An embedded system engineering off-road surface recognition method based on vehicle response

Abstract

The application provides an embedded system engineering off-road surface identification method based on vehicle response, which comprises the steps of data labeling and time synchronization, data feature extraction, data feature engineering, roadside identification classifier construction and the like, the sum average step in the calculation of the time domain and frequency domain statistical features commonly used in the time series data feature extraction is decomposed into several unit sum averages by the unit normalization method, the data cache amount required by the sum average step is reduced, the dynamic large array cache for the dynamic storage pressure of the embedded system is effectively solved, the original signal is decomposed into F different frequency band signals by constructing low-pass, band-pass and high-pass filters of different frequency bands through the frequency band energy method, and the amplitude in the continuous K sampling points is calculated based on the unit normalization method for each frequency band signal, the storage pressure of the original data dynamic cache is reduced, and the occupied space of the time series data amplitude frequency spectrum calculation result is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to an off-road surface recognition method, in particular to an embedded system engineering off-road surface recognition method based on vehicle response. BACKGROUND

[0002] Road surface information is a necessary input for intelligent off-road vehicle planning, decision-making and control. With the development of machine learning technology, off-road surface recognition methods are used to provide intelligent off-road vehicles with road surface category information and corresponding road surface feature experience mapping. In related technologies, off-road surface recognition methods for vehicles mainly include methods based on vehicle response and methods based on environmental sensors. The method based on environmental sensors relies on a large amount of data and a complex road surface recognition network, and its computing power requirement is difficult to meet the computing power limit of the vehicle embedded system and the real-time requirement in actual operation. In the method based on vehicle response, although there are some methods with lower computing power requirements such as finite state machine, fuzzy rule, decision tree, and random forest, these methods are mostly based on the statistical characteristics of a certain length of historical data, requiring the system to cache a certain length of historical data, i.e. a dynamic large array, which puts pressure on the storage capacity of the embedded system. In addition, for frequency domain statistical characteristics, the Fourier transform involved in frequency domain calculation also consumes a lot of computing resources of the embedded system.

[0003] Therefore, it is necessary to establish an embedded system engineering off-road surface recognition method based on vehicle response, which can provide road surface recognition results based on vehicle response while reducing the historical data caching requirement and the computing power requirement of frequency domain calculation. SUMMARY

[0004] To solve the above technical problems, the present application provides an embedded system engineering off-road surface recognition method based on vehicle response, comprising the following steps:

[0005] (1) Data labeling and time synchronization

[0006] The original vehicle CAN bus signals and inertial navigation system signals collected by the vehicle data collection platform on typical off-road surfaces are used as inputs. The collected vehicle CAN bus signals and inertial navigation system signals are time-synchronized, and an interpolation method is used to make all signal sampling periods f Hz. According to the road surface type of the data collection site, the interpolated vehicle CAN bus signals and inertial navigation system signals are divided into several continuous data segments, and the ground category to which each data segment belongs is labeled. The interpolated signals and the corresponding actual ground categories form an interpolated typical terrain driving data set.

[0007] (2) Data feature extraction

[0008] The vehicle driving state related signals in the interpolated typical terrain driving data set are taken as inputs to estimate the vehicle-wheel-road interaction state, the time series data of the vehicle driving state and the vehicle-wheel-road interaction state are extracted into a plurality of samples with a specified data length by a sliding window method, a total of m0 time domain and frequency domain features and the corresponding actual ground categories in each sample are extracted to form a first feature data set; the first feature data set is subjected to feature analysis and screening, and the main features related to the ground category are screened out, and a total of m1 features in the screened first feature data set and the corresponding actual ground categories form a second feature data set;

[0009] (3) Data feature engineering

[0010] The second feature data set formed after feature analysis and screening is taken as input, and m1 engineering features corresponding to the selected m1 time domain and frequency domain statistical features are calculated: the engineering features of the time domain features are calculated based on the unit normalization method; the effective frequency band of the signal is divided into a plurality of contiguous but non-overlapping frequency bands, and the engineering features of the frequency domain features are calculated based on the sub-band energy method and the unit normalization method; the m1 engineering feature data calculated from the second feature data set together with the corresponding actual ground categories form an engineering feature data set;

[0011] The engineering features of the time domain features calculated based on the unit normalization method take time series data as input, and the summation and averaging step in the time domain and frequency domain statistical feature calculation commonly used in feature extraction is decomposed into a plurality of unit summation and averaging, and the sampling point number K of the statistical data segment is reduced to k is the sampling point number of the engineering data unit;

[0012] The engineering features of the frequency domain features calculated based on the sub-band energy method and the unit normalization method take time series signals as input, divide the effective frequency band of the signal into a plurality of contiguous but non-overlapping frequency bands, decompose the original signal into F different frequency band signals by constructing low-pass, band-pass and high-pass filters for different frequency bands, and calculate the amplitude within K consecutive sampling points for each frequency band signal based on the unit normalization method to obtain a frequency band-amplitude table as an engineering representation of the amplitude spectrum obtained after fast Fourier transform of the original signal, and the time series data amplitude spectrum calculation result is reduced to F.

[0013] (4) Road surface recognition classifier construction

[0014] The m1 engineering features in the engineering feature data set are taken as input, the corresponding actual ground categories are taken as labels, the difference between the labels and the off-road road surface classifier prediction results is taken as a loss function to train the off-road road surface classifier, and the trained off-road road surface classifier takes the m1 engineering features in the engineering feature data set as input and takes the recognized off-road road surface category as output. ​

[0015] Furthermore, in step (2) data feature extraction, the vehicle driving state related signals in the interpolated typical terrain driving dataset are used as input to construct vehicle kinematics and dynamics models, perform vehicle-wheel-road interaction state estimation, and extract vehicle driving state and vehicle-wheel-road interaction state data segments of a specified data length using the sliding window method. Each data segment is regarded as a sample. For a sample with a data length of T seconds, the data in each sample corresponds to T*f sampling points. The m0-dimensional time domain and frequency domain statistical features of the n-dimensional data in each sample are extracted respectively, and together with the corresponding actual ground category, they form the first feature dataset.

[0016] Feature analysis and filtering are performed on the first feature dataset. Based on the correlation between features and the information entropy of features, the main features related to the ground category are filtered out. After filtering, a total of m1 features in the first feature dataset and the corresponding actual ground categories form the second feature dataset.

[0017] Furthermore, for a certain dimension of information X of a vehicle driving state and vehicle-wheel-road interaction state sample, its time domain characteristics include, but are not limited to, mean, peak-to-peak, variance, and root mean square; a fast Fourier transform is performed on a certain dimension of information X of a vehicle driving state and vehicle-wheel-road interaction state sample to obtain the spectrum S(f) of the vehicle-wheel-road interaction state information, and its calculated frequency domain characteristics include, but are not limited to, centroid frequency, mean square frequency, and root mean square frequency.

[0018] Furthermore, in step (3) data feature engineering, the engineering feature steps for calculating time-domain features based on the unit normalization method include:

[0019] (3.1.1) Select a unit length U0, where the unit length is the measurement length of the unit data. For time measurement, the unit length is U0 = t (s), and for spatial measurement, the unit length is U0 = l (m).

[0020] (3.1.2) Calculate the interval length u between two sampling points. t Establish a unit cumulative counter: U t =U t-1 +u t The aforementioned unit cumulative counter, for a time metric such as a timer, u t The sampling period is u; for spatial metrics, odometry is u. t The distance between the two sampling points in space;

[0021] (3.1.3) In the calculation of statistical characteristics, forms such as The summation steps use a loop accumulator to accumulate the sum: X t =X t-1 +xt ;

[0022] (3.1.4) Unit accumulation count is incremented each time the unit length is reached, the unit sum result and the current unit accumulation count are emitted, while the unit accumulation counter and the loop accumulator are reset:

[0023] if U t >=U0

[0024] return U unit =U t ,X unit =X t

[0025] U t ←0,X t ←0

[0026] else

[0027] U t =U t-1 +u t ,X t =X t-1 +x t

[0028] (3.1.5) The emitted unit sum result X unit is normalized according to the unit accumulation count U unit :

[0029]

[0030] (3.1.6) According to the required statistical data segment length U stat , the buffered unit normalized sum result X′ unit is computed:

[0031] (3.1.7) The buffered unit normalized sum results are summed again and averaged as an engineered replacement for the averaging step in the statistical feature computation, like:

[0032]

[0033] (3.1.8) Engineering features for time domain features are computed.

[0034] Further, in the step (3) data feature engineering, the engineering feature step of computing frequency domain features based on the sub-band energy method and unit normalization method comprises:

[0035] (3.2.1) The signal effective frequency band (f begain , f end ​​​), is divided into several bands that are connected but not overlapped:

[0036] (f begain ,f end )→{(f begain ,f1),(f1,f2),...(f n ,f end )}

[0037] (3.2.2) Design a band-pass filter for each band, and decompose the original signal into several signals belonging to different bands;

[0038] (3.2.3) Calculate the amplitude of the sub-band signal xf in the statistical data segment using the unit normalization method:

[0039]

[0040] (3.2.4) For each band and its sub-band signal amplitude, form a band-amplitude table as an engineering representation of the amplitude spectrum obtained after the original signal is subjected to fast Fourier transform:

[0041]

[0042] (3.2.5) Based on the band-amplitude table, calculate the engineering features of the frequency domain features.

[0043] The beneficial effects of the present application are:

[0044] The present application decomposes the summation average step in the calculation of time domain and frequency domain statistical features commonly used in time series data feature extraction into several unit summation averages, reducing the data cache required for the summation average step from K to K is the number of sampling points of the statistical data segment, and k is the sampling point number of the engineering data unit, effectively solving the dynamic storage pressure of dynamic large array cache on embedded systems.

[0045] The present application uses the sub-band energy method to decompose the original signal into F different band signals by constructing low-pass, band-pass and high-pass filters for different bands, and calculates the amplitude in the continuous K sampling points based on the unit normalization method for each band signal, obtaining a band-amplitude table as an engineering representation of the amplitude spectrum obtained after the original signal is subjected to fast Fourier transform, reducing the storage pressure of the original data dynamic cache while reducing the occupied space of the time series data amplitude spectrum calculation result from to F, effectively solving the problem of insufficient running efficiency caused by the calculation resource occupation pressure of fast Fourier transform on embedded systems in the frequency domain calculation of time series data. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0047] Figure 2 This is a schematic diagram of the data feature extraction process of the present invention.

[0048] Figure 3 This is a flowchart illustrating the engineering features of the time-domain features calculated based on the unit normalization method of the present invention.

[0049] Figure 4 This is a schematic diagram of the engineering features of the time-domain features calculated based on the unit normalization method of the present invention.

[0050] Figure 5 This is a flowchart illustrating the engineering characteristics of frequency domain features calculated based on the frequency band energy method and the unit normalization method of the present invention.

[0051] Figure 6 This is a schematic diagram comparing the engineered alternative amplitude spectrum and the original data amplitude spectrum of the frequency domain characteristics calculated based on the frequency band energy method and the unit normalization method of this invention. Detailed Implementation

[0052] See Figures 1-6 As shown:

[0053] This embodiment provides an engineered off-road surface recognition method for embedded systems based on vehicle response, comprising the following steps:

[0054] (1) Data labeling and time synchronization

[0055] Using the original vehicle CAN bus signal and inertial navigation system signal collected by the vehicle data acquisition platform on a typical off-road surface as input, the collected vehicle CAN bus signal and inertial navigation system signal are time-synchronized. Interpolation is used to make the sampling period of all signals f Hz. According to the road surface type of the data acquisition location, the interpolated vehicle CAN bus signal and inertial navigation system signal are divided into several continuous data segments, and the ground category to which each data segment belongs is labeled. The interpolated signal and the corresponding actual ground category together form the interpolated typical terrain driving dataset.

[0056] (2) Data feature extraction

[0057] The vehicle driving state related signals in the interpolated typical terrain driving data set are taken as inputs, a vehicle kinematics and dynamics model is constructed, vehicle-wheel-road interaction state estimation such as driving resistance and wheel speed fluctuation is performed, a sliding window method is used to slide and extract vehicle driving state and vehicle-wheel-road interaction state data segments with a specified data length, each data segment is regarded as a sample, for a sample with a data length of T seconds, the data in each sample corresponds to T*f sampling points, n-dimensional data in each sample is extracted to obtain m0-dimensional time domain and frequency domain statistical features, and the corresponding actual ground categories are used to form a first feature data set, the time domain features of a certain dimension X of a vehicle driving state and vehicle-wheel-road interaction state sample include but are not limited to mean value, peak-peak value, variance, and root mean square, the frequency spectrum S(f) of the vehicle-wheel-road interaction state information is obtained by performing fast Fourier transform on a certain dimension X of a vehicle driving state and vehicle-wheel-road interaction state sample, and the frequency domain features calculated according to the frequency spectrum S(f) include but are not limited to center of gravity frequency, mean square frequency, and root mean square frequency.

[0058] Feature analysis and screening are performed on the first feature data set, main features related to the ground categories are screened according to the correlation between the features and the information entropy of the features, a total of m1 features in the screened first feature data set and the corresponding actual ground categories form a second feature data set.

[0059] It should be noted that the feature analysis and screening method can be any feature analysis and screening method such as principal component analysis and factor analysis, which is a common technology known to those skilled in the art. The vehicle-wheel-road interaction state estimation method and the time domain and frequency domain statistical feature calculation method are common technologies mastered by those skilled in the related field. Therefore, the feature analysis and screening process, the vehicle-wheel-road interaction state estimation method, and the basic calculation method of the time domain and frequency domain statistical features will not be described in detail.

[0060] (3) Data feature engineering

[0061] The second feature data set formed after feature analysis and screening is taken as input, m1 engineering features corresponding to the selected m1 time domain and frequency domain statistical features are calculated: the engineering features of the time domain features are calculated based on the unit normalization method; the effective frequency band of the signal is divided into several connected but non-overlapping frequency bands, the engineering features of the frequency domain features are calculated based on the sub-band energy method and the unit normalization method; and the m1 engineering feature data calculated from the second feature data set are used to form an engineering feature data set together with the corresponding actual ground categories.

[0062] The engineered feature based on the unit normalization method for calculating time-domain features uses time-series data as input. It decomposes the summation and averaging steps commonly used in time-domain and frequency-domain statistical feature extraction into summation and averaging of several units, reducing the number of sampling points K of the statistical data segment to [missing information]. (Generally, K>50), where k is the number of sampling points for the engineered data unit; in this embodiment, an engineered calculation method for the time-domain statistical characteristic of data Z—variance var(Z)—is provided, and the length of the statistical data U is... stat =5s, cell length U0=1s, data sampling time is 0.1s, the engineering feature steps for calculating time-domain features based on cell normalization method include:

[0063] (3.1.1) Select an element length U0 = 1s;

[0064] (3.1.2) Calculate the time interval u between the two sampling points. t =0.1s, establish the unit cumulative counter U t =U t-1 +u t The unit cumulative counter described in this embodiment, for a time metric of time, is a timer, u t The sampling period;

[0065] (3.1.3) The formula for calculating the variance of the time-domain statistical characteristics of data Z is: Among them, make Summation steps Use a loop accumulator to sum up X. t =X t-1 +x t ;make Summation steps Use a loop accumulator to sum up Y. t =Y t-1 +y t ;

[0066] (3.1.4) Each time the cell cumulative count reaches the cell length, the cell summation result and the current cell cumulative count are sent out, and the cell cumulative counter and the cyclic accumulator are reset:

[0067] if U t >=1s

[0068] return U unit =U t ,X unit =X t ,Y unit =Y t

[0069] U t ←0,Xt ← 0, Y t ← 0

[0070] else

[0071] U t = U t-1 + u t , X t = X t-1 + x t , Y t = Y t-1 + y t

[0072] (3.1.5) Summing the emitted cell results X unit According to the cell accumulation count U unit Normalizing:

[0073]

[0074] (3.1.6) According to the statistical data segment length U stat = 5s, buffer cell normalized sum results X u ' nit,i , Y u ' nit,i , i = 0, 1, 2, 3, 4;

[0075] (3.1.7) Summing and averaging the buffered cell normalized sum results again as an engineered replacement for the averaging step in the statistical feature calculation, of the form

[0076]

[0077] (3.1.8) Computing the engineering calculation of the time domain statistical feature variance of the data Z at the statistical data length U stat = 5s as:

[0078] The engineering feature based on the frequency band energy method and the cell normalization method for calculating the frequency domain feature, taking the time sequence signal as the input, divides the effective frequency band of the signal into several connected but non-overlapping frequency bands, decomposes the original signal into F different frequency band signals (generally, F < 10) by constructing low-pass, band-pass, and high-pass filters for different frequency bands, and calculates the amplitude within K consecutive sampling points (generally, K > 50) for each frequency band signal based on the cell normalization method, to obtain a frequency band-amplitude table as an engineering representation of the amplitude spectrum obtained after fast Fourier transform of the original signal. The calculation result of the time sequence data amplitude spectrum is ​The reduction is F; in this embodiment, the engineering calculation method of the frequency domain statistical feature of the data Z, the center of gravity frequency FC, is provided, and the statistical data length U stat = 5s, the unit length U0 = 1s, and the data sampling time is 0.1s. The engineering feature step of calculating the frequency domain feature based on the frequency band energy method and the unit normalization method is as follows:

[0079] (3.2.1) The signal effective frequency band (f begain = 1Hz, f end = 10Hz) is divided into several connected but non-overlapping frequency bands:

[0080] (1, 10) → {(1, 4), (4, 7), (7, 10)}

[0081] (3.2.2) A band-pass filter is designed for each frequency band to decompose the original signal into several signals Z (1,10) → {Z (1,4) ,Z (4,7) ,Z (7,10)};

[0082] (3.2.3) The unit normalization method is used to calculate the amplitude of the frequency band signal Z (1,4) ,Z (4,7) ,Z (7,10) in the statistical data segment:

[0083]

[0084]

[0085]

[0086] (3.2.4) For each frequency band and the amplitude of the frequency band signal, a frequency band-amplitude table is formed as an engineering representation of the amplitude spectrum obtained after the fast Fourier transform of the original signal:

[0087]

[0088] (3.2.5) The engineering calculation of the frequency domain statistical feature of the data Z in the statistical data length U stat = 5s, the center of gravity frequency FC, is:

[0089]

[0090] (4) Road identification classifier construction

[0091] The off-road surface classifier is trained by taking m1 engineering features in the engineering feature dataset as input, taking the corresponding actual ground category as label, and taking the difference between the label and the prediction result of the off-road surface classifier as the loss function. The trained off-road surface classifier takes m1 engineering features in the engineering feature dataset as input and outputs the identified off-road surface category.

[0092] It should be noted that the off-road surface classifier can be any basic classifier model such as random forest, support vector machine, etc., which is a common technology known to those skilled in the art. Therefore, the selection and training process of the off-road surface classifier will not be described here.

Claims

1. An engineered off-road surface recognition method for embedded systems based on vehicle response, characterized in that: Includes the following steps: (1) Data labeling and time synchronization Using the original vehicle CAN bus signal and inertial navigation system signal collected by the vehicle data acquisition platform on a typical off-road surface as input, the collected vehicle CAN bus signal and inertial navigation system signal are time-synchronized. Interpolation is used to make the sampling period of all signals f Hz. According to the road surface type of the data acquisition location, the interpolated vehicle CAN bus signal and inertial navigation system signal are divided into several continuous data segments, and the ground category to which each data segment belongs is labeled. The interpolated signal and the corresponding actual ground category together form the interpolated typical terrain driving dataset. (2) Data feature extraction Using the vehicle driving state-related signals from the interpolated typical terrain driving dataset as input, the vehicle-wheel-road interaction state is estimated. The time-series data of vehicle driving state and vehicle-wheel-road interaction state are extracted into several samples of specified lengths using the sliding window method. A total of m0 time-domain and frequency-domain features and their corresponding actual ground categories are extracted from each sample to form the first feature dataset. The first feature dataset is subjected to feature analysis and filtering to select the main features related to the ground category. The m1 features in the first feature dataset and their corresponding actual ground categories are used to form the second feature dataset. (3) Data feature engineering Using the second feature dataset formed after feature analysis and filtering as input, calculate the corresponding m1 engineered features for the selected m1 time-domain and frequency-domain statistical features: engineered features for time-domain features calculated based on the unit normalization method; The effective frequency band of the signal is divided into several connected but non-overlapping frequency bands. The engineering features of the frequency domain characteristics are calculated based on the frequency band energy method and the unit normalization method. The engineering feature dataset is formed by the m1 types of engineering feature data calculated based on the second feature dataset and the corresponding actual ground categories. The engineered feature based on the unit normalization method for calculating time-domain features uses time-series data as input. It decomposes the summation and averaging steps commonly used in time-domain and frequency-domain statistical feature extraction into summation and averaging of several units, reducing the number of sampling points K of the statistical data segment to [missing information]. k is the number of sampling points for the engineered data unit; The engineering feature based on the frequency band energy method and the unit normalization method for calculating frequency domain characteristics takes a time-series signal as input. It divides the effective frequency band of the signal into several interconnected but non-overlapping frequency bands. By constructing low-pass, band-pass, and high-pass filters for different frequency bands, the original signal is decomposed into F different frequency band signals. For each frequency band signal, the amplitude within K consecutive sampling points is calculated based on the unit normalization method, resulting in a frequency band-amplitude table as an engineering representation of the amplitude spectrum obtained after the original signal undergoes a Fast Fourier Transform. The amplitude spectrum calculation results of the time-series data are obtained from... Reduced to F; (4) Construction of Road Surface Recognition Classifier Using m1 engineered features from the engineered feature dataset as input and the corresponding actual ground category as label, the off-road surface classifier is trained using the difference metric between the label and the prediction result of the off-road surface classifier as loss function. The trained off-road surface classifier takes m1 engineered features from the engineered feature dataset as input and the identified off-road surface category as output.

2. The embedded system engineering off-road surface recognition method based on vehicle response according to claim 1, characterized in that: In step (2) data feature extraction, the vehicle driving state related signals in the interpolated typical terrain driving dataset are used as input to construct vehicle kinematics and dynamics models, perform vehicle-wheel-road interaction state estimation, and extract vehicle driving state and vehicle-wheel-road interaction state data segments of a specified data length by sliding window method. Each data segment is regarded as a sample. For a sample with a data length of T seconds, the data in each sample corresponds to T*f sampling points. For each sample, m0-dimensional time-domain and frequency-domain statistical features are extracted from the n-dimensional data, and these features, together with the corresponding actual ground category, form the first feature dataset. Feature analysis and filtering are performed on the first feature dataset. Based on the correlation between features and the information entropy of features, the main features related to the ground category are filtered out. After filtering, a total of m1 features in the first feature dataset and the corresponding actual ground categories form the second feature dataset.

3. The embedded system engineering off-road surface recognition method based on vehicle response according to claim 2, characterized in that: For a vehicle driving state and a vehicle-wheel-road interaction state sample, one dimension of information X has time-domain features including but not limited to mean, peak-to-peak value, variance, and root mean square (RMS). A fast Fourier transform is performed on one dimension of information X to obtain the spectrum S(f) of the vehicle-wheel-road interaction state information. The frequency domain features calculated based on the spectrum S(f) include but are not limited to centroid frequency, MMS frequency, and RMS frequency.

4. The embedded system engineering off-road surface recognition method based on vehicle response according to claim 1, characterized in that: In step (3) of data feature engineering, the engineering feature steps based on the unit normalization method for calculating time-domain features include: (3.1.1) Select a unit length U0, where the unit length is the measurement length of the unit data. For time measurement, the unit length is U0 = t (s), and for spatial measurement, the unit length is U0 = l (m). (3.1.2) Calculate the interval length u between two sampling points. t Establish a unit cumulative counter: U t =U t-1 +u t The aforementioned unit cumulative counter, for a time metric such as a timer, u t The sampling period is u; for spatial metrics, odometry is u. t The distance between the two sampling points in space; (3.1.3) In the calculation of statistical characteristics, forms such as The summation steps use a loop accumulator to accumulate the sum: X t =X t-1 +x t ; (3.1.4) Each time the cell cumulative count reaches the cell length, the cell summation result and the current cell cumulative count are sent out, and the cell cumulative counter and the cyclic accumulator are reset: (3.1.5) Sum the issued cell result X unit According to the cumulative count U of the units unit Normalization: (3.1.6) Based on the statistical data of the demand, the segment length U stat ,cache The result of the normalized summation of each unit X u ′ nit ; (3.1.7) Regarding caching The normalized summation results of each unit are summed again and averaged, and used as a statistical feature in calculations, in the form of... Engineering alternatives to the average steps: (3.1.8) Engineering features for calculating time-domain features.

5. The embedded system engineering off-road surface recognition method based on vehicle response according to claim 1, characterized in that: In step (3) of data feature engineering, the engineering feature steps for calculating frequency domain features based on the frequency band energy method and the unit normalization method include: (3.2.1) The effective frequency band of the signal (f begain ,f end The frequency bands are divided into several interconnected but non-overlapping bands: (f begain ,f end )→{(f begain ,f1),(f1,f2),...(f n ,f end )} (3.2.2) Design a bandpass filter for each frequency band to decompose the original signal into several signals belonging to different frequency bands; (3.2.3) Calculate the amplitude of the frequency band signal xf within the statistical data segment using the unit normalization method: (3.2.4) For each frequency band and its sub-band signal amplitude, a frequency band-amplitude table is formed as an engineering representation of the amplitude spectrum obtained after the original signal undergoes a fast Fourier transform: (3.2.5) Based on the frequency band-amplitude table, calculate the engineering characteristics of the frequency domain features.

Images