signal segment, the following processing procedures are sequentially executed: extracting the local signal-to-noise ratio and signal inherent characteristics of the signal segment; the signal inherent characteristics are used to represent the signal time-frequency structure and distribution characteristics; the local signal-to-noise ratio characteristics and the signal inherent characteristics are taken as inputs and input into a preset filtering parameter decision model to output a target filtering method suitable for the signal segment and the corresponding key filtering parameters; the target filtering method and the corresponding at least one key filtering parameter are used to perform filtering processing on the signal segment to obtain a denoised signal segment. According to the present application, the filtering strategy and parameters can be dynamically adjusted according to the local characteristics of the signal.

Description

Technical Field

[0001] This invention relates to the technical field of signal noise reduction, and in particular to an adaptive noise reduction filtering method for road mechanical response signals based on signal-to-noise ratio. Background Technology

[0002] In long-term performance monitoring of road engineering, sensors embedded within the pavement structure are often used to collect mechanical response signals under vehicle loads, such as strain, acceleration, or deflection, to assess the pavement structure condition and service performance. However, due to factors such as inherent sensor noise, electromagnetic interference, environmental vibration, and signal transmission, the acquired raw time-domain signals are usually mixed with significant noise, resulting in a low signal-to-noise ratio. Especially under light load or low-speed conditions, the effective signal amplitude is weak and more easily masked by noise.

[0003] Existing noise reduction filtering methods generally employ fixed-parameter filtering techniques, such as moving average, low-pass filtering, or wavelet transform. These methods typically apply the same filtering intensity to the entire signal segment based on a pre-defined filtering approach and uniform parameters. However, road surface mechanical response signals exhibit significant non-stationarity and local time-varying characteristics. The time-frequency structure and distribution characteristics of different signal segments vary. Relying solely on fixed-parameter filtering strategies results in insufficient adaptability to signal segments with different time-frequency characteristics, making it difficult to balance global applicability with local fidelity, thus affecting the overall noise reduction effect.

[0004] Therefore, there is an urgent need for an adaptive noise reduction method that can dynamically adjust the filtering strategy and parameters based on the local characteristics of the signal. Summary of the Invention

[0005] This invention provides an adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio, which can effectively solve the problems in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio includes:

[0008] Collect the original mechanical response time-domain signal of the pavement structure under vehicle load;

[0009] The original mechanical response time-domain signal is segmented to obtain multiple continuous signal segments;

[0010] For each of the aforementioned signal segments, the following processing flow is executed sequentially:

[0011] The local signal-to-noise ratio and inherent characteristics of the signal segment are extracted; these inherent characteristics are used to characterize the time-frequency structure and distribution properties of the signal.

[0012] The local signal-to-noise ratio features and the inherent features of the signal are taken as inputs and fed into a preset filter parameter decision model, which outputs a target filtering method applicable to the signal segment and its corresponding key filtering parameters.

[0013] The target filtering method and at least one corresponding key filtering parameter are used to filter the signal segment to obtain a noise-reduced signal segment.

[0014] All the denoised signal segments are reassembled in sequence to generate the final adaptive denoised signal.

[0015] Furthermore, the inherent characteristics of the signal include at least one of local frequency characteristics and local statistical characteristics.

[0016] Furthermore, the extraction of the local frequency features includes: performing time-frequency analysis on the signal segment to calculate the dispersion of its energy-dominant frequency band and / or frequency component distribution;

[0017] The extraction of the local statistical features includes: calculating at least one of the amplitude standard deviation, kurtosis, or waveform factor of the signal segment.

[0018] Further, the local signal-to-noise ratio of the signal segment is extracted, including:

[0019] Identify and separate the background noise signal segment representing the absence of vehicle load from the original mechanical response time-domain signal;

[0020] Based on the background noise signal segment, the amplitude statistics of the background noise are calculated as the background noise benchmark;

[0021] The local signal-to-noise ratio is calculated based on the background noise benchmark and the amplitude statistics of the current signal segment.

[0022] Furthermore, the background noise signal segment characterizing the absence of vehicle load is identified and separated from the original mechanical response time-domain signal, including:

[0023] Identify the start and end points of the vehicle load action in the original mechanical response time-domain signal;

[0024] The signal segments within the first preset time period before the starting point and the signal segments within the second preset time period after the ending point are determined as the background noise signal segments.

[0025] Furthermore, the identification of the starting point and the ending point is based on at least one of the following criteria:

[0026] The signal amplitude exceeds the preset amplitude threshold;

[0027] The zero-crossing points of the first derivative of a signal are combined with the extreme points of the second derivative of a signal.

[0028] Furthermore, the filter parameter decision model is a machine learning model or a deep learning model trained based on historical data;

[0029] The training data for the filter parameter decision model includes a sample set consisting of signal segments with different signal-to-noise ratios, local frequency characteristics, and local statistical characteristics, as well as the optimal filtering method and filter parameter labels for each sample set, determined by experts or optimization algorithms.

[0030] Furthermore, the target filtering method is any one of the moving average method, low-pass filtering method, and wavelet transform method.

[0031] Furthermore, when the target filtering method is the moving average method, the corresponding key filtering parameter is the adaptive step size;

[0032] When the target filtering method is a low-pass filter, the corresponding key filtering parameter is the adaptive cutoff frequency.

[0033] When the target filtering method is wavelet transform, the corresponding key filtering parameters are adaptive threshold and / or number of decomposition layers.

[0034] Furthermore, all the denoised signal segments are reassembled sequentially to generate the final adaptive denoised signal, including:

[0035] Initialize an output signal sequence with the same length as the original mechanical response time-domain signal;

[0036] Based on the time index of each signal segment in the original mechanical response time domain signal, the data of each denoised signal segment is sequentially filled into the corresponding position of the output signal sequence;

[0037] Generates a temporally continuous and complete adaptive noise reduction signal.

[0038] The technical solution of this invention achieves the following technical effects: by using the local signal-to-noise ratio and inherent signal characteristics as inputs, a preset filtering parameter decision model is driven to dynamically output the target filtering method and its key parameters applicable to each signal segment, thereby realizing refined and adaptive noise reduction processing of the entire original mechanical response signal. Specifically, local analysis units are established through segmented processing, and the local signal-to-noise ratio and inherent characteristics characterizing the time-frequency structure and distribution of the signal are extracted synchronously within each signal segment. The local signal-to-noise ratio reflects the relative intensity of the effective components and noise in the current signal segment, while the inherent signal characteristics depict its inherent dynamic characteristics, such as frequency concentration, impulsivity, or waveform complexity. Together, they constitute a multidimensional description of the local state of the signal and serve as the input basis for the filtering parameter decision model. The filtering parameter decision model is trained based on historical data or prior knowledge and can intelligently match the optimal filtering method and its corresponding key parameters according to the input feature combination.

[0039] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart illustrating the adaptive noise reduction and filtering method for road mechanical response signals based on signal-to-noise ratio according to the present invention.

[0042] Figure 2 The original signal data curve;

[0043] Figure 3 The signal data curve after processing using the moving average method;

[0044] Figure 4 A schematic diagram of the dynamic response curve under tension-compression transformation;

[0045] Figure 5 This is a schematic diagram of the dynamic response curve with tension as the primary characteristic.

[0046] Figure 6 This is a schematic diagram of the dynamic response curve dominated by pressure. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0048] 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 invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0049] like Figure 1 As shown, the adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio of the present invention specifically includes the following steps:

[0050] Step S100: Collect the original mechanical response time-domain signal of the road surface structure under vehicle load;

[0051] Step S200: The original mechanical response time-domain signal is segmented to obtain multiple continuous signal segments;

[0052] Step S300: For each of the signal segments, the following processing flow is executed sequentially:

[0053] Step S310: Extract the local signal-to-noise ratio and inherent signal features of the signal segment; the inherent signal features are used to characterize the time-frequency structure and distribution characteristics of the signal;

[0054] Step S320: Input the local signal-to-noise ratio features and the inherent features of the signal into a preset filter parameter decision model, and output the target filtering method applicable to the signal segment and its corresponding key filtering parameters.

[0055] Step S330: Using the target filtering method and at least one corresponding key filtering parameter, the signal segment is filtered to obtain a noise-reduced signal segment;

[0056] Step S400: Reassemble all the noise-reduced signal segments in sequence to generate the final adaptive noise-reduced signal.

[0057] In this embodiment, by using both the local signal-to-noise ratio (SNR) and inherent signal characteristics as input, a preset filter parameter decision model is driven to dynamically output the target filtering method and its key parameters applicable to each signal segment, thereby achieving refined and adaptive noise reduction processing of the entire original mechanical response signal. Specifically, local analysis units are established through segmented processing, and the local SNR and inherent characteristics characterizing the signal's time-frequency structure and distribution are extracted synchronously within each signal segment. The local SNR reflects the relative intensity of the effective components and noise in the current signal segment, while the inherent signal characteristics depict its inherent dynamic properties, such as frequency concentration, impulsivity, or waveform complexity. Together, they constitute a multidimensional description of the signal's local state and serve as the input basis for the filter parameter decision model. The filter parameter decision model is trained based on historical data or prior knowledge and can intelligently match the optimal filtering method and its corresponding key parameters according to the input feature combination.

[0058] Because the selection of filtering strategies and parameters considers both noise levels and the structural characteristics of the signal itself, misjudgments caused by simply judging based on the signal-to-noise ratio (SNR) are avoided. For example, in signal segments with low SNR but typical impact characteristics, the model can choose a filtering method that preserves transient characteristics; in segments with high SNR and stable waveforms, a lightweight filter with higher computational efficiency can be used, so that the noise reduction process can effectively suppress noise while preserving the key information reflecting the road structure response to the greatest extent. By sequentially recombining each noise-reduced signal segment, a continuous and complete adaptive noise-reduced signal is finally generated, improving noise reduction accuracy and enhancing the method's adaptability to mechanical response signals under different working conditions and road conditions.

[0059] In summary, this invention achieves adaptive and high-fidelity noise reduction of road surface mechanical response signals by combining the synergistic effect of local signal-to-noise ratio and inherent signal characteristics with a preset filter parameter decision model. This effectively solves the technical problem of insufficient adaptability of existing fixed-parameter filtering methods when processing non-stationary and locally time-varying signals, and improves the noise reduction effect.

[0060] In a specific implementation, as one example, step S100 is carried out by acquiring the original mechanical response time-domain signal of the road surface structure under vehicle load in the following manner:

[0061] Based on the engineering monitoring requirements, monitoring locations were selected at key sections of the wheel track zone of the road surface structure, and mechanical response sensors were installed at these locations. The sensors were selected from one of the following: resistance strain gauges, piezoelectric accelerometers, or fiber optic strain gauges, with their sensing axes aligned with the direction of the force being measured. Waterproofing, insulation, and mechanical protection were completed during the sensor installation process.

[0062] Connect the sensor's output signal to the corresponding input channel of the multi-channel dynamic signal acquisition instrument via a cable. Set the instrument's operating parameters according to signal characteristics and engineering standards. Set the sampling frequency to no less than 1000Hz, which is determined based on the Nyquist sampling theorem and the highest frequency component of the road surface mechanical response. Set the acquisition range according to the sensor's range and the estimated signal amplitude. Set the input coupling mode to DC coupling.

[0063] When a vehicle travels over the area above the sensor, the vehicle load excites the road surface structure to produce a dynamic response. This response is sensed by the sensor and converted into an analog electrical signal. The analog electrical signal is then subjected to anti-aliasing filtering and analog-to-digital conversion by the data acquisition instrument to form a discrete digital signal sequence indexed by time. This digital signal sequence is transmitted in real time and stored completely. The parameter settings of the sampling frequency and range ensure the complete acquisition of the signal frequency components and amplitude range. The installation of the sensor on the wheel track ensures the effective characterization of the signal on the vehicle load.

[0064] In some embodiments of the present invention, step S200 employs a non-uniform segmentation strategy based on load event detection to segment the original mechanical response time-domain signal to obtain multiple continuous signal segments; the specific implementation is as follows:

[0065] Step S210: Perform load event identification on the original mechanical response time-domain signal obtained in step S100; the identification process is based on the dual threshold comparison method, setting a first threshold and a second threshold, with the first threshold being greater than the second threshold; when the signal amplitude rises from a state below the second threshold and exceeds the first threshold, and remains there for a preset minimum duration, this point is recorded as the start point of a load event; when the signal amplitude drops from a high amplitude state to below the second threshold, and remains there for a preset stabilization time, this point is recorded as the end point of the load event; the values ​​of the first threshold, the second threshold, the minimum duration, and the stabilization time are determined based on statistical analysis of typical load responses and background noise in historical monitoring data;

[0066] Step S220: After identifying the start and end points of all load events, perform an adaptive segmentation operation; take the start and end points of each load event as nodes, extend forward and backward by a preset protection time, and define the time interval thus defined as the event-dominant signal segment; the setting of the protection time refers to the typical time scale of the mechanical response transition process of the pavement structure before and after the vehicle load, in order to ensure that a single complete load excitation and structural relaxation process is contained within the same signal segment;

[0067] Step S230: For temporally adjacent dominant event signal segments, if their extended time intervals overlap, they are merged into a continuous composite event segment; for time intervals in the original signal that are not included in any dominant event signal segment or composite event segment, they are classified as pure noise background periods; for these pure noise background periods, a sliding window of fixed time length is used to uniformly divide them, generating background noise signal segments of equal length; the fixed time length is set according to the minimum data length requirement required by the system for background noise stationarity statistics;

[0068] Step S240: Arrange and splice all generated event-dominant signal segments, composite event segments, and background noise signal segments according to their chronological order in the original signal to form a sequence of analysis units that continuously cover the entire original signal on the time axis and have relatively consistent internal signal characteristics, thus completing the segmentation process.

[0069] In this embodiment, through the above settings, each event-dominant signal segment mainly contains a physically complete load response process, and the non-stationary changes of the signal are mainly reflected in the continuous evolution of amplitude and frequency rather than step changes; each background noise signal segment mainly contains statistically stable noise; based on the local signal-to-noise ratio and inherent signal characteristics calculated by the analysis unit, the true characteristics of the signal under the corresponding physical state can be more accurately reflected, thereby providing a more reliable feature input for filter parameter decision-making.

[0070] In some embodiments of the present invention, step S310 is specifically implemented as follows to extract the quantization features necessary for subsequent filtering decisions from each signal segment:

[0071] Step S311, the extraction of local signal-to-noise ratio, is as follows:

[0072] The background noise signal segment representing the absence of vehicle load is identified and separated from the original mechanical response time-domain signal; the identification of the start and end points of vehicle load action is based on the time-domain or differential characteristics of the signal as a criterion; in one embodiment, an amplitude threshold criterion is adopted: when the signal amplitude rises from a baseline state below a low threshold and exceeds a preset high threshold, it is determined as the start point of load action; when the signal amplitude falls back from a high state to below the low threshold and remains stable for a preset period of time, it is determined as the end point of load action; the high threshold and low threshold are determined based on the statistical distribution of signal peak value and background noise level in long-term monitoring data; in another embodiment, identification is performed by combining differential signals, using the zero-crossing point of the first derivative of the signal to identify the turning point of the signal change trend, and combining the extreme point of the second derivative of the signal to identify the acceleration extreme point of the signal change as a reference for the boundary of the load event;

[0073] After identifying the start and end points, background noise signal segments are extracted. Continuous signal segments within a first preset time period before the start point and continuous signal segments within a second preset time period after the end point are extracted and merged. The first preset time period can be set to 0.5 to 1 second before the start point, and the second preset time period can be set to 0.5 to 1 second after the end point. These two time periods are within a defined vehicle load interval, and their signals are considered to mainly consist of sensor inherent noise, environmental background vibration, and other noise sources. The merged signal segment is determined as the background noise signal segment used to calculate the noise benchmark.

[0074] The background noise statistics are calculated based on the aforementioned background noise signal segment. Since the original mechanical response data characterizes numerical changes as voltage variations, the effective amplitude of the signal and the effective amplitude of the noise are used in the calculation. The root mean square of the signal voltage value in this segment is calculated as the effective value A of the background noise voltage. noise,RMS It characterizes the amplitude level of background noise; it calculates the root mean square of the voltage value of the current signal segment, which is taken as the effective voltage value A of the mixed signal and noise in that segment. mix,RMS Then the local signal-to-noise ratio (SNR) of the current signal segment is... local The result is calculated using the following formula, expressed in decibels (dB):

[0075] SNR local (dB) = 20 × log 10 (A mix,RMS / A noise,RMS );

[0076] The above calculation formula directly reflects the intensity ratio of the useful signal component in the current signal segment relative to the background noise reference.

[0077] More specifically, in another implementation, the identification of the start and end points of the vehicle load is achieved by analyzing the differential characteristics of the original mechanical response time-domain signal; the core of this method lies in using the mathematical characteristics of the signal change rate (first derivative) and the change acceleration (second derivative) to detect the moment of abrupt change in the mechanical state; the specific implementation steps are as follows:

[0078] The original mechanical response time-domain signal is smoothed and filtered to obtain a preprocessed signal; the first-order numerical derivative of the preprocessed signal is calculated, and the sequence of the first-order numerical derivatives characterizes the instantaneous rate of change of the signal at each sampling point;

[0079] The identification of the starting point is accomplished by locating the zero-crossing points in the first derivative sequence that meet specific conditions; the search for zero-crossing points that satisfy the following logic is performed: before the point, the value of the first derivative is less than or equal to a negative tolerance threshold close to zero; after the point, the value of the first derivative is greater than a set positive detection threshold; the zero-crossing points that satisfy this condition are marked as starting candidate points.

[0080] The identification of the end point is accomplished by locating another type of zero-crossing point in the first derivative sequence; the zero-crossing point is searched for that satisfies the following logic: before the point, the value of the first derivative is greater than or equal to a positive tolerance threshold close to zero; after the point, the value of the first derivative is less than a set negative detection threshold; the zero-crossing point that satisfies this condition is marked as a candidate end point.

[0081] Furthermore, the second-order numerical derivative of the preprocessed signal is calculated. The sequence of second-order numerical derivatives characterizes the trend of the instantaneous rate of change of the signal, i.e., acceleration. The local extrema of the second-order derivative correspond to the moment when the acceleration of the signal change is the greatest. The candidate points identified by the zero-crossing points of the first-order derivative are compared and correlated with the local extrema of the second-order derivative in the adjacent time period. If a starting candidate point or an ending candidate point is sufficiently close to a local extrema of the second-order derivative in time, the position of the candidate point can be corrected or confirmed by the position of the extrema point to determine the final starting point or ending point position.

[0082] The specific values ​​of the positive detection threshold, negative detection threshold, and tolerance threshold are determined by analyzing and statistically analyzing the signal derivatives corresponding to known load events in historical data. The above method is based on the physical fact that the rate of change of the mechanical response signal must change direction when the vehicle load begins to act on the road structure and leaves. The event boundary is identified by detecting the change in its mathematical derivative characteristics.

[0083] Step S312: The inherent features of the signal are used to quantitatively characterize the time-frequency structure and distribution characteristics within the signal segment. The extraction of these features includes local frequency features and local statistical features.

[0084] Specifically, local frequency features are extracted by performing time-frequency analysis on the signal segment, as follows:

[0085] Perform spectral analysis on the discrete-time series of the current signal segment, for example, by converting it to the frequency domain using a Fast Fourier Transform to obtain the spectrum of the signal segment; identify the frequency range with the most concentrated energy in the spectrum; specifically, locate the frequency component with the largest amplitude or energy in the spectrum, and determine the frequency corresponding to this component or the narrow frequency range centered on it as the energy-dominant frequency band of the signal segment; to quantify the dispersion of the signal frequency components, analyze the spectrum to calculate a frequency component distribution dispersion index; the frequency component distribution dispersion index is obtained by evaluating the breadth of effective energy in the frequency domain in the spectrum; one implementation is to calculate the standard deviation of the spectral amplitude sequence, the larger the standard deviation value, the more dispersed the frequency component distribution, and the smaller the value, the more concentrated the energy is near the dominant frequency band.

[0086] Local statistical features are extracted by performing statistical calculations on the time-domain amplitude sequence of the current signal segment, as follows:

[0087] The standard deviation of the amplitude sequence is calculated to quantify the degree of fluctuation of the signal amplitude around its average level; the kurtosis of the amplitude sequence is calculated to describe the tail thickness of the signal amplitude probability distribution curve, and its magnitude reflects the probability of the presence of large amplitude impact components in the signal; the waveform factor of the amplitude sequence is calculated, which is defined as the ratio of the root mean square value of the signal to its absolute average value, and is used to characterize the shape difference characteristics between the waveform of the signal and a pure sine wave.

[0088] By performing the above-mentioned extraction operations of local frequency features and local statistical features, one or more specific quantitative indicators are obtained to describe the time-frequency structure and distribution characteristics of the signal segment; the above-mentioned quantitative indicators and the aforementioned local signal-to-noise ratio together constitute a feature set characterizing the state of the signal segment.

[0089] In some embodiments of the present invention, the extraction of inherent signal features in step S312 further includes the analysis and extraction of local response morphological features of the signal segment, achieved by processing the time-domain waveform data and its numerical derivative sequence of the current signal segment; based on the positive and negative change trend of the data sequence value within the signal segment relative to a preset baseline value, the local response morphology of the signal segment is classified into one of three categories: tension-compression transformation type, tension-dominant type, or compression-dominant type; based on the first-order and second-order numerical derivative sequences of the signal segment, key feature points corresponding to the above-mentioned morphological categories are identified; the specific identification method is as follows:

[0090] like Figure 4 As shown, for signal segments classified as pull-compression transformation types, the following three characteristic points are identified:

[0091] Feature point A: The point where the first derivative of the strain curve is less than zero and the second derivative is close to zero during the wheel approach process, corresponding to the maximum compressive strain.

[0092] Feature point B: The point where the maximum tensile strain is identified within one wheel load cycle;

[0093] Feature point C: The point where the first derivative of the strain curve is less than zero and the second derivative is close to zero during the wheel's departure process, corresponding to the maximum compressive strain.

[0094] like Figure 5 As shown, for signal segments classified as pull-dominant, three characteristic points are identified:

[0095] Feature point D: The point where the first derivative of the strain curve is less than zero and the second derivative is close to zero during the wheel approach process;

[0096] Feature point E: The point where the maximum tensile strain is identified within this signal segment;

[0097] Feature point F: The point where the first derivative of the strain curve is less than zero and the second derivative is close to zero during the wheel's departure process.

[0098] like Figure 6 As shown, for signal segments classified as pressure-dominant, three characteristic points are identified:

[0099] Feature point G: The point where the first derivative of the strain curve is greater than zero and the second derivative is close to zero during the wheel approach process;

[0100] Feature point H: The point with the maximum compressive strain identified within this signal segment;

[0101] Feature point I: The point where the first derivative of the strain curve is greater than zero and the second derivative is close to zero during the wheel's departure process.

[0102] The extracted output of local response morphological features includes: signal segment morphological classification identifiers, and the relative position indices of the identified key feature points within the time series of the signal segment, such as the index numbers of feature points A, E, and H. Local response morphological features, together with local frequency features and local statistical features, constitute a more comprehensive description of the inherent characteristics of the signal segment.

[0103] In this embodiment, by introducing fine classification of response patterns and feature point recognition based on derivative analysis, the features extracted in step S310 can more deeply reflect the physical essence of the signal segment, such as the symmetry of the load action and the dominant response mode. They can be used as additional inputs to the filter parameter decision model, which helps the model to consider not only the signal-to-noise ratio and spectral characteristics when making decisions, but also the engineering significance contained in the signal waveform. For example, when filtering signal segments containing clear tension and compression extreme points, special attention should be paid to peak fidelity to improve the fit between the adaptive filtering strategy and the physical mechanism of road mechanical response.

[0104] In some embodiments of the present invention, in order to automatically map the local features of a signal segment to the optimal filtering strategy, this embodiment deploys a preset filtering parameter decision model, which receives signal features and outputs directly executable filtering instructions.

[0105] Specifically, for each signal segment divided in step S200, the filter parameter decision model organizes all the features of the signal segment extracted in step S310 into a structured input feature vector. This input feature vector includes the local signal-to-noise ratio and optionally includes one or more features such as energy-dominant frequency, frequency distribution dispersion, amplitude standard deviation, and kurtosis. This input feature vector is then input into the filter parameter decision model. The filter parameter decision model processes and parses the input feature vector according to its internal algorithm and parameters, and generates a structured decision output. This decision output includes an identifier representing the category of the filtering method and a set of numerical key filtering parameters corresponding to the identifier. This output is the target filtering method and its key filtering parameters applicable to the current signal segment.

[0106] As a further refinement of the above embodiments, the filter parameter decision model is a machine learning model or deep learning model trained based on historical data. Its construction and training is an offline data-driven process, and the specific steps are as follows:

[0107] Step 321: Collect a historical pavement mechanical response dataset covering a wide range of working conditions, including different road grades, vehicle types, vehicle speeds, sensor types, and environmental noise; perform preprocessing, segmentation (step S200), and feature extraction (step S310) on the dataset in accordance with the online process to generate a basic feature sample library; each sample is a feature vector.

[0108] Step 322: Generate a corresponding optimal filtering strategy label for each feature vector sample in the sample library. The generation methods include the following two:

[0109] a) Multiple experts in signal processing and road engineering independently analyze the original signal segment corresponding to each sample in a double-blind environment. The experts comprehensively consider the signal fidelity requirements, noise type, and engineering application purpose, select a filtering method from the candidate method library, and manually adjust its parameters until the optimal filtering effect under their subjective judgment is obtained. The final determined method and parameters are recorded as the label of the sample. The results of multiple experts can be combined by voting or by averaging the parameters.

[0110] b. Define objective evaluation functions, such as the comprehensive signal-to-noise ratio improvement rate and waveform structure similarity index; for each sample's corresponding original signal segment, use automated optimization algorithms, such as grid search, genetic algorithm, or Bayesian optimization, to search in the parameter space composed of moving average method (parameter: step size range 1-20), low-pass filtering method (parameter: cutoff frequency range 5-100Hz, order fixed at 4) and wavelet transform method (parameter: threshold range 0.1-5 times noise standard deviation, decomposition level range 3-10, wavelet basis fixed at db8). The filtering method and its parameter combination that maximizes the evaluation function value will be determined as the label of the sample.

[0111] Step 323: Randomly divide the labeled sample set into training, validation, and test sets; select one or more model architectures for training; one implementation is a machine learning model using a random forest model, the training process of which includes: setting the number of trees to 100, setting the number of features considered when each tree grows to the square root of the total number of features, using Gini impurity as the node splitting criterion, and constructing multiple decision trees on the training set to form a forest; another implementation is a deep learning model, constructing a fully connected feedforward neural network, the number of input layer nodes equal to the feature vector dimension, the output layer designed as a multi-task output, that is, a Softmax classification head outputs the probability distribution of three filtering methods; three independent regression heads correspond to the moving average step size, low-pass cutoff frequency, and wavelet threshold / number of layers respectively; the hidden layer can be set to 2 layers, each with 64 nodes, using the ReLU activation function, and introducing a Dropout layer to prevent overfitting;

[0112] During training, the loss function is minimized using the training set data. During training, performance is monitored on the validation set, and training is stopped when performance no longer improves to prevent overfitting. Finally, the model's accuracy and the mean relative error of parameter predictions are evaluated on an independent test set to confirm the model's generalization ability.

[0113] As a specific limitation on the output result of the aforementioned filtering parameter decision model, the target filtering method is any one of the moving average method, low-pass filtering method, and wavelet transform method; the target filtering method, namely the moving average method, low-pass filtering method, and wavelet transform method, are commonly used and effective noise reduction methods for processing road mechanical response signals in engineering, and together constitute the output option set of the model decision; the moving average method smooths noise by performing moving average calculation in the time domain; the low-pass filtering method filters out high-frequency noise components by setting a cutoff frequency; the wavelet transform method separates noise and signal through multi-scale decomposition and coefficient processing; the model selects the most suitable one from the candidate method set as the output based on the characteristics of the input.

[0114] As a further specification of the output, the filter parameter decision model, while outputting the identifier of the target filtering method, must also output the specific values ​​of the core adjustable parameters required to execute the method, and these values ​​must be adapted to the characteristics of the current signal segment; specifically:

[0115] When the target filtering method identifier of the model output is the moving average method, the key filtering parameter it outputs is the specific adaptive step size N, where N is a positive integer; the adaptive step size is used to determine the width of the sliding window, which affects the smoothing strength and the degree of detail preservation.

[0116] When the target filtering method identifier output by the model is low-pass filtering, the key filtering parameter output simultaneously is the specific adaptive cutoff frequency. The adaptive cutoff frequency is a positive real number in Hertz; the adaptive cutoff frequency is used to define the boundary between the filter's passband and stopband.

[0117] When the target filtering method identifier of the model output is wavelet transform, the key filtering parameters output simultaneously are adaptive threshold and / or adaptive decomposition level; the adaptive threshold is a value used to perform soft or hard thresholding on the wavelet detail coefficients; the adaptive decomposition level is the number of wavelet decomposition levels used to determine the scale range of the analysis.

[0118] The above parameter values ​​are not fixed constants, but rather results inferred in real time by the filter parameter decision model based on the input feature vector. For example, for signal segments with low local signal-to-noise ratio and high frequency distribution dispersion, the filter parameter decision model may output a lower adaptive cutoff frequency value for low-pass filtering or a higher adaptive threshold for wavelet transform to achieve more aggressive noise reduction; conversely, it outputs more lenient parameters to protect the signal.

[0119] In this embodiment, a filter parameter decision model trained on historical data is deployed, and a matching filter method and parameters are generated based on the local signal-to-noise ratio and multi-dimensional inherent characteristics of the signal segment. This data-driven model learns the complex mapping relationship between signal features and optimal filter strategies in massive labeled samples offline. Its decision-making basis comes from the induction of historical experience. Compared with traditional methods that rely on fixed rules or simple thresholds, it can make more accurate judgments on diverse signal patterns and improve the rationality of filter method selection. When making decisions, the model comprehensively considers multiple dimensions of features such as signal-to-noise ratio, frequency distribution, and statistical characteristics, so that the output filter method and key parameters can more precisely adapt to the specific time-frequency structure and noise level of the current signal segment, and realize differentiated processing of signal segments with different signal-to-noise ratios and morphologies.

[0120] In some embodiments of the present invention, step S330 receives the target filtering method identifier and its key filtering parameters output by step S320, and performs corresponding filtering operations on the current signal segment; according to the method identifier output by step S320, a corresponding execution path is selected from the preset filtering algorithm modules; the filtering algorithm modules include a moving average filtering submodule, a low-pass filtering submodule, and a wavelet transform noise reduction submodule; the original data sequence of the current signal segment and the key filtering parameters output by step S320 are transmitted together to the selected submodule.

[0121] When the target filtering method is identified as the moving average method, its key parameter is the adaptive step size N; the filtering execution process is as follows: Let the current signal segment be a discrete sequence S. orig =[x1,x2,…,x M Create a sliding window of length N; starting from the beginning of the sequence, calculate the arithmetic mean of all data points within the window using the following formula:

[0122] ;

[0123] Among them, SMA k This is the new data point in the filtered sequence corresponding to the starting position k of the original window; slide the window backward by one sampling point, repeat the above calculation, until the window covers the end of the sequence, finally obtaining the denoised sequence SMA of length M−N+1; specifically, place the window at the starting position, covering the first N points: [x1,x2,…,x N ; Calculate the arithmetic mean of these N points; Use this mean as the first data point of the filtered SMA sequence; Move the window forward by one sampling point, at which point the points covered by the window become: [x2, x3, ..., x N+1 ]; Calculate the arithmetic mean of these N new points and use it as the second data point for the SMA; Repeat the sliding and calculation process above until the window covers the last N points of the original sequence [x M-N+1 ,…,x M [and complete the calculation; after the above process, the original sequence S of length M is...] orig It is converted into a new SMA sequence of length M-N+1, which is the noise-reduced signal segment.

[0124] When the target filtering method is identified as low-pass filtering, the low-pass filtering submodule is invoked. The low-pass filtering submodule reads the adaptive cutoff frequency value from the key filtering parameters. Based on this adaptive cutoff frequency value, the pre-set sampling frequency, and a fixed filter order, such as 4th order, the Butterworth filter design method is used to calculate the corresponding digital filter coefficients. Subsequently, a digital filter is constructed using these digital filter coefficients to perform zero-phase digital filtering on the data sequence of the current signal segment, including one forward filter and one reverse filter, to eliminate the phase shift introduced by the filter. The numerical value of the adaptive cutoff frequency defines the passband boundary of the filter; signal components with frequencies higher than the adaptive cutoff frequency will be attenuated.

[0125] When the target filtering method is identified as wavelet transform, the wavelet transform denoising submodule is invoked. This submodule reads the adaptive threshold λ and the adaptive decomposition level L from the key filtering parameters, and uses pre-selected wavelet basis functions to perform L-level discrete wavelet decomposition on the data sequence of the current signal segment, obtaining a set of low-frequency approximation coefficients and L sets of high-frequency detail coefficients. Then, thresholding is performed on all high-frequency detail coefficients from level 1 to level L: the absolute value of each detail coefficient is compared with λ; if its absolute value is less than λ, the coefficient is set to zero; if its absolute value is greater than or equal to λ, the coefficient value is shrunk towards zero by the absolute value of λ. Afterward, wavelet reconstruction is performed using the thresholded detail coefficients and the original approximation coefficients to generate the denoised signal sequence. The adaptive threshold λ controls the degree of noise coefficient removal; the larger the λ value, the more coefficients are set to zero or shrunk, resulting in greater denoising strength. The adaptive decomposition level L determines the number of scales at which the signal is decomposed, affecting the depth of analysis for noise in different frequency ranges.

[0126] In this embodiment, by transforming the abstract decision of step S320 into a specific and repeatable signal processing operation and strictly executing it according to the quantization parameters of the decision output, the intensity of the filtering behavior is directly related to the local signal-to-noise ratio level of the signal segment evaluated in step S310. At the same time, the selection of the filtering method is matched with the time-frequency structure characteristics and statistical morphology of the signal segment, ensuring consistency from feature analysis to decision-making and then to the final signal transformation, thus realizing adaptive processing of the signal segment.

[0127] In practical applications, taking the moving average method as an example of the target filtering method, the original signal data curve is as follows: Figure 2 As shown, with sliding step sizes of 5 and 20 respectively, the signal data curves after applying the moving average method are as follows: Figure 3 As shown in Table 1, the differences between the processed data and the original data after different time lengths are as follows:

[0128]

[0129] As shown in Table 1, when the step size is set to 5, the difference between the peak value and the original signal is about 1.2%, which is within the acceptable peak clipping range; while when the step size is 20, the difference reaches 12.0%, and the signal distortion is obvious. Therefore, the adaptive step size N of the model output is preferably dynamically selected around 5 to achieve a balance between noise reduction and fidelity.

[0130] In some embodiments of the present invention, step S400 receives all the denoised signal segments output in step S330 and their time position information in the original signal, and reassembles them into a complete time-domain signal; the execution of step S400 depends on a global signal buffer and a mapping table recording the start and end time indices of each signal segment; according to the order in which the signal segments are processed or according to their time index, the data of each denoised signal segment is written to the corresponding position in the global signal buffer; this position is uniquely determined by the start sampling point index and end sampling point index of the signal segment in the original mechanical response time-domain signal; the specific implementation process is as follows:

[0131] First, initialize an empty array of the same length as the original mechanical response time-domain signal as the final output signal sequence; then, iterate through all signal segments generated in step S200 and processed in step S300; for each signal segment, read its corresponding original starting sampling point index t from the mapping table. start and the original end sampling point index t end Next, the denoised data sequence of the signal segment output in step S330 is sequentially assigned to the index t in the final output signal sequence according to its internal order. start to t end The corresponding position; since the segmentation strategy of step S200 ensures that each signal segment is continuous in time and does not overlap, this assignment process will not cause data conflict or overwriting.

[0132] After traversing all signal segments and completing the assignment, all index positions in the output signal sequence are filled with the corresponding denoised data, thus forming a continuous and complete adaptive denoised signal on the time axis. The total duration and sampling rate of this signal are consistent with the original signal, but the data in each local time period has undergone filtering processing adapted to its own characteristics.

[0133] In this embodiment, through precise mapping and data replacement based on time index, the local denoising results processed in the previous steps are restored to a globally continuous signal without distortion, preserving all time coordinate information of the original signal, so that any time-point-based engineering analysis can be accurately performed on the denoised signal.

[0134] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. An adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio, characterized in that, include: Collect the time-domain signal of the original mechanical response of the pavement structure under vehicle load; The original mechanical response time-domain signal is segmented to obtain multiple continuous signal segments; For each of the aforementioned signal segments, the following processing flow is executed sequentially: Extract the local signal-to-noise ratio and inherent signal characteristics of the signal segment; The inherent features of the signal are used to characterize the time-frequency structure and distribution characteristics of the signal; extracting the local signal-to-noise ratio of the signal segment includes: identifying and separating the background noise signal segment characterizing the absence of vehicle load from the original mechanical response time-domain signal; calculating the amplitude statistics of the background noise as a background noise reference based on the background noise signal segment; and calculating the local signal-to-noise ratio based on the background noise reference and the amplitude statistics of the current signal segment. The local signal-to-noise ratio features and the inherent features of the signal are taken as inputs and fed into a preset filter parameter decision model, which outputs a target filtering method applicable to the signal segment and its corresponding key filtering parameters. The target filtering method and at least one corresponding key filtering parameter are used to filter the signal segment to obtain a noise-reduced signal segment. All the denoised signal segments are reassembled in sequence to generate the final adaptive denoised signal.

2. The adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio according to claim 1, characterized in that, The inherent characteristics of the signal include at least one of local frequency characteristics and local statistical characteristics.

3. The adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio according to claim 2, characterized in that, The extraction of the local frequency features includes: performing time-frequency analysis on the signal segment and calculating the dispersion of its energy-dominant frequency band and / or frequency component distribution; The extraction of the local statistical features includes: calculating at least one of the amplitude standard deviation, kurtosis, or waveform factor of the signal segment.

4. The adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio according to claim 3, characterized in that, Identify and separate the background noise signal segment representing the absence of vehicle load from the original mechanical response time-domain signal, including: Identify the start and end points of the vehicle load action in the original mechanical response time-domain signal; The signal segments within the first preset time period before the starting point and the signal segments within the second preset time period after the ending point are determined as the background noise signal segments.

5. The adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio according to claim 4, characterized in that, The identification of the starting point and the ending point is based on at least one of the following criteria: The signal amplitude exceeds the preset amplitude threshold; The zero-crossing points of the first derivative of a signal are combined with the extreme points of the second derivative of a signal.

6. The adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio according to claim 1, characterized in that, The filter parameter decision model is a machine learning model or a deep learning model trained based on historical data; The training data for the filter parameter decision model includes a sample set consisting of signal segments with different signal-to-noise ratios, local frequency characteristics, and local statistical characteristics, as well as the optimal filtering method and filter parameter labels for each sample set, determined by experts or optimization algorithms.

7. The adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio according to claim 6, characterized in that, The target filtering method is any one of the following: moving average method, low-pass filtering method, and wavelet transform method.

8. The adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio according to claim 7, characterized in that, When the target filtering method is the moving average method, the corresponding key filtering parameter is the adaptive step size; When the target filtering method is a low-pass filter, the corresponding key filtering parameter is the adaptive cutoff frequency. When the target filtering method is wavelet transform, the corresponding key filtering parameters are adaptive threshold and / or number of decomposition layers.

9. The adaptive noise reduction and filtering method for road surface mechanical response signals based on signal-to-noise ratio according to claim 1, characterized in that, All the denoised signal segments are reassembled in sequence to generate the final adaptive denoised signal, including: Initialize an output signal sequence with the same length as the original mechanical response time-domain signal; Based on the time index of each signal segment in the original mechanical response time domain signal, the data of each denoised signal segment is sequentially filled into the corresponding position of the output signal sequence; Generates a temporally continuous and complete adaptive noise reduction signal.

Images