Multi-strategy road flatness detection method

Abstract

The invention provides a multi-strategy road flatness detection method, which comprises the following steps of: (1) updating and correcting acceleration data acquired and measured by an acceleration sensor of a vehicle-mounted smart phone in real time by adopting an attitude self-correction algorithm so as to adapt to the dynamic change of a vehicle in a driving process; (2) utilizing a multi-scale fuzzy entropy algorithm to carry out multi-scale processing on measurement data which is collected by a vehicle-mounted smart phone and reflects the road flatness in a vehicle driving process, constructing a road flatness representation model, and combining global features and local features; and (3) adopting a self-attention mechanism convolutional neural network to carry out road flatness identification on acceleration data collected by a vehicle-mounted smart phone in a vehicle driving process. According to the invention, real-time accurate detection of the road flatness can be well realized.

Description

Technical Field

[0001] The present invention relates to the technical field of road detection, and in particular to a multi-strategy road flatness detection method. Background Art

[0002] Roads are infrastructure, and the road surface is the direct carrier of moving vehicles in the highway network. Therefore, the road surface is required to meet certain functional and structural performances.

[0003] The functional performance of the road surface mainly includes safety performance and driving quality. From the perspective of road engineering, the flatness of the road surface is the main factor affecting the driving quality of the road surface. In addition to affecting the driving quality of the road surface, the uneven road surface will also cause additional dynamic loads on the vehicle tires, leading to accelerated damage to automobile parts, tire wear and increased fuel consumption, and reduced vehicle service life; in addition, it is not conducive to road drainage, accelerating road water damage and affecting driving safety. At the same time, because the additional dynamic load of the vehicle on the road surface becomes larger and the road surface is flooded, the road surface damage is aggravated and rutting occurs. For this reason, the accuracy of road flatness detection is extremely important, which has guiding significance for building an effective long-term system for road maintenance and management. Summary of the invention

[0004] The purpose of the present invention is to provide a multi-strategy road flatness detection method to achieve real-time and accurate detection of road flatness.

[0005] The present invention adopts the following technical solutions:

[0006] A multi-strategy road roughness detection method comprises the following steps:

[0007] (1) Using a posture self-correction algorithm, the acceleration data collected and measured by the vehicle-mounted smartphone acceleration sensor is updated and corrected in real time to adapt to the dynamic changes of the vehicle during driving;

[0008] (2) Using a multi-scale fuzzy entropy algorithm, the measurement data reflecting the road roughness collected by the on-board smartphone during vehicle driving is processed at multiple scales to construct a road roughness representation model and combine global features with local features.

[0009] (3) A convolutional neural network with a self-attention mechanism is used to identify the road roughness based on the acceleration data collected by the onboard smartphone during vehicle driving.

[0010] The specific implementation process of step (1) is as follows:

[0011] S1. Establish the vehicle coordinate system and the mobile phone coordinate system:

[0012] In the vehicle coordinate system, the longitudinal vibration of the vehicle is defined as the Y direction, with the driving direction as positive, the lateral vibration of the vehicle is defined as the X direction, with the right side of the driving direction as positive, and the vertical vibration of the vehicle is defined as the Z direction, with the upward direction as positive; in the mobile phone coordinate system, the vehicle-mounted smartphone is set vertically. Ideally, the screen of the vehicle-mounted smartphone is perpendicular to the direction of vehicle travel. In non-ideal situations, the direction from the bottom to the top of the vehicle-mounted smartphone is defined as the positive y direction, from left to right as the positive x direction, and perpendicular to the screen outward as the positive z direction;

[0013] S2. Vertical posture correction of the in-vehicle smartphone:

[0014] 1. Measured vehicle-mounted smartphone acceleration data, vector representation (a x ,a y ,a z, ), then the acceleration value in the vehicle coordinate system is expressed as (a X ,a Y ,a Z, );

[0015] 2. Vertical acceleration correction: There is an angle deviation between the vehicle-mounted smartphone and the vehicle body coordinate system. When the vehicle is parked, the acceleration vector V of the vehicle-mounted smartphone in three directions is expressed as (x g ,y g ,z g ), the resultant vector is the gravitational acceleration, recorded as:

[0016]

[0017] Where: (x g ,y g ,z g ) T They are the three components of gravity acceleration in the x, y, and z coordinates of the smartphone;

[0018] 3. The vertical unit vector Z of the vehicle coordinate system is (0, 0, 1). The mobile phone acceleration vector V is rotated to the vertical unit vector Z. The transformation matrix R is determined so that the relationship shown in formula (2) is satisfied:

[0019] Z T =R·V (2)

[0020] According to the Rodriguez rotation formula,

[0021] R(k,φ)=I·cosφ+Ksinφ+(1―cosφ)u·u T (3)

[0022] Among them: after rotation, u is the cross product of the vertical unit vector Z and the mobile phone acceleration vector V, and is normalized:

[0023] K is the cross product matrix of the rotation axis, then:

[0024]

[0025] The rotation angle φ between the unit vector Z and the mobile phone acceleration vector v is:

[0026] φ=arccos(VZ / |V||Z|) (6)

[0027] The multi-scale fuzzy entropy algorithm in step (2) first processes the corrected original acceleration according to the time series, and then calculates the fuzzy entropy value of the sequence. The calculation process is as follows:

[0028] S1: original acceleration sequence, original acceleration sequence U with N points after rotation correction i ={u1,u,…,u N} According to the scale factor τ=1,2,…,N, we can get the new vector:

[0029]

[0030] When τ is 1, y j (1) is the original sequence U i , τ converts the original sequence U i Split into τ sequences y of length N / τ each i (τ);

[0031] S2: For the sequence Y(i)={y(1),y(2)…,y(τ)} after rotation correction and segmentation, first determine the vector dimension m, and then reconstruct the time series Y to obtain a new sequence:

[0032]

[0033] Subtract the mean y0(i) from the continuous m elements starting from the i-th point, where m is the embedding dimension, and we get:

[0034]

[0035] S3: Define two reconstruction vectors and The maximum distance between

[0036]

[0037] S4: The fuzzy membership function is:

[0038]

[0039] Among them, n is its boundary gradient, r is the similarity tolerance, which is generally 0.1 to 0.25SD (SD is the standard deviation of the original data;

[0040] S5: Define the metric function to evaluate the relevance:

[0041]

[0042] S6: Calculate the mean of the metrics:

[0043]

[0044] S7: After the vector dimension is m+1, repeat Step1-Step5 to get δ m+1 (n,r):

[0045]

[0046] S8: Define fuzzy entropy as:

[0047]

[0048] The specific steps of step (3) are as follows:

[0049] S1: Visualization of acceleration information features. The acceleration information is transformed into a time-frequency spectrum using a short-time Fourier transform, and then the time-frequency spectrum is transformed into a Mel-scaled time-frequency spectrum through a Mel-scaled filter bank. To enhance data validity, the spectrum is pre-emphasized.

[0050] S2: Spectral feature extraction, using the feature extraction capability of convolutional neural network, setting the convolution kernel according to the characteristic characteristics, and performing linear sliding operation on the time-frequency spectrum to obtain the features;

[0051] S3: Feature enhancement, using the self-attention mechanism to enhance the spectrogram features and improve the effectiveness of acceleration information features;

[0052] S4: Use equations (16)-(17) to express the established road roughness information:

[0053] y(t)=x(t)+a×x(t―1) (16)

[0054]

[0055] Where: x(t) is the acceleration value after rotation multi-scale processing at the tth moment; a is the pre-emphasis coefficient; y(t) is the acceleration value after pre-emphasis; w(t-s) is the Hamming window, whose center is at s; x(f,s) is the two-dimensional time-frequency spectrum matrix obtained after short-time Fourier transform; f is the number of matrix rows, indicating frequency; s is the frame obtained by transformation in each window; the time-frequency matrix x(f,s) can be passed through the Mel-scale filter bank to obtain the Mel-frequency time-frequency spectrum x(fmel,s), which is specifically implemented as shown in the following equation (18):

[0056]

[0057] S5: A 2D convolutional neural network is used to extract the characteristic information contained in the acceleration spectrum. According to the time-domain continuity characteristics of the acceleration spectrum, a vector convolution kernel of size [N, 1] is used to slide linearly on the spectrum. At the same time, a pooling kernel of [N, 1] is used to downsample the frequency domain dimension to 1, and finally sort them in chronological order.

[0058] S6: Self-attention mechanism convolutional neural network, which uses the self-attention mechanism to process the spectrogram features of the last pooling layer of each channel. The specific implementation method is shown in equations (19)-(21):

[0059] Score = query × key (19)

[0060]

[0061] output=S×value (21)

[0062] Among them: query, key, value are the feature matrices obtained after 1×1 convolution of the current spectrogram features, which are consistent with the shape of the original spectrogram features; Score is the result of dot multiplication of the query and key feature matrices; n is the number of numerical points in the feature matrix; S i is the attention weight value calculated based on the current Score element; S is S i The attention weight matrix is ​​obtained according to the position combination in the original feature matrix; the output is the result obtained by multiplying the spectrogram features and the attention weights;

[0063] S7: After the above S1-S6 operations, the enhanced features are used as classifier input, and then the road flatness is identified by the classifier.

[0064] Beneficial effects of the present invention: The multi-strategy road roughness detection method provided by the present invention combines the traditional road detection principle with the environmental conditions of temporal and spatial consistency, carries out road condition information analysis and extraction based on cloud storage technology, makes full use of low-cost multivariate methods and technologies, and provides an effective method for road roughness detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a road priority analysis report.

[0066] Figure 2 It is the experimental result of the cross-validation score visualization. DETAILED DESCRIPTION

[0067] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0068] This embodiment discloses a multi-strategy road roughness detection method, comprising the following steps:

[0069] 1. Posture self-correction Posture correction

[0070] ①Basic principles of smartphone posture correction

[0071] In the vehicle coordinate system, the longitudinal vibration of the vehicle is defined as the Y direction, with the driving direction as positive, the lateral vibration of the vehicle is defined as the X direction, with the right side of the driving direction as positive, and the vertical vibration of the vehicle is defined as the Z direction, with the upward direction as positive; in the mobile phone coordinate system, the vehicle-mounted smartphone is set vertically.

[0072] Ideally, the screen of the smartphone in the vehicle is perpendicular to the direction of the vehicle. In non-ideal situations, the y positive direction is defined as the direction from the bottom to the top of the smartphone, the x positive direction is defined as the direction from left to right, and the z positive direction is defined as the direction perpendicular to the screen. When the vehicle is driving, the vibration of the vehicle is transmitted to the phone and recorded by the built-in sensor. It is very difficult or even impractical to completely keep the smartphone coordinate system orthogonal to the vehicle coordinate system, so correction is required to ensure the convenience of smartphone detection.

[0073] ②Mobile phone vertical posture correction

[0074] The measured output of the in-vehicle smartphone is acceleration, expressed as a vector (a x ,a y ,a z ,), the acceleration value in the vehicle coordinate system is expressed as (a X ,a Y ,a Z ,). Correction of vertical acceleration: the mobile phone coordinate system is the xy plane, the vehicle coordinate system is the XY plane, and the vehicle-mounted smartphone has an angle deviation relative to the vehicle-mounted coordinate system. When the vehicle is parked, the acceleration vector v of the three directions of the vehicle-mounted smartphone is expressed as (x g ,yg ,z g ), the resultant vector is the gravitational acceleration, recorded as:

[0075]

[0076] Where: (x g ,y g ,z g ) T They are the three components of gravity acceleration in the x, y, and z coordinates of the smartphone;

[0077] ③ The vertical unit vector Z of the vehicle coordinate system is (0, 0, 1). The mobile phone acceleration vector V is rotated to the vertical unit vector Z. The transformation matrix R is determined so that the relationship shown in formula (2) is satisfied:

[0078] Z T =R·V (2)

[0079] According to the Rodriguez rotation formula,

[0080] R(k,φ)=I·cosφ+Ksinφ+(1―cosφ)u·u T (3)

[0081] Among them: after rotation, u is the cross product of the vertical unit vector Z and the mobile phone acceleration vector v, and is normalized:

[0082] K is the cross product matrix of the rotation axis, then we have

[0083]

[0084] The rotation angle φ between the unit vector Z and the mobile phone acceleration vector v is:

[0085] φ=arccos(VZ / |V||Z|) (6)

[0086] 2. Multi-scale fuzzy entropy road roughness information representation

[0087] According to the measurement characteristics and error characteristics of intelligent sensors, theoretical and technical research is carried out on sensor measurement characteristics, measurement quality and accuracy degradation mechanism, and road roughness information is represented.

[0088] The multi-scale fuzzy entropy algorithm first processes the original time series and then calculates the fuzzy entropy value of the series. The calculation process is as follows.

[0089] S1: original acceleration sequence, original acceleration sequence U with N points after rotation correction i = {u1,u,…,u N} According to the scale factor τ = 1, 2, ..., N, we get the new vector:

[0090]

[0091] When τ is 1, y j (1) is the original sequence U i τ converts the original sequence U i Split into τ sequences y of length N / τ each i (τ).

[0092] S2: For the sequence Y(i)={y(1),y(2)…,y(τ)} after rotation correction and segmentation, first determine the vector dimension m, and then reconstruct the time series Y to obtain a new sequence:

[0093]

[0094] Subtract the mean y0(i) from the m consecutive elements starting from the i-th point, where m is the embedding dimension, and we get:

[0095]

[0096] S3: Define two reconstruction vectors and The maximum distance between

[0097]

[0098] S4: The fuzzy membership function is:

[0099]

[0100] Among them, n is the boundary gradient, r is the similarity tolerance, which is generally 0.1 to 0.25SD (SD is the standard deviation of the original data.

[0101] S5: Define the metric function to evaluate the relevance:

[0102]

[0103] S6: Calculate the mean of the metrics:

[0104]

[0105] S7: After the vector dimension is m+1, repeat Step1-Step5 to get δ m+1 (n,r):

[0106]

[0107] S8: Define fuzzy entropy as:

[0108]

[0109] 3. Use the self-attention mechanism convolutional neural network to identify the road roughness based on the acceleration data collected by the onboard smartphone during vehicle driving. The specific steps are as follows:

[0110] S1: Visualization of acceleration information features. The acceleration information is transformed into a time-frequency spectrum using a short-time Fourier transform, and then the time-frequency spectrum is transformed into a Mel-scaled time-frequency spectrum through a Mel-scaled filter bank. To enhance data validity, the spectrum is pre-emphasized.

[0111] S2: Spectral feature extraction, using the feature extraction capability of convolutional neural network, setting the convolution kernel according to the characteristic characteristics, and performing linear sliding operation on the time-frequency spectrum to obtain the features;

[0112] S3: Feature enhancement, using the self-attention mechanism to enhance the spectrogram features and improve the effectiveness of acceleration information features;

[0113] S4: Use equations (16)-(17) to express the established road roughness information:

[0114] y(t)=x(t)+a×x(t-1)(16)

[0115]

[0116] Where: x(t) is the acceleration value after rotational multi-scale processing at the tth moment; a is the pre-emphasis coefficient; y(t) is the acceleration value after pre-emphasis; w(t-s) is the Hamming window, whose center is at s; x(f,s) is the two-dimensional time-frequency spectrum matrix obtained after short-time Fourier transform; f is the number of matrix rows, indicating frequency; s is the frame obtained by transformation in each window; the time-frequency matrix x(f,s) is passed through the Mel-scale filter bank to obtain the Mel-frequency time-frequency spectrum x(fmel,s), which is specifically implemented as shown in the following equation (18):

[0117]

[0118] S5: Use a 2D convolutional neural network to extract the characteristic information hidden in the acceleration spectrum. According to the temporal continuity of the acceleration spectrum, a vector convolution kernel of size [N, 1] is used to slide linearly on the spectrum, and a pooling kernel of size [N, 1] is used to downsample the frequency domain dimension to 1, and finally sort them in chronological order. Among them: the input size is 80×600, the number of filters in the convolution layer is 8, the number of filters in the pooling layer is not specified, and the filter size is [2,1]. After the above operation, the spectrum feature is obtained with a size of 1×600.

[0119] S6: Convolutional Neural Networks

[0120] The self-attention mechanism can enhance features. The self-attention mechanism is used to process the spectrogram features of the last pooling layer of each channel. The specific implementation method is shown in equations (19)-(21):

[0121] Score = query × key (19)

[0122]

[0123] output=S×value (21)

[0124] Among them: query, key, value are the feature matrices obtained after 1×1 convolution of the current spectrogram features, which are consistent with the shape of the original spectrogram features; Score is the result of dot multiplication of the query and key feature matrices; n is the number of numerical points in the feature matrix; S i is the attention weight value calculated based on the current Score element; S is S i The attention weight matrix is ​​obtained according to the position combination in the original feature matrix; the output is the result obtained by multiplying the spectrogram features and the attention weights.

[0125] After the above operations, the enhanced features are used as classifier input, and then the road roughness is identified by the classifier.

[0126] Experimental results:

[0127] In this embodiment, a total of 442 data were collected, covering a distance of 15.24 kilometers. The obtained data were uploaded to the Huawei cloud platform, and data analysis was performed using the posture self-correction posture correction algorithm, multi-scale fuzzy entropy road roughness information representation, and self-attention mechanism convolutional neural network road roughness recognition algorithm.

[0128] The data is divided into training set and test set to obtain the best road roughness model and cross-validation score, and the cross-validation score is visualized. The experimental results are as follows Figure 1 , Figure 2 shown.

[0129] A certain area is randomly selected to collect and verify the data of different levels of road surface roughness. When the vehicle is driving, a smart terminal is used to collect road surface roughness information to evaluate the road surface quality.

[0130] In this area, the road surface roughness of a designated lane of a highway is detected by a road surface roughness fast tester, and the IRI value of each kilometer pile number on a specific lane is calculated by the road condition information element, as shown in Table 1.

[0131] Table 1 Experimental equipment

[0132]

[0133] A verification experiment was conducted on a highway. During the verification experiment, the driving speed was stabilized at about 80km / h. In order to better verify the accuracy of the experiment, three smart terminals of the vehicle-mounted smartphone were used to simultaneously collect road condition information.

[0134] For convenience, the experiment uses the pile number signs along the road as the representation. The IRI calculation results of the pile number per kilometer in the right test section of a certain highway are shown in Table 2.

[0135] The road surface roughness test index is the International Roughness Index (IRI, unit: m / km), and then according to the formula RQI = 100 / (1+0.026e 0.65IRI ) is converted to RQI.

[0136] Based on a highway experiment, the vehicle speed was 80km / h and three intelligent terminals collected road condition information at the same time. The measured results showed that a multi-strategy road roughness detection method can effectively and quickly identify road roughness.

[0137] Table 2 IRI of a certain highway right test section

[0138]

[0139]

Claims

1. A multi-strategy road roughness detection method, characterized in that: The following steps are involved: (1) Using a posture self-correction algorithm, the acceleration data collected and measured by the vehicle-mounted smartphone acceleration sensor is updated and corrected in real time to adapt to the dynamic changes of the vehicle during driving; (2) Using a multi-scale fuzzy entropy algorithm, the measurement data reflecting the road roughness collected by the on-board smartphone during vehicle driving is processed at multiple scales to construct a road roughness representation model and combine global features with local features. (3) A convolutional neural network with a self-attention mechanism is used to identify the road roughness based on the acceleration data collected by the onboard smartphone during vehicle driving.

2. A multi-strategy road roughness detection method according to claim 1, characterized in that: The specific implementation process of step (1) is as follows: S1. Establish the vehicle coordinate system and the mobile phone coordinate system: In the vehicle coordinate system, the longitudinal vibration of the vehicle is defined as the Y direction, with the driving direction as positive, the lateral vibration of the vehicle is defined as the X direction, with the right side of the driving direction as positive, and the vertical vibration of the vehicle is defined as the Z direction, with the upward direction as positive; in the mobile phone coordinate system, the vehicle-mounted smartphone is set vertically. Ideally, the screen of the vehicle-mounted smartphone is perpendicular to the direction of vehicle travel. In non-ideal situations, the direction from the bottom to the top of the vehicle-mounted smartphone is defined as the positive y direction, from left to right as the positive x direction, and perpendicular to the screen outward as the positive z direction; S2. Vertical posture correction of the in-vehicle smartphone: (1) Measured vehicle-mounted smartphone acceleration data, vector representation is (a x ,a y ,a z, ), then the acceleration value in the vehicle coordinate system is expressed as (a X ,a Y ,a Z, ); (2) Vertical acceleration correction: There is an angle deviation between the vehicle-mounted smartphone and the vehicle body coordinate system. When the vehicle is parked, the acceleration vector V of the vehicle-mounted smartphone in three directions is expressed as (x g ,y g ,z g ), the resultant vector is the gravitational acceleration, recorded as: Where: (x g ,y g ,z g ) T They are the three components of gravity acceleration in the x, y, and z coordinates of the smartphone; (3) The vertical unit vector Z of the vehicle coordinate system is (0, 0, 1). The mobile phone acceleration vector V is rotated to the vertical unit vector Z. The transformation matrix R is determined so that the relationship shown in formula (2) is satisfied: With T =R·V (2) According to the Rodriguez rotation formula, R(k,φ)=I·cosφ+Ksinφ+(1―cosφ)u·u T (3) Among them: after rotation, u is the cross product of the vertical unit vector Z and the mobile phone acceleration vector V, and is normalized: K is the cross product matrix of the rotation axis, then: The rotation angle φ between the unit vector Z and the mobile phone acceleration vector v is: φ=arccos(VZ / ||V||Z|) (6) 3. A multi-strategy road roughness detection method according to claim 1, characterized in that: The multi-scale fuzzy entropy algorithm in step (2) first processes the corrected original acceleration according to the time series, and then calculates the fuzzy entropy value of the sequence. The calculation process is as follows: S1: original acceleration sequence, original acceleration sequence U with N points after rotation correction i = {u1,u,…,u N } According to the scale factor τ = 1, 2, ..., N, we get the new vector: When τ is 1, y j (1) is the original sequence U i , τ converts the original sequence U i Split into τ sequences y of length N / τ each i (τ); S2: For the sequence Y(i)={y(1),y(2)…,y(τ)} after rotation correction and segmentation, first determine the vector dimension m, and then reconstruct the time series Y to obtain a new sequence: Subtract the mean y0(i) from the m consecutive elements starting from the i-th point, where m is the embedding dimension, and we get: S3: Define two reconstruction vectors and The maximum distance between S4: The fuzzy membership function is: Among them, n is its boundary gradient, r is the similarity tolerance, which is generally 0.1 to 0.25SD (SD is the standard deviation of the original data; S5: Define the metric function to evaluate the relevance: S6: Calculate the mean of the metrics: S7: After the vector dimension is m+1, repeat Step1-Step5 to get δ m+1 (n,r): S8: Define fuzzy entropy as: FE(m,n,r,τ)=lnφ m (n,r)―lnφ m+1 (n,r)(15) 4. A multi-strategy road roughness detection method according to claim 1, characterized in that: The specific steps of step (3) are as follows: S1: Visualization of acceleration information features. The acceleration information is transformed into a time-frequency spectrum using a short-time Fourier transform, and then the time-frequency spectrum is transformed into a Mel-scaled time-frequency spectrum through a Mel-scaled filter bank. To enhance data validity, the spectrum is pre-emphasized. S2: Spectral feature extraction, using the feature extraction capability of convolutional neural network, setting the convolution kernel according to the characteristic characteristics, and performing linear sliding operation on the time-frequency spectrum to obtain the features; S3: Feature enhancement, using the self-attention mechanism to enhance the spectrogram features and improve the effectiveness of acceleration information features; S4: Use equations (16)-(17) to express the established road roughness information: y(t)=x(t)+a×x(t-1)(16) Where: x(t) is the acceleration value after rotation multi-scale processing at the tth moment; a is the pre-emphasis coefficient; y(t) is the acceleration value after pre-emphasis; w(t-s) is the Hamming window, whose center is at s; x(f,s) is the two-dimensional time-frequency spectrum matrix obtained after short-time Fourier transform; f is the number of matrix rows, indicating frequency; s is the frame obtained by transformation in each window; the time-frequency matrix x(f,s) can be passed through the Mel-scale filter bank to obtain the Mel-frequency time-frequency spectrum x(fmel,s), which is specifically implemented as shown in the following equation (18): S5: A 2D convolutional neural network is used to extract the characteristic information contained in the acceleration spectrum. According to the time-domain continuity of the acceleration spectrum, a vector convolution kernel of size [N, 1] is used to slide linearly on the spectrum. At the same time, a pooling kernel of size [N, 1] is used to downsample the frequency domain dimension to 1, and finally sort them in chronological order. S6: Self-attention mechanism convolutional neural network, which uses the self-attention mechanism to process the spectrogram features of the last pooling layer of each channel. The specific implementation method is shown in equations (19)-(21): Score = query × key (19) output=S×value (21) Among them: query, key, value are the feature matrices obtained after 1×1 convolution of the current spectrogram features, which are consistent with the shape of the original spectrogram features; Score is the result of dot multiplication of the query and key feature matrices; n is the number of numerical points in the feature matrix; S i is the attention weight value calculated based on the current Score element; S is S i The attention weight matrix is ​​obtained according to the position combination in the original feature matrix; the output is the result obtained by multiplying the spectrogram features and the attention weights; S7: After the above S1-S6 operations, the enhanced features are used as classifier input, and then the road flatness is identified through the classifier.

Images