Method and system for classifying and calculating depth of pavement rut based on deep learning

Abstract

The application provides a kind of based on deep learning's road surface rut classification and rut depth calculation method and system, it is related to road surface disease detection technical field.The method is first for the characteristics of high-dimensional rut cross section elevation data, carry out smoothing, inclination correction, and dimensionality reduction such as pretreatment, and according to the different classification of rut cross section two kinds of data sets are formed, the establishment of rut type and shape correlation;Then construct a variety of structure neural network model for rut shape classification, and select the best classification model by ablation experiment;Finally, based on the classification results of different categories of rut, different virtual ruler is established, and the rut depth is calculated.The method selects CNN-BiLSTM model as the best classification model, and different virtual straight is established for different types of rut to calculate the rut depth, which improves the efficiency of rut depth calculation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pavement disease detection, and in particular to a pavement rut classification and rut depth calculation method and system based on deep learning. BACKGROUND

[0002] As one of the common and serious diseases of asphalt pavement, rutting has become a key problem in road safety and maintenance management. Rutting detection, as an important part of road quality evaluation, has received extensive attention in recent years. Rutting detection technology can be divided into two categories: traditional detection technology and modern detection technology. Traditional detection technology mainly relies on manual or simple instruments to measure rut depth. Modern detection technology uses high-precision sensors, laser scanning, image processing and other technical means to achieve rapid and automated rut detection.

[0003] Manual measurement is the earliest method, which usually measures the rut depth directly by using tools such as steel ruler and tape measure. Although this method is simple and cost-effective, it has the disadvantages of low measurement accuracy, low efficiency and high manual intervention, which cannot meet the needs of large-scale road detection. With the development of technology, non-contact detection of rutting has been widely used. This method realizes zero interference operation of traffic flow, thereby improving detection efficiency, and is particularly suitable for sensitive scenes such as highways. Common non-contact detection methods include ultrasonic detection, laser detection and image detection. Ultrasonic detection technology has some limitations, such as interference of external noise with the sensor and high requirements for the position of the sensor during detection, which will affect the detection accuracy. Laser detection has certain errors due to the number of laser points. The increase in the number of laser points is not unlimited, so the detection device can only obtain the cross-section elevation data of the rut with a large interval, which limits the comprehensive and accurate acquisition of the rut shape.

[0004] The commonly used method for calculating rut depth assumes that the pavement cross-section has only two rut grooves, and their respective highest and lowest points fall exactly in the specified area. According to the actual situation of rut occurrence, the rut shape is complex and diverse, and the "one-size-fits-all" method has certain limitations.

[0005] Currently, three-dimensional line laser scanning technology is widely used in pavement detection, and a large amount of pavement cross-section data is obtained. Deep learning method can effectively solve the problem of recognizing unknown complex patterns by using multi-layer neural network combined with massive data training to extract features. Three-dimensional line laser scanning technology, as a non-contact measurement method, can efficiently obtain the elevation data of the road surface and provide rich information data for rut detection. With the popularization of intelligentization and automation and the continuous improvement of transportation network, these data should be reasonably utilized to provide high-quality data sets for intelligent transportation. SUMMARY

[0006] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a method and system for road rut classification and rut depth calculation based on deep learning. The method first classifies the rut morphology and then calculates the rut depth based on the classification results, thereby effectively improving the accuracy of rut depth detection.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] On the one hand, this invention provides a method for road rut classification and rut depth calculation based on deep learning, including:

[0009] Elevation data of rut cross sections were acquired and preprocessed. Two datasets were established based on different classification methods of rut cross sections, and the correlation between rut type and shape was established.

[0010] We constructed neural network models with various structures for rut shape classification and selected the best classification model through ablation experiments.

[0011] The rut depth is calculated based on the rut classification results; different virtual rulers are established for different categories of rut classification results to calculate the rut depth.

[0012] Furthermore, the acquisition and preprocessing of the cross-sectional elevation data of the ruts includes:

[0013] A three-dimensional line laser detection system was used to obtain cross-sectional elevation data of vehicle ruts;

[0014] The acquired rut cross-sectional elevation data were processed by smoothing and noise reduction, cross-sectional tilt correction, and dimensionality reduction.

[0015] Furthermore, the specific method for tilt correction of the rut cross section is as follows:

[0016] The elevation difference method or angle rotation method is used to eliminate non-realistic deformation data caused by equipment posture deviation during the measurement process, and to ensure that the horizontal projection baseline formed by the boundary points on both sides of the cross section is completely coincident with the X-axis.

[0017] Furthermore, the specific method for establishing two datasets based on different classification methods of rut cross-sections is as follows:

[0018] Two datasets, A and B, were constructed based on different classification methods for rut cross-sections. Dataset A categorizes rut ​​shapes into three types based on the distribution of the central peak point in the rut cross-section: U-shaped, concave W-shaped, and convex W-shaped. The U-shaped rut cross-section is characterized by the absence of a central peak point, exhibiting a smooth concave curve with rut depth concentrated in the middle. The concave W-shaped rut cross-section is characterized by two valley points and a central peak point, exhibiting two grooves with the central convex point lower than the road surface. The convex W-shaped rut cross-section is characterized by two grooves and a central peak point, with the central peak point higher than the sides, exhibiting a raised central section and relatively lower sides.

[0019] Dataset B is classified according to the number of peak points at both ends of the rut cross section, specifically into three types: Type 0, Type 1, and Type 2. Type 0 rut cross sections are characterized by no peak points on both sides, with a smooth shape or depth concentrated in the middle; Type 1 rut cross sections are characterized by only one peak point at each end, exhibiting asymmetrical geometric characteristics; Type 2 rut cross sections are characterized by one peak point at each end, exhibiting symmetrical double-convex characteristics.

[0020] Furthermore, the construction of neural network models with various structures for rut shape classification specifically includes CNN models, BiLSTM models, CNN-LSTM models, and CNN-BiLSTM models; the CNN-BiLSTM model was selected as the best classification model through ablation experiments.

[0021] Furthermore, the method selects the cross-entropy loss function as the loss function for the rut shape classification task and uses the AdamW optimizer as the model parameter update mechanism.

[0022] Furthermore, the method introduces a self-attention mechanism and a learning rate scheduler to optimize the CNN-BiLSTM model;

[0023] A self-attention mechanism is introduced to construct the CNN-BiLSTM-Att model based on the CNN-BiLSTM model. The self-attention mechanism calculates the attention of a sampling point in the rut cross-section data to itself and other sampling points, analyzes the correlation between sampling points in the rut cross-section elevation data sequence, and adjusts the representation of each sampling point to further improve the model performance. The learning rate scheduler is used to dynamically adjust the learning rate of the CNN-BiLSTM model, so that the model can gradually optimize its convergence effect and training speed during the training process.

[0024] Furthermore, the method establishes different virtual rulers for different categories of rut classification results and calculates the rut depth, specifically as follows:

[0025] When the classification results obtained by the CNN-BiLSTM-ATt model are based on the three rut types—U-shaped, concave W-shaped, and convex W-shaped—in dataset A, a local extremum index function is first designed to determine whether there are peak points within 20 data points on each side of the rut curve, and the coordinates of these peak points are saved. Simultaneously, based on the rut type, the coordinates of valley points are further searched and saved. Different virtual rulers are established based on the peak and valley points of different rut types to calculate the rut depth. Specifically, the following three scenarios apply:

[0026] (1) Convex W-shaped ruts: When calculating the rut depth, first determine the coordinates of the middle peak point, and then select the virtual ruler construction method according to whether there are peak points on both sides; if there are peak points on both sides, the virtual ruler is connected to the middle peak point by the peak points on both sides respectively, or by one peak point and the other end point connected to the middle peak point respectively; if there are no peak points on both sides, the virtual ruler is constructed by the middle peak point and the two end points; after the virtual ruler is constructed, the height difference between the valley point on the left and right sides and the virtual ruler is calculated respectively, and then the depth of the ruts on the left and right sides is obtained.

[0027] (2) Concave W-shaped ruts: A virtual ruler is established based on whether there are peak points on both sides of the valley point. If there are peak points, the virtual ruler is connected by two peak points, or by one peak point and the other end point. If there are no peak points, the virtual ruler is established through the two end points. After the virtual ruler is constructed, the height difference between the two valley points on the left and right sides and the virtual ruler is calculated to obtain the rut depth on the left and right sides respectively.

[0028] (3) U-shaped ruts: First, establish a virtual ruler through the two ends, and then calculate the height difference between the virtual ruler and the valley point to determine the rut depth;

[0029] When the classification result obtained by the CNN-BiLSTM-ATt model is based on three rut types (Type 0, Type 1, and Type 2) in dataset B, a local extremum index function is first designed to determine whether there are peak points in the rut curve between the 50th and 80th data points, and the coordinates of these peak points are saved. Furthermore, the coordinates of valley points are found and saved according to the rut type. Different virtual rulers are established based on the peak and valley points of different rut types to calculate the rut depth. Specifically, the following three scenarios apply:

[0030] (1) Type 0 ruts: First, determine whether there is a peak point in the middle of the cross-sectional curve. If not, the virtual ruler is directly connected from both ends. If there is, compare the value of the middle peak point with the value of the left end point. When the value of the middle peak point is greater than the value of the left end point, the virtual ruler is directly connected from both ends. When the value of the middle peak point is less than the value of the left end point, the virtual ruler is constructed by combining the middle peak point and the two ends. Then, the height difference between the valley point and the virtual ruler is calculated through the virtual ruler to obtain the depth of the rut.

[0031] (2) Type 1 ruts: First, determine which side of the rut cross-section curve has a peak point among 20 data points on each side, and save the coordinates of the peak point; then compare the value of the middle peak point with the value of the left endpoint. When the value of the middle peak point is greater than the value of the left endpoint, the virtual ruler is connected to the other endpoint by one peak point; when the value of the middle peak point is less than the value of the left endpoint, the virtual ruler is constructed by connecting one peak point and the other endpoint to the middle peak point respectively; then, the height difference between the valley points on the left and right sides and the ruler is calculated by the two virtual rulers respectively, so as to obtain the depth of the ruts on the left and right sides.

[0032] (3) Type 2 ruts: First, determine which side of the rut cross section curve has a peak point among 20 data points on each side, and save the coordinates of the peak point; then compare the value of the middle peak point with the value of the left endpoint. When the value of the middle peak point is greater than the value of the left endpoint, the virtual ruler is connected by the peak points on both sides; when the value of the middle peak point is less than the value of the left endpoint, the virtual ruler is constructed by connecting the peak points on both sides to the middle peak point respectively; then, the height difference between the valley points on the left and right sides and the ruler is calculated by the two virtual rulers respectively, so as to obtain the depth of the ruts on the left and right sides.

[0033] On the other hand, the present invention provides a road rut classification and rut depth calculation system based on deep learning, including a data acquisition module, a data processing module, a rut shape classification module, a display module, a rut depth calculation module and a data storage module;

[0034] The data acquisition module is used to acquire the cross-sectional elevation data of the ruts; a three-dimensional line laser detection system is used to acquire the cross-sectional elevation data of the ruts, which is used to realize the rut shape recognition and classification and the rut depth;

[0035] The data processing module includes a three-level preprocessing architecture. First, noise reduction is achieved based on B-spline curve fitting. Second, a tilt correction model is constructed, and the height difference method and angle rotation method are used to eliminate the systematic deviation caused by road camber. Finally, the interval sampling method is used to reduce the dimensionality of the data.

[0036] The rut shape classification module classifies ruts according to different classification methods;

[0037] The display module visualizes the cross-section of ruts, including untreated rut cross-sections and cross-sections at any pre-processing stage.

[0038] The rut depth calculation module calculates rut ​​depth by establishing different virtual rulers based on the rut shape classification results.

[0039] The data storage module is used to save the rut shape classification and depth calculation results.

[0040] Thirdly, the present invention proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned deep learning-based method for road rut classification and rut depth calculation.

[0041] The beneficial effects of adopting the above technical solution are as follows: The deep learning-based method and system for road rut classification and rut depth calculation provided by this invention uses rut ​​cross-sectional elevation data obtained by three-dimensional line laser scanning as a foundation, and combines deep learning to identify and classify rut shapes and calculate rut depth. Multiple data preprocessing methods are employed to preserve data features and improve computational efficiency. Two datasets, A and B, are established based on different classification methods to verify the model's adaptability and provide data support for rut shape classification research. The CNN-BiLSTM model is determined to be the optimal classification model through ablation experiments, and the optimal activation function and standardization are selected for the model. To further improve the model's classification accuracy, a self-attention mechanism is introduced into the CNN-BiLSTM model for optimization. Different virtual straight lines are established for different types of ruts to calculate rut depth, thereby improving the efficiency of rut depth calculation. Attached Figure Description

[0042] Figure 1 The flowchart shows the deep learning-based method for classifying road ruts and calculating rut depth provided in Embodiment 1 of the present invention.

[0043] Figure 2 The image shows the cross-sectional curve data of the ruts before and after the abrupt change point provided in Embodiment 1 of the present invention, wherein (a) is the cross-sectional rut before the abrupt change point and (b) is the cross-sectional rut after the abrupt change point.

[0044] Figure 3 The following is a comparison chart of the denoising effects of six different smoothing methods provided in Embodiment 1 of the present invention. Among them, (a) is a comparison chart of the denoising effects of three smoothing methods: moving average method, weighted moving average method and exponential smoothing method, and (b) is a comparison chart of the denoising effects of three smoothing methods: LOESS smoothing method, B-spline smoothing method and wavelet transform.

[0045] Figure 4 This is a schematic diagram of the tilt of the rut cross section caused by the lateral slope of the road provided in Embodiment 1 of the present invention;

[0046] Figure 5 Schematic diagrams of cross-sections of ruts of different shapes provided in Embodiment 1 of the present invention;

[0047] Figure 6 This is a diagram of the CNN-BiLSTM model structure provided in Embodiment 1 of the present invention;

[0048] Figure 7 The confusion matrix of the CNN-BiLSTM model provided in Embodiment 1 of the present invention for classification on the test set;

[0049] Figure 8 This is a schematic diagram of rut depth calculation based on dataset A provided in Embodiment 1 of the present invention, wherein (a) is the calculation of convex W-shaped rut depth, (b) is the calculation of concave W-shaped rut depth, and (c) is the calculation of U-shaped rut depth.

[0050] Figure 9 This is a schematic diagram of rut depth calculation based on dataset B provided in Embodiment 1 of the present invention, wherein (a) is the rut depth calculation of type 0, (b) is the rut depth calculation of type 1, and (c) is the rut depth calculation of type 2.

[0051] Figure 10 This is a block diagram of the road rut classification and rut depth calculation system based on deep learning provided in Embodiment 2 of the present invention. Detailed Implementation

[0052] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0053] Example 1:

[0054] This embodiment provides a method for road rut classification and rut depth calculation based on deep learning, such as... Figure 1 As shown, it includes the following steps:

[0055] Step 1: Obtain the cross-sectional elevation data of the ruts and perform preprocessing;

[0056] A three-dimensional line laser detection system was used to acquire cross-sectional elevation data of ruts. The collected cross-sectional elevation data was analyzed, and noise reduction, cross-sectional tilt correction, and dimensionality reduction were applied based on its characteristics. Two datasets were established according to different classification methods to establish the correlation between rut type and shape, providing a foundation for a deep learning-based rut classification model.

[0057] Step 1.1: The elevation data of the rut cross section acquired by the three-dimensional line laser detection system is processed to remove noise;

[0058] The elevation data of rut cross sections acquired using a three-dimensional line laser scanning system can predict the trend of the next point based on the changing relationship, belonging to one-dimensional sequential data. The line laser scanning system used in this study has a uniform intensity distribution of the scanning lines, resulting in elevation data with high spatial continuity. However, due to the influence of factors such as natural light, road surface material, and interference spots of the excitation source, rut cross section elevation data usually contains high-frequency noise, requiring smoothing processing. Commonly used smoothing methods for ordered elevation data include moving average, weighted moving average, exponential smoothing, spline smoothing, LOESS smoothing, and wavelet transform.

[0059] This embodiment uses Figure 2 Using the rutted pavement cross-sectional curve data as an example, the denoising effects of different smoothing methods are compared. The denoising effect will be evaluated based on the high-frequency oscillation noise and system background noise coexisting in the rutted pavement elevation data. Figure 3 The denoising effects of different smoothing methods are presented.

[0060] pass Figure 3 The comparison of denoising effects shows that while exponential smoothing effectively reduces data noise, it fails to accurately capture data trends when processing details. Moving average, weighted moving average, and wavelet transform methods effectively filter noise, but they still retain some minor variations in texture during the denoising process, resulting in overly coarse details. In contrast, LOESS smoothing and B-spline smoothing have similar denoising effects. B-spline smoothing effectively removes noise while preserving the detailed features of the data, making it more consistent with the characteristics of rut cross-sections. Therefore, after comprehensive comparative analysis, this embodiment selects B-spline smoothing as the denoising method for rut cross-section elevation data.

[0061] Step 1.2: Correct the tilt of the cross-section of the rut;

[0062] During the process of collecting cross-sectional elevation data of road ruts using a three-dimensional line laser detection system, the reconstructed rut shape often has a certain angle with the road reference plane due to the influence of the road's transverse slope and scanning vibration. Figure 4 This diagram illustrates how the lateral slope of a road causes a rut shape to have an angle with the reference plane. When the X-axis is used as the zero-elevation reference plane for the road, the endpoints of the cross-section should meet the collinearity requirement. Tilt correction involves adjusting the spatial position of the rut cross-section using coordinate transformation algorithms. Specific operations include eliminating inaccurate deformation data caused by equipment posture deviations during measurement and ensuring that the horizontal projection reference line formed by the boundary points on both sides of the cross-section completely coincides with the X-axis. Commonly used rut tilt correction methods include the elevation difference method and the angle rotation method.

[0063] The "Specifications for Field Testing of Highway Subgrade and Pavement" (JTG 3450—2019) specifies seven common rut shapes, such as... Figure 5 As shown. These rut shapes are classified according to the geometric characteristics of the cross-sectional curve, and are mainly extracted and analyzed through the following key features: (1) Peak points: Peak points are local maximum points in the cross-section of the rut. Their quantity and distribution characteristics reflect the geometric undulation of the rut and are an important basis for distinguishing different shapes; (2) Valley points: Valley points mainly represent the relatively low points in the cross-sectional curve. Their distribution is closely related to calculating the rut depth and identifying the rut shape. In particular, the rut depth characteristics can be measured by the height difference between the valley point and the virtual ruler.

[0064] In this embodiment, based on the geometric characteristics of the rut cross-sections, two rut shape datasets, A and B, were constructed using different classification methods. Dataset A classifies rut ​​shapes into three categories based on the distribution of the intermediate peak point: U-shaped, concave W-shaped, and convex W-shaped, as shown in Table 1. The U-shaped rut cross-section is characterized by the absence of a peak point in the middle, exhibiting a smooth concave curve, with the rut depth typically concentrated in the middle. The concave W-shaped rut cross-section is characterized by two valley points and a peak point in the middle, exhibiting two grooves with a slightly convex middle section that is lower than the road surface. The convex W-shaped rut cross-section is characterized by two grooves and a distinct intermediate peak point, with the peak point higher than the sides, exhibiting a raised middle section and relatively lower sides.

[0065] Table 1. Rut Shape Dataset A;

[0066] ;

[0067] Dataset B is classified according to the number of peak points at both ends of the cross-section, specifically into three types: Type 0, Type 1, and Type 2, as shown in Table 2. Type 0 rut cross-sections are characterized by the absence of obvious peak points on both sides, typically exhibiting a smooth shape or depth concentrated in the middle; Type 1 rut cross-sections are characterized by only one obvious peak point at each end, displaying asymmetrical geometric characteristics; Type 2 rut cross-sections are characterized by one obvious peak point at each end, typically exhibiting symmetrical double-convex characteristics.

[0068] Table 2. Rut Shape Dataset B;

[0069] ;

[0070] The two datasets describe the geometric characteristics of rut cross-sections from different perspectives. The shape features of rut cross-sections are also influenced by different types of damage, and their correlations can be summarized as follows: convex W-shaped ruts (convex in the middle, concave at the wheel track areas on both sides) are mainly affected by unstable ruts; concave W-shaped ruts (lower convex in the middle, concave at the wheel track areas on both sides) are strongly correlated with compacted ruts; while U-shaped ruts (uniform subsidence) are most significantly affected by structural ruts. The classification results of dataset B can also determine the degree of rut boundary damage, and combined with dataset A, can specifically identify the rut type, allowing for targeted maintenance solutions such as milling and repaving or structural reinforcement for subsequent pavement repairs. These two classification methods not only provide a systematic approach for rut shape feature extraction and identification classification tasks, but also provide two datasets for deep learning models. This dual-standard classification approach adapts to different research needs and can also verify the model's adaptability to datasets with different classification standards.

[0071] Step 1.3: Perform dimensionality reduction processing on the cross-sectional elevation data of the vehicle ruts;

[0072] In processing rut cross-section elevation data, the high dimensionality of the data often leads to excessive computational burden. This is especially true when inputting high-resolution data into neural network training, which significantly increases training time and resource consumption.

[0073] To effectively reduce computational burden and improve training efficiency, this embodiment employs interval sampling to reduce the dimensionality of rut cross-section elevation data. Interval sampling involves selecting a point at fixed intervals between the original data points, thereby reducing the data's dimensionality. This method reduces data resolution while preserving the overall trend and characteristics of the data, alleviating computational pressure on the subsequent neural network model. In this embodiment, each complete rut cross-section contains 4096 sampling points, which are reduced to 128 sampling points using interval sampling. While reducing data dimensionality, interval sampling preserves the geometric shape of the rut cross-section elevation data, ensuring that key features are not lost during dimensionality reduction. Simultaneously, during neural network training, the model can effectively extract important features from the simplified data. This method effectively controls computational resource consumption, accelerates training, and reduces the risk of overfitting.

[0074] Step 2: Construct a neural network model to classify road ruts;

[0075] We constructed neural network models with various structures for rut shape classification, selected the optimal classification model through ablation experiments, including the selection of activation functions and standardization, and verified the classification accuracy and stability of the model using cross-validation experiments.

[0076] In this embodiment, the CNN model consists of multiple layers, and the network structure is shown in Table 3. Here, Layer represents the name of each layer in the CNN model; Parameter design refers to the parameter settings used in the CNN model; Output Shape represents the output dimension; k_size represents the kernel size used; s represents the stride; Pool_size represents the pooling region size; and u represents the number of neurons. In this embodiment, the CNN model consists of two convolutional layers, a pooling layer, and a fully connected layer. The network design aims to fully extract spatial and local features from the input data. The feature vector size of the input sample is (128, 1). The first convolutional layer uses 32 kernels, each with a size of 3×1; the second convolutional layer uses 64 kernels, also with a kernel size of 3×1. Through these two convolutional layers, the CNN model can progressively extract low-level and high-level features from the data. After each convolutional layer, a 2×1 global max pooling operation is applied for downsampling, allowing the feature dimension and computational complexity to be reduced simultaneously. After two convolution-pooling cascades, the data is mapped to a fully connected layer with 64 neurons. This layer transforms the abstract feature vector into the final classification probability distribution through nonlinear transformation.

[0077] Table 3 CNN model structure;

[0078] ;

[0079] To improve network performance, this embodiment introduces several optimization strategies into the CNN model. First, by selecting an appropriate learning rate, the network training process is ensured to be smooth, avoiding instability or overfitting caused by an excessively large learning rate. After each convolutional layer, the ReLU activation function is used to enable the CNN model to learn more complex features. Furthermore, to improve the network's convergence speed, a normalization method is used, which ensures that the input of each layer remains within a relatively stable range, thus promoting smooth training. In the network's output stage, the Softmax activation function is used to output the classification result, converting the output into a probability distribution. The CNN model ultimately classifies based on the maximum probability value. This design enables the network to effectively distinguish between multiple categories and ensures more accurate final classification results.

[0080] In this embodiment, the constructed BiLSTM model comprises two mapping layers, two BiLSTM layers, one fully connected layer, and one output layer. Its network structure is shown in Table 4, with parameter definitions identical to those in Table 3. The input feature vector has dimensions (128, 1), passes through two mapping layers with 64 and 128 channels respectively, and then enters two BiLSTM layers with 64 channels each. The output data is fed into a fully connected layer with 32 neurons to map the extracted features to the final classification result.

[0081] Table 4. BiLSTM model structure;

[0082] ;

[0083] To improve the training performance and convergence speed of the BiLSTM network, this embodiment adds the ReLU activation function to the mapping layer and each layer of the BiLSTM. Furthermore, normalization is also introduced into the BiLSTM network to further enhance the model's training stability and accelerate convergence. The output layer uses the Softmax function to output the classification results and obtain the predicted labels.

[0084] Wheel rut cross-sectional elevation data not only contains local spatial features such as peaks and grooves, but also exhibits temporal variations that change with distance or time. The CNN-LSTM model combines the advantages of CNN and LSTM to provide an effective solution for the characteristics of wheel rut cross-sectional data. In this model, the CNN module extracts local spatial features, such as convexities and grooves, from the wheel rut cross-section through convolution and pooling operations, accurately capturing local changes in the data; the LSTM module addresses long-term dependencies in temporal data, identifying trends in wheel rut shape changes with distance or time, and performs particularly well in handling dependencies over long time spans. Therefore, combining the two enables the CNN-LSTM model to handle tasks with both spatial and temporal dependencies, comprehensively processing wheel rut cross-sectional elevation data, thereby improving classification accuracy and overcoming the shortcomings of traditional methods in handling complex wheel rut data.

[0085] The CNN-LSTM model constructed in this embodiment consists of two convolutional layers, one pooling layer, one LSTM layer, one fully connected layer, and one output layer, as shown in Table 5. The parameter definitions are the same as those in the model network structure in Table 3. The model input is a feature vector of size (128,1). After processing by convolution, pooling, and LSTM layers, the final classification result is obtained through the fully connected layer. The model uses 32 convolutional kernels in the first convolutional layer, each with a size of 3×1; the second convolutional operation uses 64 convolutional kernels, each with a size of 5×1, and both convolutions have a stride of 2. Through these two convolutional layers, the model can progressively extract low-level and high-level features from the data. Simultaneously, after the last convolutional layer, a 2×1 global max pooling operation is applied for downsampling, simultaneously reducing feature dimensionality and computational complexity. The pooled data is then passed to a 64-channel LSTM layer for temporal feature extraction. After the LSTM layer, the output data is fed into a fully connected layer with 32 neurons to map the extracted features to the final classification result.

[0086] Table 5. CNN-LSTM model structure;

[0087] ;

[0088] To improve the performance of the CNN-LSTM model, this embodiment employs the ReLU activation function in every convolutional and LSTM layer. This activation function actively suppresses gradient decay through a threshold truncation mechanism. To further improve the model's convergence speed and training stability, normalization is added to both the CNN and LSTM layers. This optimizes the input distribution of each layer, promoting efficient model training. Furthermore, the Softmax activation function used in the output layer converts the output of the fully connected layer into class probabilities, thereby enabling multi-class classification tasks.

[0089] The CNN-BiLSTM model introduces a bidirectional LSTM module, enabling it to consider both forward and backward temporal information simultaneously. Based on this characteristic, the model can not only learn past temporal information from the data but also capture future trends, thus better classifying rut shapes. In rut cross-sectional data, changes in local morphology are often influenced by both preceding and following points; the introduction of bidirectional information flow significantly improves the predictive ability for these changes. The bidirectional structure of BiLSTM provides a more comprehensive understanding of the temporal dependencies of the data, reducing the risk of the model missing crucial information. Compared to the CNN-LSTM model with only a unidirectional LSTM, the CNN-BiLSTM model more effectively captures details and patterns across time steps when processing datasets like rut data, which exhibit strong temporal and spatial variations.

[0090] The CNN-BiLSTM model constructed in this embodiment consists of two convolutional layers, one pooling layer, one BiLSTM layer, one fully connected layer, and one output layer, as shown in Table 6. The parameter definitions are the same as those in the model network structure in Table 3. The model input is a feature vector of size (128,1). After processing by convolutional, pooling, and BiLSTM layers, the final classification result is obtained through a fully connected layer. The model structure is as follows: Figure 6 As shown, the model uses 32 convolutional kernels in the first convolutional layer, each with a size of 3×1; the second convolutional operation uses 64 convolutional kernels, each with a size of 5×1, and both convolutions have a stride of 2. Through these two convolutional operations, the model can extract local features from the input data. Simultaneously, a 2×1 global max pooling operation is applied after each convolutional layer for downsampling, simultaneously reducing feature dimensionality and computational complexity. The pooled data is then passed to a 64-channel BiLSTM layer for temporal feature extraction. After the BiLSTM layer, the output data is fed into a fully connected layer with 32 neurons to map the extracted features to the final classification result.

[0091] Table 6. CNN-BiLSTM model structure;

[0092] ;

[0093] To optimize the CNN-BiLSTM model, this embodiment adds ReLU activation and normalization to both convolutional layers. These accelerate network training and promote efficient model training. The output layer uses the Softmax activation function to convert the output of the fully connected layer into class probabilities, thereby achieving multi-class classification tasks.

[0094] Activation functions play a crucial role in neural networks, providing nonlinear expressive capabilities. Their mathematical essence lies in constructing nonlinear decision boundaries. Removing activation functions degenerates multi-layered network architectures into linear regression models, rendering the neural network incapable of handling complex tasks. Common activation functions include ReLU (Rectified Linear Unit) and Sigmoid.

[0095] The ReLU activation function is a simple piecewise linear function. When the input is less than zero, the output is zero; when the input is greater than or equal to zero, the output equals the input. The advantages of ReLU are its simple computation and its constant gradient of 1 in the positive region. This characteristic avoids the vanishing gradient problem, allowing for better training of deep networks. However, the negative part of ReLU is truncated to zero, resulting in sparse activation, meaning only a portion of neurons are activated. While ReLU is widely used in forward propagation due to its sparse activation, the forced zeroing of neuron activation values ​​in the negative region limits the nonlinear representation of the feature space. To overcome the shortcomings of ReLU, a new activation function—the Mish activation function—is introduced. The Mish activation function is a self-regularized non-monotonic smooth function; it is differentiable over the entire real domain, making it smoother than ReLU in the negative region, preserving more information and gradient flow. The Mish activation function effectively solves the negative suppression defect of the traditional ReLU function through its non-monotonic response characteristics. The Mish function has advantages in feature capture: its information integrity in the negative range allows the hidden layers of the network to extract richer rut morphology features; its continuous differentiability ensures the stability of gradient flow in backpropagation and effectively suppresses gradient vanishing in deep networks.

[0096] To verify the effectiveness of the proposed model and the Mish activation function, two sets of ablation experiments were conducted in this embodiment: the first set was...

[0097] Ablation experiments were conducted using CNN-BiLSTM, CNN-LSTM, CNN, and BiLSTM models; the second group consisted of ablation experiments using Mish and ReLU activation functions. The results of the ablation experiments on dataset A are shown in Table 7. The model ablation experiments were based on an effective combination of CNN, LSTM, and BiLSTM modules. The CNN-BiLSTM model first extracts the spatial morphological features of the rut cross-sectional elevation data through convolutional layers, and then uses BiLSTM to obtain the rut shape waveform evolution pattern. Compared to other models, the CNN-BiLSTM model achieved a classification accuracy of 97.14%, with recall, precision, and F1-score all exceeding 90%, thus reducing false positives and false negatives in the classification task. Furthermore, compared to the conventional ReLU activation function, the network structure using the Mish activation function exhibits superior feature extraction capabilities, and its continuously differentiable nature preserves subtle gradient information of the input features. This provides a more complete parameter update path for the network to learn nonlinear relationships in complex feature spaces.

[0098] Table 7. Comparison of classification performance of activation functions and models (dataset A);

[0099] ;

[0100] The results of the ablation experiment on dataset B are shown in Table 8. The conclusions are the same as those for dataset A: the CNN-BiLSTM model performs better in terms of sensitivity and specificity, with a classification accuracy of 96.07%, and recall, precision, and F1-score all exceeding 85%. However, since the rut shapes for each class in this dataset are not significantly different, the number of rounds was increased to 30 during the experiment, while other metrics remained unchanged, allowing the model to reach a proper fit.

[0101] Table 8. Comparison of classification performance of activation functions and models (dataset B);

[0102] ;

[0103] Five-fold cross-validation is a commonly used model evaluation method, primarily used to assess a model's stability and generalization ability across different datasets. This method first divides the dataset into five equally sized subsets (or "folds"), and then comprehensively evaluates the model's performance through multiple training and testing iterations. Five-fold cross-validation not only provides a more intuitive evaluation result but also effectively reduces errors caused by the randomness of data partitioning. Its basic steps are as follows:

[0104] (1) Data partitioning: First, all sample data are randomly divided into 5 equal-sized subsets (folds) to ensure that the data contained in each subset is representative and to avoid bias. If the total number of samples is N, the size of each subset is N / 5, and the remaining 4 subsets are used as the training set.

[0105] (2) Model training and testing: In each round of cross-validation, the test set is one subset, and the other four subsets are the training set. The training set is then used to train the model, and the test set is used to evaluate the model performance.

[0106] (3) Repeated evaluation: Repeat the above training and testing steps 5 times. The purpose is to use each subset as a test set and the other 4 subsets as training sets. The model performance metrics obtained from each training will be recorded.

[0107] (4) Performance measurement calculation: After 5 evaluations, the results of each round of testing can be summarized and the average value of each evaluation index can be calculated to obtain the overall performance of the model.

[0108] (5) Final evaluation result: The final evaluation result is generally the average of the performance metrics of all five tests. This method, through multiple different training and testing sessions, can provide a comprehensive and stable estimate of the model performance, reducing the impact of random factors on the model evaluation results.

[0109] Therefore, through multiple training and testing iterations, five-fold cross-validation effectively reduces bias caused by data partitioning, fully assesses the model's performance on all data and its stability, and minimizes the risk of overfitting to a specific data partition. Unlike a single partitioning of the training and test sets, cross-validation ensures that every data point participates in both training and testing, resulting in a more reliable performance evaluation. Each training round utilizes 80% of the data, ensuring efficient data utilization. Cross-validation better leverages limited datasets, deriving more representative model performance metrics through multiple evaluations of different training and test data combinations, and avoiding evaluation bias caused by random data partitioning.

[0110] In this embodiment, after five-fold cross-validation, the mean values ​​of each category and the overall performance index of dataset A were calculated using statistical methods. The experimental results are shown in Table 9. As can be seen from the table, the overall precision, recall, accuracy, and composite score of dataset A in the three categories after five-fold cross-validation were 97.14%, 91.94%, 88.49%, and 88.45%, respectively.

[0111] Table 9 Evaluation results of the CNN-BiLSTM model on the test set (dataset A);

[0112] ;

[0113] Figure 7 The confusion matrix of the CNN-BiLSTM model on the test set is presented. According to the confusion matrix, the proposed model correctly identified 6804 U-shaped classifications with only 36 errors, and correctly identified 6897 convex W-shaped classifications. However, it correctly identified 6253 concave W-shaped classifications, with a relatively high number of errors. This is because the overall shape of convex W-shaped and concave W-shaped classifications is similar, and the only difference between them is the height of the middle peak point. This results in a relatively lower model evaluation result compared to the U-shaped classification. However, the method achieves good classification performance for the overall classification.

[0114] Step 3: Determine the loss function and optimizer for the neural network classification model;

[0115] The task of classifying tire rut shapes uses a neural network to train a model for feature learning, and the model ultimately outputs a category prediction. A loss function can be used to measure the difference between the model's predicted features and the actual samples; this function reflects the degree of model optimization. The smaller the loss function value, the better the model fits the data, and the higher the classification accuracy.

[0116] In this embodiment, the cross-entropy loss function is selected as the loss function for the tire rut shape classification task. The cross-entropy loss function quantifies the uncertainty of the classifier's output by calculating the information entropy difference between the model's predicted probability distribution and the true label distribution. The convergence direction of the loss value is negatively correlated with the model's optimization objective; when the loss approaches its minimum, it indicates that the log-likelihood estimate of the predicted probability and the true distribution has reached its optimal state.

[0117] Besides choosing an appropriate loss function, the optimization algorithm plays a crucial role in model training. The AdamW optimizer is adopted as the parameter update mechanism. Its algorithm architecture integrates historical gradient first-moment estimation from the momentum method with second-moment correction using the adaptive learning rate. Compared to the traditional Adam algorithm, AdamW effectively solves the parameter convergence bias problem that may be caused by L2 regularization in adaptive optimizers by explicitly decoupling the weighted loss term (weightdecay) from the gradient update operation. This characteristic gives it superior generalization performance in rut classification tasks, especially when dealing with non-stationary objective functions and sparse gradient scenarios.

[0118] Step 4: Introduce a self-attention mechanism and a learning rate scheduler to optimize the CNN-BiLSTM neural network model;

[0119] To further improve classification accuracy, a self-attention mechanism is introduced to construct a CNN-BiLSTM-Att model based on the CNN-BiLSTM neural network model. The self-attention mechanism calculates the attention of a sampling point in the rut data to itself and other sampling points, analyzes the correlation between sampling points in the rut cross-section elevation data sequence, adjusts the representation of each sampling point, and further improves the model performance.

[0120] The CNN-BiLSTM-Att model with self-attention mechanism integrates CNN, BiLSTM and self-attention mechanism. It uses CNN module to extract local features, BiLSTM to capture the contextual information of the sequence, and the self-attention mechanism to supplement the capture of global information.

[0121] The learning rate (LR) is a crucial hyperparameter in deep learning model training, controlling the step size for parameter updates. A reasonable learning rate directly influences the model's convergence dynamics and final performance. When the learning rate is too large, gradient explosion can be triggered at any time during training, leading to uncontrolled weight oscillations during parameter updates. Conversely, when the learning rate is too small, training becomes slow and may even get stuck in local optima. To balance these two factors, a learning rate scheduler is typically used to dynamically adjust the learning rate, allowing the model to gradually optimize its convergence and training speed during training.

[0122] Learning rate schedulers adjust the learning rate based on the progress of model training. Typically, the learning rate is gradually reduced according to certain rules during training to improve the model's generalization ability and accelerate convergence. ReduceLROnPlateau is a common learning rate scheduling strategy used to dynamically adjust the learning rate based on the model's performance (usually the loss on the validation set). Unlike other learning rate scheduling methods based on time steps or periods, ReduceLROnPlateau determines whether to adjust the learning rate based on the model's performance during training. When the validation loss or other evaluation metrics stop improving, it helps the model escape local optima and accelerate convergence by reducing the learning rate. Specifically, it monitors a specific metric (usually the loss on the validation set or other performance metrics). If this metric does not improve significantly within several training periods, the learning rate is reduced by multiplying it by a scaling factor. Reducing the learning rate helps the optimizer fine-tune parameters later in training, preventing overfitting or lingering near local optima.

[0123] Step 5: Calculate the rut depth based on the rut classification results;

[0124] By processing the cross-sectional elevation data of ruts using the optimal CNN-BiLSTM-ATt model, classification results of rut shapes can be obtained. For different rut classification results, various rut ​​depth calculation methods are designed. When dataset A covers three rut types—U-shaped, concave W-shaped, and convex W-shaped—a local extremum index function is first designed to determine whether there are peak points within 20 data points on each side of the rut curve, and the coordinates of these peak points are saved, such as... Figure 8 (a) shows the blue dots. Simultaneously, based on the rut type, the coordinates of the valley points are further located and saved, as shown below. Figure 8 (As shown by the red dots in (a)), these extreme points are crucial for constructing the virtual ruler during depth calculation, and the pseudocode is shown in Table 10. Specific situations are categorized into the following three types:

[0125] (1) Convex W-shaped ruts: Their significant characteristic is a peak point in the middle of the curve, with possible additional peak points on both sides. When calculating the rut depth, first determine the coordinates of the middle peak point, then select an appropriate virtual ruler construction method based on whether peak points exist on both sides. If peak points exist on both sides, the virtual ruler can be constructed by connecting the peak points on both sides to the middle peak point, or by connecting one peak point and the endpoint on the other side to the middle peak point; if there are no peak points on both sides, the virtual ruler is constructed by combining the middle peak point with the endpoints on both sides. After completing the virtual ruler construction, calculate the height difference between the valley points on the left and right sides and the ruler, thus obtaining the depth of the ruts on both sides, such as... Figure 8 As shown in (a).

[0126] (2) Concave W-shaped ruts: These typically contain two valley points, and may have peak points on both sides. A virtual ruler is constructed based on the presence of peak points on both sides. If peak points exist, the virtual ruler can be formed by connecting the two peak points, or by connecting one peak point to the endpoint on the other side; if no peak points exist, the virtual ruler is constructed through the endpoints on both sides. After constructing the virtual ruler, the height difference between the two valley points and the ruler is calculated to determine the rut depth on both sides, such as... Figure 8 As shown in (b).

[0127] (3) U-shaped ruts: Their cross-sectional curve has only one valley point. In calculation, a virtual ruler is first established at both ends, and then the height difference between the virtual ruler and the valley point is calculated to determine the rut depth, such as... Figure 8 As shown in (c).

[0128] Table 10 Pseudocode for rut depth calculation (Dataset A);

[0129] ;

[0130] When the classification result obtained by the CNN-BiLSTM-ATt model is based on three rut types (Type 0, Type 1, and Type 2) in dataset B, a local extremum indexing function is first designed to determine whether there are peak points in the data points between the 50th and 80th intervals of the rut curve, and the coordinates of these peak points are saved. Figure 9 (a) shows the blue dots. Additionally, the coordinates of the valley points are located and saved based on the rut type, as shown below. Figure 9 As shown by the red dots in (a), these extreme points will serve as an important basis for constructing the virtual ruler during the depth calculation process, and the pseudocode is shown in Table 11. Specifically, there are three types:

[0131] (1) Type 0 ruts: Characterized by the absence of peak points on both sides of the cross-sectional curve. During calculation, first determine if there is a peak point in the middle. If not, connect the two endpoints directly using a virtual ruler. If there is, compare the value of the middle peak point with the value of the left endpoint. When the value of the middle peak point is greater than the value of the left endpoint, connect the two endpoints directly using a virtual ruler. When the value of the middle peak point is less than the value of the left endpoint, construct a virtual ruler by combining the middle peak point with both endpoints. Then, calculate the height difference between the valley point and the ruler using the virtual ruler to obtain the rut depth, such as... Figure 9 As shown in (a).

[0132] (2) Type 1 ruts: Characterized by a peak point on only one side of the cross-sectional curve. During calculation, first determine which side of the rut curve has a peak point among 20 data points on each side, and save the coordinates of that peak point. Then compare the value of the middle peak point with the value of the left endpoint. When the value of the middle peak point is greater than the value of the left endpoint, a virtual ruler is constructed by connecting one peak point to the other endpoint. When the value of the middle peak point is less than the value of the left endpoint, a virtual ruler is constructed by connecting one peak point and the other endpoint to the middle peak point. Then, the height difference between the valley points on the left and right sides and the ruler is calculated using two virtual rulers, thus obtaining the depth of the ruts on both sides, such as... Figure 9 As shown in (b).

[0133] (3) Type 2 ruts: These ruts have peak points on both sides of their cross-sectional curve. During calculation, first determine which side of the rut curve has a peak point within each of the 20 data points, and save the coordinates of that peak point. Then, compare the value of the middle peak point with the value of the left endpoint. When the value of the middle peak point is greater than the value of the left endpoint, a virtual ruler is constructed by connecting the peak points on both sides. When the value of the middle peak point is less than the value of the left endpoint, a virtual ruler is constructed by connecting the peak points on both sides to the middle peak point. Subsequently, the height difference between the valley points on the left and right sides and the ruler is calculated using two virtual rulers, thus obtaining the depth of the ruts on both sides, such as... Figure 9 As shown in (c).

[0134] Table 11 Pseudocode for rut depth calculation (Dataset B).

[0135] ;

[0136] Example 2:

[0137] This embodiment provides a deep learning-based system for road rut classification and rut depth calculation, such as... Figure 10 As shown, it includes a data acquisition module, a data processing module, a rut shape classification module, a display module, a rut depth calculation module, and a data storage module;

[0138] The data acquisition module is used to acquire rut cross-sectional elevation data; a three-dimensional line laser detection system is used to acquire rut cross-sectional elevation data to realize rut shape recognition and classification and rut depth calculation functions. In this embodiment, the rut data records road cross-sectional elevation information and stores the data in a .csv format file;

[0139] The data processing module includes a three-level preprocessing architecture. First, noise reduction is achieved based on B-spline curve fitting. Second, a tilt correction model is constructed, and the height difference method and angle rotation method are used to eliminate the systematic deviation caused by road camber. Finally, due to the heavy computational burden of high-dimensional data, the interval sampling method is used to reduce the dimensionality of the data, thereby reducing the computational complexity while ensuring data quality.

[0140] The rut shape classification module classifies ruts according to different classification methods, providing preparation for subsequent depth calculation and maintenance decisions, and saves the classification results to the last column of the file;

[0141] The display module visualizes the cross-sections of ruts, including untreated rut cross-sections and cross-sections at any pre-processing stage. This function helps workers intuitively grasp the next processing steps during batch processing of rut cross-sections.

[0142] The rut depth calculation module calculates rut ​​depth based on the rut shape classification results and uses a virtual ruler algorithm.

[0143] The data storage module is used to save the rut shape classification and depth calculation results. In this embodiment, the rut shape classification results and depth calculation results are stored in tabular form to provide data support for rut maintenance decision analysis.

[0144] Example 3:

[0145] This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the deep learning-based road rut classification and rut depth calculation method.

[0146] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the deep learning-based road rut classification and rut depth calculation method as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.

[0147] The processor is used to execute all or part of the steps in the deep learning-based road rut classification and rut depth calculation method described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0148] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the deep learning-based road rut classification and rut depth calculation method described in the above embodiments.

[0149] Example 4:

[0150] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0151] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the deep learning-based road rut classification and rut depth calculation method described in the various embodiments of this application.

[0152] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) app stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned deep learning-based road rut classification and rut depth calculation method.

[0153] Example 5:

[0154] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the deep learning-based method for road rut classification and rut depth calculation.

[0155] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0156] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. A method for classifying road ruts and calculating rut depth based on deep learning, characterized in that, include: Elevation data of rut cross sections were acquired and preprocessed. Two datasets were established based on different classification methods of rut cross sections, and the correlation between rut type and shape was established. Two datasets, A and B, were constructed based on different classification methods for rut cross-sections. Dataset A categorizes rut ​​shapes into three types based on the distribution of the central peak point in the rut cross-section: U-shaped, concave W-shaped, and convex W-shaped. The U-shaped rut cross-section is characterized by the absence of a central peak point, exhibiting a smooth concave curve with rut depth concentrated in the middle. The concave W-shaped rut cross-section is characterized by two valley points and a central peak point, exhibiting two grooves with the central convex point lower than the road surface. The convex W-shaped rut cross-section is characterized by two grooves and a central peak point, with the central peak point higher than the sides, exhibiting a raised central section and relatively lower sides. Dataset B is classified according to the number of peak points at both ends of the rut cross section, specifically into three types: Type 0, Type 1, and Type 2. Type 0 rut cross sections are characterized by no peak points on both sides, with a smooth shape or depth concentrated in the middle; Type 1 rut cross sections are characterized by only one peak point at each end, exhibiting asymmetrical geometric characteristics; Type 2 rut cross sections are characterized by one peak point at each end, exhibiting symmetrical double-convex characteristics. We constructed neural network models with various structures for rut shape classification and selected the best classification model through ablation experiments. The rut depth is calculated based on the rut classification results; different virtual rulers are established for different categories of rut classification results to calculate the rut depth. When the classification results obtained by the CNN-BiLSTM-Att model are based on the three rut types of U-shaped, concave W-shaped, and convex W-shaped ruts in dataset A, a local extremum index function is first designed to determine whether there are peak points in the rut curve within 20 data points on each side, and the coordinates of these peak points are saved. At the same time, based on the rut type, the coordinates of the valley points are further found and saved. Different virtual rulers are established according to the peak points and valley points of different rut types, and then the rut depth is calculated. When the classification result obtained by the CNN-BiLSTM-Att model is based on the three rut types 0, 1 and 2 in dataset B, a local extremum index function is first designed to determine whether there are peak points in the rut curve between the 50th and 80th data points, and the coordinates of these peak points are saved. In addition, the coordinates of valley points are found and saved according to the rut type. Different virtual rulers are established according to the peak points and valley points of different rut types, and then the rut depth is calculated.

2. The method for road rut classification and rut depth calculation based on deep learning according to claim 1, characterized in that, The process of acquiring and preprocessing the cross-sectional elevation data of wheel ruts includes: A three-dimensional line laser detection system was used to obtain cross-sectional elevation data of vehicle ruts; The acquired rut cross-sectional elevation data were processed by smoothing and noise reduction, cross-sectional tilt correction, and dimensionality reduction.

3. The method for road rut classification and rut depth calculation based on deep learning according to claim 2, characterized in that, The specific method for tilt correction of the cross section of the rut is as follows: The elevation difference method or angle rotation method is used to eliminate non-realistic deformation data caused by equipment posture deviation during the measurement process, and to ensure that the horizontal projection baseline formed by the boundary points on both sides of the cross section is completely coincident with the X-axis.

4. The method for road rut classification and rut depth calculation based on deep learning according to claim 3, characterized in that, The construction of various neural network models for rut shape classification specifically includes CNN model, BiLSTM model, CNN-LSTM model and CNN-BiLSTM model; the CNN-BiLSTM model was selected as the best classification model through ablation experiments.

5. The method for road rut classification and rut depth calculation based on deep learning according to claim 4, characterized in that, The method selects the cross-entropy loss function as the loss function for the rut shape classification task and uses the AdamW optimizer as the model parameter update mechanism.

6. The method for road rut classification and rut depth calculation based on deep learning according to claim 5, characterized in that, The method introduces a self-attention mechanism and a learning rate scheduler to optimize the CNN-BiLSTM model; A self-attention mechanism is introduced to construct the CNN-BiLSTM-Att model based on the CNN-BiLSTM model. The self-attention mechanism calculates the attention of a sampling point in the rut cross-section data to itself and other sampling points, analyzes the correlation between sampling points in the rut cross-section elevation data sequence, and adjusts the representation of each sampling point to further improve the model performance. The learning rate scheduler is used to dynamically adjust the learning rate of the CNN-BiLSTM model, so that the model can gradually optimize its convergence effect and training speed during the training process.

7. The method for road rut classification and rut depth calculation based on deep learning according to claim 6, characterized in that, When the classification result obtained by the CNN-BiLSTM-ATt model is based on the three rut types—U-shaped, concave W-shaped, and convex W-shaped—in dataset A, the rut depth is calculated in the following three cases: (1) Convex W-shaped ruts: When calculating the rut depth, first determine the coordinates of the middle peak point, and then select the virtual ruler construction method according to whether there are peak points on both sides; if there are peak points on both sides, the virtual ruler is connected to the middle peak point by the peak points on both sides respectively, or by one peak point and the other end point connected to the middle peak point respectively; if there are no peak points on both sides, the virtual ruler is constructed by the middle peak point and the two end points; after the virtual ruler is constructed, the height difference between the valley point on the left and right sides and the virtual ruler is calculated respectively, and then the depth of the ruts on the left and right sides is obtained. (2) Concave W-shaped ruts: A virtual ruler is established based on whether there are peak points on both sides of the valley point. If there are peak points, the virtual ruler is connected by two peak points, or by one peak point and the other end point. If there are no peak points, the virtual ruler is established through the two end points. After the virtual ruler is constructed, the height difference between the two valley points on the left and right sides and the virtual ruler is calculated to obtain the rut depth on the left and right sides respectively. (3) U-shaped ruts: First, establish a virtual ruler through the two ends, and then calculate the height difference between the virtual ruler and the valley point to determine the rut depth; When the classification result obtained by the CNN-BiLSTM-ATt model is based on the three rut types (Type 0, Type 1, and Type 2) in dataset B, the rut depth is calculated in the following three cases: (1) Type 0 ruts: First, determine whether there is a peak point in the middle of the cross-sectional curve. If not, the virtual ruler is directly connected from both ends. If there is, compare the value of the middle peak point with the value of the left end point. When the value of the middle peak point is greater than the value of the left end point, the virtual ruler is directly connected from both ends. When the value of the middle peak point is less than the value of the left end point, the virtual ruler is constructed by combining the middle peak point and the two ends. Then, the height difference between the valley point and the virtual ruler is calculated through the virtual ruler to obtain the depth of the rut. (2) Type 1 ruts: First, determine which side of the rut cross-section curve has a peak point among 20 data points on each side, and save the coordinates of the peak point; then compare the value of the middle peak point with the value of the left endpoint. When the value of the middle peak point is greater than the value of the left endpoint, the virtual ruler is connected to the other endpoint by one peak point; when the value of the middle peak point is less than the value of the left endpoint, the virtual ruler is constructed by connecting one peak point and the other endpoint to the middle peak point respectively; then, the height difference between the valley points on the left and right sides and the ruler is calculated by the two virtual rulers respectively, so as to obtain the depth of the ruts on the left and right sides. (3) Type 2 ruts: First, determine which side of the rut cross section curve has a peak point among 20 data points on each side, and save the coordinates of the peak point; then compare the value of the middle peak point with the value of the left endpoint. When the value of the middle peak point is greater than the value of the left endpoint, the virtual ruler is connected by the peak points on both sides; when the value of the middle peak point is less than the value of the left endpoint, the virtual ruler is constructed by connecting the peak points on both sides to the middle peak point respectively; then, the height difference between the valley points on the left and right sides and the ruler is calculated by the two virtual rulers respectively, so as to obtain the depth of the ruts on the left and right sides.

8. A deep learning-based system for classifying road ruts and calculating rut depth, implemented based on the deep learning-based method for classifying road ruts and calculating rut depth as described in claim 1, characterized in that, It includes a data acquisition module, a data processing module, a rut shape classification module, a display module, a rut depth calculation module, and a data storage module; The data acquisition module is used to acquire the cross-sectional elevation data of the ruts; a three-dimensional line laser detection system is used to acquire the cross-sectional elevation data of the ruts, which is used to realize the rut shape recognition and classification and the rut depth; The data processing module includes a three-level preprocessing architecture. First, noise reduction is achieved based on B-spline curve fitting. Second, a tilt correction model is constructed, and the height difference method and angle rotation method are used to eliminate the systematic deviation caused by road camber. Finally, the interval sampling method is used to reduce the dimensionality of the data. The rut shape classification module classifies ruts according to different classification methods; The display module visualizes the cross-section of ruts, including untreated rut cross-sections and cross-sections at any pre-processing stage. The rut depth calculation module calculates rut ​​depth by establishing different virtual rulers based on the rut shape classification results. The data storage module is used to save the rut shape classification and depth calculation results.

9. A computer program product for executing the deep learning-based road rut classification and rut depth calculation method according to any one of claims 1-7, characterized in that, This includes a computer program or instructions that, when executed by a processor, implement the deep learning-based method for road rut classification and rut depth calculation.

Images