Suspension vibration response-based deep learning model road surface grade identification method

Abstract

The application discloses a kind of based on suspension vibration response's deep learning model road surface grade identification method, comprising: step 1, constructs multi-working condition suspension vibration response time series dataset, step 2, adopts data set division and standardization processing of preventing data leakage, step 3, constructs extreme lightweight Transform road surface grade identification model, step 4, model verification and cross-platform deployment.The application applies Transform architecture to suspension vibration time series data, can break through the limitation of local receptive field through self-attention mechanism, directly establish the global feature association between any time step, accurately capture cross-time step long-range dependence features, effectively improve the identification accuracy of adjacent road surface grade and model generalization performance.

Description

Technical Field

[0001] This invention relates to the field of road surface detection technology, and in particular to a method for identifying road surface grades based on a deep learning model of suspension vibration response. Background Technology

[0002] Road surface roughness assessment is a crucial foundation for road maintenance, intelligent vehicle control, and traffic safety. According to the ISO 8608 standard, road surface roughness can be classified into multiple consecutive grades based on power spectral density, covering various road surfaces from highways to rural roads. Different road surface grades directly impact vehicle ride comfort, fuel economy, and safety.

[0003] Existing road surface grade identification technologies are mainly divided into three categories: manual inspection method, visual image detection method, and road surface detection method based on vehicle-mounted sensors. However, the manual inspection method relies on professionals to use road surface evenness instruments for on-site measurement, which has the disadvantages of low efficiency, high labor intensity, and strong subjectivity, making it difficult to meet the daily maintenance needs of large-scale road networks.

[0004] Visual image detection methods use vehicle-mounted cameras to collect road surface images and employ computer vision technology to identify road surface damage and smoothness. However, they are easily affected by environmental factors such as lighting, weather, and road debris, and their recognition accuracy decreases significantly under adverse conditions such as nighttime, tunnels, rain, and snow.

[0005] Road surface detection methods based on vehicle-mounted sensors offer advantages such as all-weather operation, low cost, and ease of deployment. Among these, road surface recognition methods based on suspension vibration response signals are the most mature. This method utilizes the suspension vibration response caused by road excitation during vehicle movement to invert road surface smoothness, eliminating the need for additional dedicated detection equipment and enabling large-scale, crowdsourced road surface data collection using ordinary passenger vehicles. However, existing road surface grade recognition technologies based on suspension vibration response signals still have the following areas for optimization: Data processing procedures pose a risk of information leakage: Some deep learning research uses random slicing to divide the dataset, which may result in different segments of the same original signal appearing in both the training and validation sets, causing information leakage and leading to a discrepancy between the model's performance on the validation set and its actual deployment effect. At the same time, some standardized processes do not strictly follow the "divide before compute" principle, further introducing unnecessary risks of information leakage.

[0006] There are limitations to the ability to model long-term features: traditional machine learning methods rely on manually designed features, which do not fully explore the long-term dependencies of suspension vibration response signals; convolutional neural networks are limited by the characteristics of local receptive fields, and have a natural weakness in modeling the global dependencies of long-term vibration response signals, and there is still room for improvement in the recognition accuracy of medium and low-grade road surfaces with high feature overlap.

[0007] Lightweight design and real-time performance optimization need to be strengthened: While pursuing recognition accuracy, most studies have not adequately considered the limitations of computing resources on vehicle edge devices. Some models have a large number of parameters and slow inference speed, and the efficiency of real-time deployment on low-computing vehicle platforms needs to be improved. At the same time, some methods have not been specifically optimized for the problem of unbalanced distribution of road surface grade samples, and the recall rate of low-grade roads still needs to be further improved.

[0008] Therefore, there is an urgent need for a general-purpose ISO8608 standard road surface grade identification method that balances recognition accuracy and generalization ability and has the advantage of lightweight deployment. Summary of the Invention

[0009] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a road surface grade identification method based on a deep learning model of suspension vibration response.

[0010] The technical solution adopted to achieve the purpose of this invention is: A method for identifying road surface grades based on a deep learning model of suspension vibration response, characterized by the following steps: Step 1: Simulate and generate suspension vibration response signals under different working conditions using a two-degree-of-freedom 1 / 4 vehicle suspension dynamics model. Perform sliding window slicing and data augmentation on the suspension vibration response signals to obtain time-series datasets of road surface grades corresponding to each working condition. Step 2: Divide the time series datasets of each road surface level obtained in Step 1 into training set, validation set and independent test set; calculate the statistics of each channel signal based on the training set, and perform standardization processing on all datasets based on the statistics of each channel signal. Step 3: Train the Transformer road surface level recognition model using the standardized training set from Step 2. Output the predicted probability of each road surface level in each training round and verify the accuracy using the validation set. After training, select the parameters of the round with the highest accuracy as the parameters of the best Transformer road surface level recognition model. Step 4: Export the best Transformer pavement grade recognition model to a preset format, test it using the independent test set obtained in Step 2, calculate the prediction probability of the best Transformer pavement grade recognition model for the test samples, and calculate the overall recognition accuracy and the recognition accuracy of each pavement grade.

[0011] In the above technical solution, in step 1, different road surface level excitations are set according to the ISO8608 standard, and the suspension vibration response signals under different working conditions are generated by simulation using a two-degree-of-freedom 1 / 4 vehicle suspension dynamics model. The suspension vibration response signals include three channels: suspension travel, vehicle speed, and vehicle acceleration. In the above technical solution, during the sliding window slicing process of the suspension vibration response signal in step 1, all slices of the same suspension vibration response signal share the same original suspension vibration response signal ID. The data enhancement includes adding Gaussian white noise with a preset standard deviation, random amplitude scaling, random time offset, and small time stretching or compression.

[0012] In the above technical solution, step 2, the time series dataset partitioning process is specifically as follows: extract the last predetermined proportion as an independent test set according to the arrangement order of the original suspension vibration response signal IDs, perform stratified random sampling based on road surface grade and using the original suspension vibration response signal IDs as the unit, and divide the remaining time series dataset into training set and validation set according to the predetermined proportion.

[0013] In the above technical solution, in step 2, the statistics of each channel signal are the mean and standard deviation of suspension travel, vehicle speed, and vehicle acceleration calculated based on the entire training set. Z-score standardization is then uniformly applied to all datasets using the statistics of each channel signal. The Z-score standardization process is expressed as follows: ; In the formula, The samples are Z-score standardized. As a sample, This is the average of suspension travel, vehicle speed, or vehicle acceleration. This represents the standard deviation of suspension travel, vehicle speed, or vehicle acceleration.

[0014] In the above technical solution, in step 3, the Transformer road surface level recognition model consists of an input projection layer, a position encoding layer, a Transformer encoder, and a classification head. The training process of the Transformer road surface level recognition model includes two stages. In the first stage, the Adam optimizer is used for full parameter training. In the second stage, freeze-fine-tuning is used to freeze all parameters of the input projection layer and the Transformer encoder, and only the classification head is fine-tuned.

[0015] In the above technical solution, the input projection layer includes a one-dimensional convolution, a one-dimensional batch normalization layer, a modified linear unit activation function, and a random deactivation structure; the position encoding layer adopts sine and cosine position encoding; the Transformer encoder includes a pre-normalization structure, a multi-head attention head, and a Gaussian error linear unit, with residual connections and a random deactivation structure added after each sub-layer; the classification head includes a random deactivation structure, a fully connected layer, and a Softmax layer.

[0016] A second aspect of the present invention is an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the road surface grade identification method based on a deep learning model of suspension vibration response.

[0017] A third aspect of the present invention is a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the road surface grade identification method based on a deep learning model of suspension vibration response.

[0018] A fourth aspect of the present invention is a computer program product comprising computer-executable instructions, which, when executed, implement a road surface grade identification method based on a deep learning model of suspension vibration response.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention is not limited to a specific road surface grade. By flexibly adjusting the output dimensions of the classification head, it can adapt to all road surface grade classification requirements defined by the ISO8608 standard, and has good scalability. 2. This invention adopts a dataset partitioning method based on hierarchical sampling of the original suspension vibration response signal ID. First, the independent test set is sealed, and then the training set and validation set are partitioned. At the same time, it combines a strict leak-free standardization process to effectively avoid information leakage problems during data processing. With the addition of four data augmentation strategies designed based on physical meaning, the robustness of the model in recognition under different vehicle speeds, different suspension parameters, and different road surface grades is significantly improved. The end-to-end anti-information leakage design significantly improves the generalization ability. 3. This invention deeply integrates the Transformer architecture with suspension vibration response data, offering significant technical advantages. Images are two-dimensional spatial data, with features dependent on local pixel texture associations, making them suitable for the local receptive field mechanism of convolutional neural networks. Suspension vibration response signals, on the other hand, are one-dimensional time-series data, generated by road surface excitation coupled through suspension dynamics. They feature strong coupling of multiple physical quantities and high overlap of features at adjacent levels. Applying the Transformer architecture to this type of one-dimensional suspension vibration time-series data in long-term statistical dependencies can overcome the limitations of local receptive fields through a self-attention mechanism, directly establishing global feature associations between any time steps, accurately capturing long-range dependency features across time steps, and effectively improving the recognition accuracy of adjacent road surface levels and the model's generalization performance. 4. This invention is based on the Transformer architecture and is optimized for extreme lightweight design. The total number of model parameters is only 0.21M, and the single-sample inference latency on the CPU is only 68ms, which meets the real-time requirements of automotive edge devices. It supports export to the ONNX universal format and can be quickly deployed across platforms. 5. This invention uses a labeled smooth Focal Loss loss function, which can flexibly adjust the loss parameters according to the sample distribution of the actual road surface grade, effectively alleviate the training bias caused by sample imbalance, has strong adaptability to sample imbalance, and improves the recognition accuracy of low-grade roads. 6. This invention fully leverages the inherent advantages of the Transformer architecture in long-term dependency modeling, enabling more accurate mining of temporal features in suspension vibration response signals, and demonstrating excellent long-term feature modeling capabilities. Experimental results show that model misjudgments only occur between adjacent road surface levels, with no cross-level misjudgments, fully conforming to the physical variation law of road surface unevenness. 7. While maintaining recognition accuracy comparable to traditional machine learning methods, this invention has stronger long-term feature extraction and generalization capabilities. At the same time, the model is small in size and has a fast inference speed, making it a promising application in automotive edge deployment scenarios, thus balancing accuracy and deployment efficiency. Attached Figure Description

[0020] Figure 1 The diagram shows the flowchart of the road surface grade identification method based on the deep learning model of suspension vibration response according to the present invention.

[0021] Figure 2 The diagram shown is a structural schematic of the Transformer road surface grade recognition model of the present invention.

[0022] Figure 3 The figure shows a comparison curve of training and validation losses for the Transformer road surface classification model of this invention. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] Example 1 A method for road surface grade identification based on a deep learning model of suspension vibration response, referring to... Figure 1 This includes the following steps: Step 1: Construct a multi-condition suspension vibration response time-series dataset: Based on the ISO8608 standard, different levels of road surface excitation are set, and the suspension vibration response signals under different conditions are generated by simulation using a two-degree-of-freedom 1 / 4 vehicle suspension dynamics model. The suspension vibration response signals are then processed by sliding window slicing and data augmentation to obtain the time-series datasets corresponding to the road surface level for each condition.

[0025] Furthermore, the road surface grades described in this embodiment are divided into four grades: A, B, C, and D according to the ISO8608 standard. The corresponding suspension vibration response signals are generated by simulation using a two-degree-of-freedom 1 / 4 vehicle suspension dynamics model.

[0026] Simulation parameters are set for different operating conditions. The simulation parameters include vehicle body mass, unsprung mass, suspension stiffness, and vehicle speed. In this embodiment, the vehicle body mass is set to 260~320kg, the unsprung mass is set to 40~55kg, the suspension stiffness is set to 14000~18000N / m, and the vehicle speed is set to 30~100km / h.

[0027] The suspension vibration response signal is a three-channel signal including suspension travel, vehicle speed, and vehicle acceleration. In this embodiment, the sampling frequency is set to 100Hz, and the simulation duration for a single signal is 500 seconds. Input excitation based on the corresponding road surface level is generated according to the ISO8608 standard, substituted into the two-degree-of-freedom suspension dynamics differential equation, and solved point-by-point numerically using the Euler integral method, outputting a three-channel vibration response signal with a sampling rate of 100Hz and a duration of 500 seconds.

[0028] Furthermore, in this embodiment, the window length of the sliding window slice is set to 256 sampling points (corresponding to 2.56 seconds), the step size is set to 128 sampling points, 390 samples are generated for each working condition, and a globally unique suspension vibration response signal ID is assigned to each original continuous suspension vibration response signal. All slices of the same suspension vibration response signal share the same suspension vibration response signal.

[0029] The data augmentation of the suspension vibration response signal includes: adding Gaussian white noise with a standard deviation of 0.03 (simulating sensor noise, activation probability 0.5), random amplitude scaling of 0.95~1.05 (simulating suspension characteristic differences, activation probability 0.4), random time offset of ±3 sampling points (simulating sampling synchronization error, activation probability 0.3), and small time stretching / compression of 0.95~1.05 (simulating vehicle speed fluctuations). The final result is a time-series dataset covering all ISO 8608 road surface grades.

[0030] Step 2, adopting data leakage prevention dataset partitioning and standardization processing: extract a certain proportion of the samples from the time series datasets of each road surface level obtained in Step 1 as independent test sets, and divide the remaining samples from the time series datasets into training sets and validation sets according to a predetermined ratio.

[0031] Furthermore, the samples in the time-series datasets for each road surface level under each working condition share the same original suspension vibration response signal ID, ensuring that all slices of the same original signal appear only in a single dataset, thus preventing data leakage at its source. In this embodiment, 20% of the samples from the end of each road surface level time-series dataset are extracted as an independent test set based on the original suspension vibration response signal ID. The remaining samples are then divided into a training set and a validation set in a 75%:25% ratio.

[0032] The independent test set is not involved in model training and parameter tuning throughout the entire process, but is used for final model performance verification.

[0033] The leak-free standardization of samples strictly follows the principle of "first partitioning the dataset, then calculating the standardization parameters": After partitioning the dataset, Z-score standardization is uniformly performed on all datasets based on the statistics of each channel signal of the training set samples. The statistics of each channel signal are the mean and standard deviation of suspension travel, vehicle speed, and vehicle acceleration calculated based on the entire training set. The Z-score standardization process is expressed as follows: ; In the formula, The samples are Z-score standardized. As a sample, This is the average of suspension travel, vehicle speed, or vehicle acceleration. This represents the standard deviation of suspension travel, vehicle speed, or vehicle acceleration.

[0034] By employing a two-pronged approach of segmentation strategy and standardized processes, information leakage is avoided, realistically simulating the scenario where "input data is completely unknown" during vehicle deployment.

[0035] In this embodiment, the independent test set is divided as follows: within each road surface level (A, B, C, D), the last 20% of the complete original signals are extracted according to the order of the original suspension vibration response signal IDs, and all slice samples of the corresponding signals are included in the independent test set. This set does not participate in model training, hyperparameter tuning, or model selection throughout the entire process; it is only used for the final generalization performance evaluation after training. Its data distribution is consistent with the full dataset, which can truly reflect the recognition effect when the model is deployed.

[0036] The division of the training and validation sets is as follows: All remaining raw signals after extracting the independent test set are stratified based on road surface grade and randomly sampled using the original suspension vibration response signal ID as the unit. The samples are then divided into training and validation sets at a ratio of 75%:25%, ensuring that the sample proportions for each road surface grade within the training and validation sets are consistent with the overall distribution. The training set is used for iterative learning of the model weights, while the validation set is used for performance monitoring and optimal model selection during the training process.

[0037] Step 3, construct an ultra-lightweight Transformer road surface level recognition model: use the training set after standardization in Step 2 to train the deep learning model (Transformer) road surface level recognition model. After each training round, output the predicted probability of the road surface level recognition result and use the validation set to test the recognition accuracy. After training, select the parameters of the round with the highest accuracy as the parameters of the best Transformer road surface level recognition model. Furthermore, refer to Figure 2 The Transformer road surface classification model consists of an input projection layer, a position encoding layer, a Transformer encoder, and a classification head. All of these components are based on well-known fundamental components in the field of deep learning. The model achieves a lightweight design by simplifying the number of layers and optimizing the dimensions, making it compatible with vehicle-mounted edge computing power and flexibly adaptable to different road surface classification needs.

[0038] In this embodiment, the input projection layer uses a one-dimensional convolution (kernel size 3, padding 1, stride 1), a one-dimensional batch normalization layer, a corrected linear unit, and a random deactivation structure to map the 3-channel suspension vibration response signal into a feature vector with a hidden dimension of 128, with a random deactivation rate of 0.25.

[0039] The position encoding layer uses sine and cosine position encoding with a maximum length of 5000. Position information is injected into each temporal position and added element by element to the output of the input projection layer.

[0040] The Transformer encoder adopts a 1-layer pre-normalization structure, with 4 multi-head attention heads, each head having a dimension of 32; the feedforward network has a hidden dimension of 256 and Gaussian error linear units; residual connections and random deactivation structures are added after each sub-layer, with a random deactivation rate of 0.5%.

[0041] The classification head includes a random deactivation structure, a fully connected layer, and a Softmax layer. The classification head first performs global mean pooling on the encoder output in the time dimension to obtain a 128-dimensional global feature vector. Then, it passes through the random deactivation structure and the fully connected layer to map the number of road surface class categories corresponding to the ISO 8608 standard. The predicted probability of each road surface class is output through Softmax.

[0042] The Transformer road surface classification model has only 0.21M parameters and can be adapted to various computing platforms such as vehicle edge CPUs, industrial control computers, and NVIDIA / AMD GPUs. The classification head dimension can be flexibly adjusted according to actual needs.

[0043] Furthermore, the Transformer road surface classification model training process includes two stages: the first stage employs full-parameter training, and the second stage uses freeze-and-fine-tuning. This targeted, optimized two-stage model training strategy, combined with a dedicated loss function and learning rate scheduling strategy, optimizes the model's convergence and generalization capabilities, making it adaptable to road surface classification tasks with varying numbers of road surface levels.

[0044] Furthermore, the first stage of full-parameter training (0-60 epochs in this embodiment) uses the Adam optimizer with an initial learning rate of 2×10⁻⁶. -4 Weight decay 1×10 -3 The gradient clipping maximum norm is 0.5; the learning rate scheduling adopts a 5-round warm-up + cosine annealing strategy. During the warm-up phase, the learning rate linearly increases from 0 to the initial value, and then decays to 1×10⁻⁶ in the subsequent 55 rounds. -6 The loss function uses labeled smoothing Focal Loss, which can adjust the focusing parameter γ and the class balance coefficient α according to the actual road surface grade distribution to alleviate the sample imbalance problem; the batch size is 256, and online data augmentation is used.

[0045] The second stage, freeze-fine-tuning (61-100 rounds in this embodiment), freezes all parameters of the input projection layer and Transformer encoder, and only fine-tunes the classification head, optimizing the classification decision boundary while preserving learned general features; the learning rate is reduced to 2×10. -5 Continue training for 40 rounds; Step 4, Model Validation and Cross-Platform Deployment: Export the best Transformer road surface classification model trained in Step 3 to ONNX format (Opset=14, supports dynamic batching) to achieve cross-platform deployment; use the independent test set obtained in Step 2 to test, calculate the overall accuracy of the prediction probability of road surface classification and the accuracy of the prediction probability of each road surface classification, and comprehensively evaluate the model performance; deploy and validate on an in-vehicle edge device to test single-sample inference latency and real-time performance.

[0046] Example 2 This embodiment details the verification of typical pavement grades A, B, C, and D according to the ISO 8608 standard. This embodiment is only used to explain the invention and is not intended to limit the scope of protection of the invention. Those skilled in the art can extend this method to other pavement grades according to actual needs.

[0047] The dataset was generated through simulation using commercial software; model training and validation were based on Python, PyTorch, and CUDA environments, with hardware configuration including a computer with an NVIDIA RTX 3050 6GB graphics card; model deployment and inference used the ONNX Runtime inference engine.

[0048] A method for road surface grade identification based on a deep learning model of suspension vibration response includes the following steps: Step 1: Construction of a multi-condition suspension vibration response dataset: A two-degree-of-freedom 1 / 4 vehicle suspension dynamics model was adopted, and simulation parameters were set. These parameters included: vehicle mass 260~320kg (step 20kg), unsprung mass 40~55kg (step 5kg), suspension stiffness 14000~18000N / m (step 2000N / m), damping coefficient ±10% random disturbance, and vehicle speed 30~100km / h (step 10km / h). Four levels of typical road surface excitation (A, B, C, D) were generated based on the ISO8608 standard as experimental verification samples. A single signal was simulated for 500 seconds at a sampling rate of 100Hz, and three channels of suspension vibration response signals were collected. A 256-point window and a 128-point step slice were used to generate 390 samples for each condition. Four types of data augmentation driven by physical meaning were added to the training set. The final time series dataset contains 1,536 raw signals and 597,500 samples, with each sample assigned a unique raw suspension vibration response signal ID.

[0049] Step 2, Data Leakage Prevention Preprocessing and Standardization: First, extract 20% of the original signal slices according to road surface grade as an independent test set, with a total sample size of 119,500. The remaining samples are hierarchically divided by the original suspension vibration response signal ID, into training and validation sets at a ratio of 75%:25%, with corresponding sample sizes of 358,500 and 119,500, respectively. Following the "divide first, standardize later" leakage prevention rule, only the mean and standard deviation of the 3-channel vibration data are calculated based on the training set, and Z-score standardization is performed on the entire dataset. The dataset is then uniformly converted into a PyTorch tensor of dimension (N, 3, 256) and the standardized statistical parameter file is stored.

[0050] Step 3, Transformer road surface grade recognition model construction: The input projection layer uses 1D convolution (kernel 3, padding 1, stride 1), 1D batch normalization layer, modified linear unit activation function, and random deactivation structure (random deactivation rate 0.25), with an output dimension of (batch_size, 128, 256); the position encoding layer uses sine and cosine position encoding with a maximum length of 5000, and an output dimension of (batch_size, 256, 128); the Transformer encoder is a 1-layer pre-normalization structure, 4-head attention, feedforward network dimension 256, modified linear unit activation function, and random deactivation structure (random deactivation rate 0.5), with an output dimension of (batch_size, 256, 128); the global mean pooling layer has an output dimension of (batch_size, 128); the classification head uses a random deactivation structure (random deactivation rate 0.5), fully connected layers (128→4), and Softmax, corresponding to 4 typical road surface grades. The Transformer road surface grade identification model has a total of 0.21M parameters.

[0051] The Transformer road surface classification model employs a two-stage training process: the first stage (0-60 epochs) uses the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -4 Weight decay 1×10 -3 Gradient clipping 0.5; 5 rounds of warm-up + cosine annealing learning rate; Focal Loss ( , Label smoothing was 0.15; batch size was 256, and online data augmentation was implemented. In the second stage (61-100 epochs), the input projection layer and encoder were frozen, and only the classification head was trained, with the learning rate reduced to 2×10⁻⁶. -5 Save the model parameters that achieve the highest accuracy on the validation set in each round. (Refer to...) Figure 3 During model training, both training loss and validation loss decreased rapidly and synchronously. After about 20 rounds, the model entered a stable convergence phase, and after 80 rounds, the loss stabilized at around 0.41. Throughout the training process, the validation loss did not show a continuous increase and had a good fit with the training loss, indicating that the model did not overfit and the two-stage training strategies were smoothly connected.

[0052] Step 4, Model Performance Validation and Deployment: Generalization performance was validated using an independent test set. The best Transformer road surface classification model achieved a high overall accuracy. Differences existed in the recognition performance across different road surface grades, with the smooth road surface showing the best accuracy, while the bumpy road surface recognition accuracy was moderate. The best Transformer road surface classification model exhibited short single-sample inference time on ordinary CPU devices, meeting the real-time recognition requirements for in-vehicle systems. The trained best Transformer road surface classification model was exported to the ONNX universal inference format, supporting dynamic batch input and adaptable to various in-vehicle edge hardware platforms for deployment.

[0053] Example 3 The difference between this embodiment and Embodiment 1 is that the input projection layer of the Transformer road surface grade recognition model is a two-layer 1D convolutional structure with kernel sizes of 5 and 3, which further enhances the ability to extract local features. The number of model parameters increases to 0.28M, and the inference latency is about 75ms, which still meets the real-time requirements.

[0054] In the first stage of full-parameter training, the weighted combination of the loss functions Dice Loss and Focal Loss, with a weight ratio of 1:1, can further improve the recall rate of low-level roads by about 2.1%.

[0055] When deploying the Transformer road surface classification model, TensorRT can be used to accelerate the ONNX format Transformer road surface classification model by performing INT8 quantization. After quantization, the inference latency can be reduced to 32ms, and the accuracy loss is less than 0.5%.

[0056] The output dimension of the fully connected layer of the classification head can be flexibly adjusted according to the number of ISO8608 road surface grades to be identified as needed, without modifying the main structure of the model.

[0057] Example 4 This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the road surface grade identification method based on a deep learning model of suspension vibration response as described in Embodiment 1 above.

[0058] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0059] Example 5 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the road surface grade identification method based on a deep learning model of suspension vibration response as described in Embodiment 1 above.

[0060] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0061] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0062] Example 6 This embodiment provides a computer program product including computer-executable instructions stored on a machine-readable storage medium (such as a disk, flash memory, or optical disk). When executed by at least one data processing device (such as a microprocessor or digital signal processor), the instructions cause the data processing device to implement the steps in the road surface grade identification method based on a deep learning model of suspension vibration response described in Embodiment 1.

[0063] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0064] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0065] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0066] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying road surface grades based on a deep learning model of suspension vibration response, characterized in that, Includes the following steps: Step 1: Simulate and generate suspension vibration response signals under different working conditions using a two-degree-of-freedom 1 / 4 vehicle suspension dynamics model. Perform sliding window slicing and data augmentation on the suspension vibration response signals to obtain time-series datasets of road surface grades corresponding to each working condition. Step 2: Divide the time series datasets of each road surface level obtained in Step 1 into training set, validation set and independent test set; calculate the statistics of each channel signal based on the training set, and perform standardization processing on all datasets based on the statistics of each channel signal. Step 3: Train the Transformer road surface level recognition model using the standardized training set from Step 2. Output the predicted probability of each road surface level in each training round and verify the accuracy using the validation set. After training, select the parameters of the round with the highest accuracy as the parameters of the best Transformer road surface level recognition model. Step 4: Export the best Transformer pavement grade recognition model to a preset format, test it using the independent test set obtained in Step 2, calculate the prediction probability of the best Transformer pavement grade recognition model for the test samples, and calculate the overall recognition accuracy and the recognition accuracy of each pavement grade.

2. The method for road surface grade identification based on a deep learning model of suspension vibration response as described in claim 1, characterized in that, In step 1, different road surface levels of excitation are set according to the ISO8608 standard, and the suspension vibration response signals under different working conditions are generated by simulation using a two-degree-of-freedom 1 / 4 vehicle suspension dynamics model. The suspension vibration response signals are three-channel signals including suspension travel, vehicle speed, and vehicle acceleration.

3. The method for road surface grade identification based on a deep learning model of suspension vibration response as described in claim 1, characterized in that, In step 1, during the sliding window slicing process of the suspension vibration response signal, all slices of the same suspension vibration response signal share the same original suspension vibration response signal ID. The data enhancement includes adding Gaussian white noise with a preset standard deviation, random amplitude scaling, random time offset, and small time stretching or compression.

4. The method for road surface grade identification based on a deep learning model of suspension vibration response as described in claim 1, characterized in that, In step 2, the time series dataset partitioning process is as follows: according to the arrangement order of the original suspension vibration response signal IDs, the last predetermined proportion is extracted as an independent test set. The time series dataset after extracting the independent test set is stratified according to the road surface grade and the original suspension vibration response signal ID is used as the unit for stratified random sampling. The remaining time series dataset is divided into training set and validation set according to the predetermined proportion.

5. The method for road surface grade identification based on a deep learning model of suspension vibration response as described in claim 1, characterized in that, In step 2, the statistics for each channel signal are the mean and standard deviation of suspension travel, vehicle speed, and vehicle acceleration calculated based on the entire training set. Z-score standardization is then uniformly applied to all datasets using the statistics for each channel signal. The Z-score standardization process is expressed as follows: ； In the formula, The samples are Z-score standardized. As a sample, This is the average of suspension travel, vehicle speed, or vehicle acceleration. This represents the standard deviation of suspension travel, vehicle speed, or vehicle acceleration.

6. The method for road surface grade identification based on a deep learning model of suspension vibration response as described in claim 1, characterized in that, In step 3, the Transformer road surface classification model consists of an input projection layer, a position encoding layer, a Transformer encoder, and a classification head. The training process of the Transformer road surface classification model includes two stages. In the first stage, the Adam optimizer is used for full parameter training. In the second stage, freeze-fine-tuning is used to freeze all parameters of the input projection layer and the Transformer encoder, and only the classification head is fine-tuned.

7. The method for road surface grade identification based on a deep learning model of suspension vibration response as described in claim 6, characterized in that, The input projection layer includes a one-dimensional convolution, a one-dimensional batch normalization layer, a modified linear unit activation function, and a random deactivation structure. The position encoding layer uses sine and cosine position encoding. The Transformer encoder includes a pre-normalization structure, a multi-head attention head, and Gaussian error linear units. Residual connections and random deactivation structures are added after each sub-layer. The classification head includes a random deactivation structure, a fully connected layer, and a Softmax layer.

8. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the road surface grade identification method based on a deep learning model of suspension vibration response as described in claim 1.

9. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed, are used to implement the road surface grade identification method based on a deep learning model of suspension vibration response as described in claim 1.

10. A computer program product, characterized in that, The aforementioned computer program product includes computer-executable instructions that, when executed, implement the road surface grade identification method based on a deep learning model of suspension vibration response as described in claim 1.

Images