An online road roughness identification method and system

Abstract

The embodiment of the present invention provides an online road surface roughness identification method and system. The method comprises: obtaining a time series signal to be identified of a road surface; inputting the time series signal to be identified into a pre-trained road surface roughness grade classification model to obtain a road surface roughness classification result output by the road surface roughness grade classification model; wherein the road surface roughness grade classification model is obtained by training an LSTM network based on the training set data of the road surface and the grade labels obtained by classifying according to the road surface roughness standard grade. The embodiment of the present invention obtains a simple and small amount of time series signals, avoids complex data processing, and obtains the road surface roughness grade result quickly and in real time.

Description

Technical Field

[0001] The present invention relates to the field of vehicle engineering technology, and in particular to an online road surface roughness identification method and system. Background Art

[0002] Road roughness has a great impact on vehicle driving safety and ride comfort. Understanding road roughness information is of great significance for vehicle dynamics control, especially suspension control. In suspension control, the dual requirements of ride comfort and driving safety need to be met, but the same set of control parameters cannot achieve the optimal control effect under roads with different roughness levels. Therefore, road roughness level information can provide a direct and effective basis for parameter adjustment of the control algorithm.

[0003] The existing on-board road recognition strategies can be generally divided into two methods: direct and indirect. The direct recognition method is achieved by perceiving and measuring the road surface. It usually requires the help of dedicated sensors. The surface photometer is the simplest and most convenient measurement method, but it requires special customization and is costly, and cannot be commercialized on a large scale. Visible light imaging is a widely used method. The color, texture and edge features of road images are manually extracted, and road recognition is achieved by training a neural network classifier. There is also a method that uses a dataset containing dry road, wet road, snow road and mud road images, and uses a convolutional neural network model in deep learning for end-to-end recognition. However, this method requires a huge amount of computation, and its performance is heavily dependent on the amount of training data. The road surface point cloud model method uses an on-board laser radar to obtain three-dimensional point cloud data of the road surface, and uses morphological gradients to calculate the roughness of the point set. Real vehicle experiments in complex environments have confirmed that the algorithm can adaptively achieve road recognition. However, laser radar is expensive, requires a large amount of computation and storage space when building point clouds, and has poor portability. The ultrasonic method uses ultrasonic sensors to sense different road surfaces, extracts the waveform characteristics of the echo signal and inputs it into the classifier for identification. This method needs to consider the impact of vehicle speed and the dissipation of sound wave energy in the atmosphere on the identification accuracy.

[0004] The above direct identification methods all require the use of dedicated sensors, which not only increases the cost of road identification and reduces the scalability of sensors, but also makes it impossible to achieve real-time identification due to the large amount of algorithm calculation. At present, the indirect identification method based on the response of the suspension system has attracted more attention from researchers. Roads with different roughness levels have different bandwidths and amplitudes of excitation on the moving vehicle, which is reflected in the suspension as different acceleration responses. The indirect identification method aims to reversely infer the road grade from the suspension acceleration response.

[0005] Indirect identification methods can also be divided into model-based and data-driven methods. The Kalman filter method is used to estimate the roughness of the road from simulation and experimental data. This method requires a simple model, small amount of calculation, and can be extended to use other information to improve the estimation performance. The data-driven method can ignore the physical process of road excitation to suspension system response and use the algorithm model to autonomously learn the potential patterns in the data. The vertical acceleration signal of a single suspension sprung mass is subjected to a second-order wavelet transform to extract the statistical waveform features of the time domain and transform domain signals respectively, and then input into the adaptive neural network fuzzy inference system for road grade classification, but this method requires a lot of data preprocessing. The acceleration response of the unsprung mass is more suitable for road identification because it is easier to measure and can achieve higher accuracy. Summary of the invention

[0006] The embodiment of the present invention provides an online road surface roughness identification method and system to solve the defects in the prior art that a large amount of data is needed when measuring road surface roughness, and the measurement result is not accurate enough due to the susceptibility to environmental interference.

[0007] In a first aspect, an embodiment of the present invention provides an online road roughness identification method, comprising:

[0008] Acquire the time sequence signal to be identified from the road surface;

[0009] The time series signal to be identified is input into a pre-trained road surface roughness grade classification model to obtain a road surface roughness classification result output by the road surface roughness grade classification model; wherein the road surface roughness grade classification model is obtained by training an LSTM network based on training set data of the road surface and grade labels obtained by classifying according to standard road surface roughness grades.

[0010] Furthermore, the step of obtaining the time sequence signal to be identified of the road surface specifically includes:

[0011] Collect wheel vertical acceleration;

[0012] intercepting the acceleration according to a preset time length to obtain an acceleration segment;

[0013] The acceleration segments are normalized to obtain the time series signal to be identified.

[0014] Furthermore, the road surface roughness classification model is obtained by the following steps:

[0015] Acquire training set data of a road surface, and preprocess the training set data to obtain preprocessed training set data;

[0016] Grade the preprocessed training set data according to the road roughness standard grade to obtain the grade label;

[0017] Build an LSTM network as the initial model;

[0018] The preprocessed training set data marked with corresponding grade labels is input into the initial model for training, and the network and training parameters of the initial model are adjusted to obtain the road surface roughness grade classification model.

[0019] Furthermore, the step of acquiring the training set data of the road surface and preprocessing the training set data to obtain the preprocessed training set data specifically includes:

[0020] Acquire simulation training set data and experimental training set data;

[0021] The simulation training set data and the experimental training set data are intercepted according to a preset time length to obtain training set segments;

[0022] The training set fragments are normalized to obtain the preprocessed training set data.

[0023] Furthermore, the obtaining of simulation training set data and experimental training set data specifically includes:

[0024] The simulation data set generates a set of excitation signals of several road surface roughness levels by white noise filtering, and calculates a set of wheel vertical accelerations corresponding to the several road surface roughness levels by vehicle suspension transfer characteristics;

[0025] The test data set is obtained by installing an acceleration sensor at the wheel center and arranging a data acquisition system to obtain a set of wheel vertical accelerations on the road surfaces of the several roughness levels.

[0026] Furthermore, the construction of the LSTM network as an initial model specifically includes:

[0027] Using a time series input layer as the first layer of the initial model, the time series input layer is used to input each sampling point in the time series signal into the next layer separately;

[0028] The LSTM layer is used as the second layer of the initial model, and each LSTM unit is connected to the LSTM units corresponding to the previous and next moments, and the LSTM unit state and output information of the previous moment are transmitted;

[0029] A fully connected layer is used as the third layer of the initial model, and the fully connected layer performs linear operations on the network output results of the previous layer to obtain weight values ​​of different categories;

[0030] The Softmax layer is used as the fourth layer of the initial model, the probabilities of different categories are obtained through the Softmax layer, and the one with the largest probability is selected as the classification result.

[0031] Furthermore, adjusting the network and training parameters of the initial model includes adjusting hyperparameters, and the hyperparameters include the number of iterations and the learning rate.

[0032] In a second aspect, an embodiment of the present invention provides an online road roughness identification system, comprising:

[0033] An acquisition module, used for acquiring a signal to be identified on the road surface;

[0034] A processing module is used to input the signal to be identified into a pre-trained road surface roughness grade classification model to obtain a road surface roughness classification result output by the road surface roughness grade classification model; wherein the road surface roughness grade classification model is obtained by training an LSTM network based on a training set of road surface data and grade labels obtained by classifying according to standard road surface roughness grades.

[0035] In a third aspect, an embodiment of the present invention provides an electronic device, including:

[0036] A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any step of the online road surface roughness identification method is implemented.

[0037] In a fourth aspect, an embodiment of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the online road surface roughness identification methods.

[0038] The online road surface roughness identification method and system provided by the embodiment of the present invention obtain a simple small amount of reference signals, avoid complex data processing, and obtain the road surface roughness grade results quickly and in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 A flow chart of an online road surface roughness identification method provided by an embodiment of the present invention;

[0041] Figure 2 A flow chart of a road surface roughness level identification algorithm provided by an embodiment of the present invention;

[0042] Figure 3A sequential structure diagram of an LSTM repetitive module provided in an embodiment of the present invention;

[0043] Figure 4 LSTM module structure diagram provided for an embodiment of the present invention;

[0044] Figure 5 A full network architecture diagram from time series input to classification result output provided by an embodiment of the present invention;

[0045] Figure 6 A schematic diagram of a time domain signal of roughness of a Class E road surface at 20 m / s provided by an embodiment of the present invention;

[0046] Figure 7 A schematic diagram of the power spectrum density of the E-class road surface generated by the embodiment of the present invention and the upper and lower limits of the standard E-class road surface;

[0047] Figure 8 A diagram of a quarter vehicle suspension model provided by an embodiment of the present invention;

[0048] Fig. 9 A schematic diagram of random window selection acceleration provided by an embodiment of the present invention;

[0049] Fig.10 A schematic diagram of the vertical acceleration response results of wheels on five different grades of road surfaces provided by an embodiment of the present invention;

[0050] Fig.11 Schematic diagram of training and verification accuracy (top) and loss function value (bottom) provided in an embodiment of the present invention;

[0051] Fig.12 A schematic diagram of the time domain prediction results of a test set provided by an embodiment of the present invention;

[0052] Fig.13 A schematic diagram of a confusion matrix of test set results provided by an embodiment of the present invention;

[0053] Fig.14 A schematic diagram of test result evaluation indicators provided by an embodiment of the present invention;

[0054] Fig.15 A schematic diagram of recognition accuracy at different vehicle speeds provided by an embodiment of the present invention;

[0055] Fig.16 A structural diagram of an online road surface roughness identification system provided by an embodiment of the present invention;

[0056] Fig.17 This is a structural block diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Figure 1 A flow chart of an online road roughness identification method provided by an embodiment of the present invention is as follows: Figure 1 As shown, including:

[0059] S1, obtaining the time sequence signal to be identified on the road surface;

[0060] Specifically, in order to measure the roughness of the road surface, it is necessary to obtain some actual data to be identified. The relevant signal acquisition equipment is installed at the wheel center in advance, the vertical acceleration of the wheel is collected during the vehicle's driving process, the acceleration signal fragments are intercepted, and then normalized to obtain the time series signal to be identified.

[0061] S2, inputting the time series signal to be identified into a pre-trained road surface roughness grade classification model to obtain a road surface roughness classification result output by the road surface roughness grade classification model; wherein the road surface roughness grade classification model is obtained by training the LSTM network based on the training set data of the road surface and the grade labels obtained by classifying according to the standard grade of road surface roughness.

[0062] Specifically, after the time series signal to be identified is obtained, the time series signal is input into a pre-trained real-time network, namely, a road surface roughness grade classification model, so that the final road surface roughness grade classification result can be obtained quickly and accurately.

[0063] Here, the embodiment of the present invention uses an LSTM network for model training, such as Figure 2 As shown in the figure, it is divided into the network training stage and the actual use stage. In the network training stage, the LSTM network is first built, and the initial network model is built by building a multi-layer model structure with respective functions. When obtaining the training data of the model, since deep learning requires a large amount of data for network training, and the test data is not easy to contain a large number of complete working conditions, the acceleration signal of simulation and actual measurement can be used as the source of the training data set in the network training stage. In the actual use stage, the trained network can be applied to real-time road classification.

[0064] The trained network was tested and the results showed that the method of identifying the road roughness level through acceleration signals is effective and feasible, and the performance evaluation indicators such as accuracy are very high. Compared with the previous road roughness level identification algorithms, the advantages of this algorithm are:

[0065] (1) Only one signal is needed, such as the unsprung mass acceleration, the sensor layout is simple, and the hardware cost is low; (2) The input is a time-series acceleration signal, and there is no need to perform frequency domain transformation such as Fourier transform or feature extraction, and the original signal can be directly input;

[0066] (3) No need for complex filtering, high noise tolerance and good robustness;

[0067] (4) Effective identification can be performed at different vehicle speeds;

[0068] (5) When there is insufficient test data, a training set can be generated through white noise filtering, covering a wider range of working conditions, and measured data can be incorporated into the training set to further improve the recognition accuracy in actual use;

[0069] (6) When training the network, there is no need to perform a large amount of manual feature selection offline, which reduces the difficulty of development;

[0070] (7) Online algorithms require low computing power and do not require data preprocessing. The trained network can run directly at a very high speed, which ensures the real-time performance of the algorithm.

[0071] The algorithm for identifying the road roughness level based on LSTM through vehicle acceleration response in the embodiment of the present invention is effective and highly accurate, and can perform real-time road surface identification through measured acceleration in actual engineering applications. Road roughness level information can provide an important basis for adjusting parameters of adaptive algorithms for vehicle control such as semi-active suspension control, and has high practical value.

[0072] Based on the above embodiment, the step of obtaining the time sequence signal to be identified of the road surface specifically includes:

[0073] Collect wheel vertical acceleration;

[0074] intercepting the acceleration according to a preset time length to obtain an acceleration segment;

[0075] The acceleration segments are normalized to obtain the time series signal to be identified.

[0076] Specifically, the embodiment of the present invention installs an acceleration sensor and a data acquisition system at the center of the wheel, collects the vertical acceleration of the wheel during vehicle driving, and intercepts a 1-second acceleration signal segment and performs normalization processing to obtain a set of data to be identified.

[0077] Based on any of the above embodiments, the road surface roughness classification model is obtained by the following steps:

[0078] Acquire training set data of a road surface, and preprocess the training set data to obtain preprocessed training set data;

[0079] Grade the preprocessed training set data according to the road roughness standard grade to obtain the grade label;

[0080] Build an LSTM network as the initial model;

[0081] The preprocessed training set data marked with corresponding grade labels is input into the initial model for training, and the network and training parameters of the initial model are adjusted to obtain the road surface roughness grade classification model.

[0082] The step of obtaining the training set data of the road surface and preprocessing the training set data to obtain the preprocessed training set data specifically includes:

[0083] Acquire simulation training set data and experimental training set data;

[0084] The simulation training set data and the experimental training set data are intercepted according to a preset time length to obtain training set segments;

[0085] The training set fragments are normalized to obtain the preprocessed training set data.

[0086] The obtaining of simulation training set data and experimental training set data specifically includes:

[0087] The simulation training set generates a set of excitation signals of several road surface roughness levels by white noise filtering, and calculates a set of wheel vertical accelerations corresponding to the several road surface roughness levels by vehicle suspension transfer characteristics;

[0088] The test training set is obtained by installing an acceleration sensor at the wheel center and arranging a data acquisition system to obtain a set of vertical accelerations of the wheels on the roads of the several roughness levels, thereby obtaining the test training set data.

[0089] The construction of the LSTM network as the initial model specifically includes:

[0090] Using a time series input layer as the first layer of the initial model, the time series input layer is used to input each sampling point in the time series signal into the next layer separately;

[0091] The LSTM layer is used as the second layer of the initial model, and each LSTM unit is connected to the LSTM units corresponding to the previous and next moments, and the LSTM unit state and output information of the previous moment are transmitted;

[0092] A fully connected layer is used as the third layer of the initial model, and the fully connected layer performs linear operations on the network output results of the previous layer to obtain weight values ​​of different categories;

[0093] The Softmax layer is used as the fourth layer of the initial model, the probabilities of different categories are obtained through the Softmax layer, and the one with the largest probability is selected as the classification result.

[0094] Wherein, adjusting the network and training parameters of the initial model includes adjusting hyperparameters, and the hyperparameters include the number of iterations and the learning rate.

[0095] Specifically, the measurement data used in the embodiment of the present invention is an acceleration signal, which is a time series unstructured data, while the classic classification problem usually uses structured data as input. Structured data consists of different features, and converting time series signals into structured data requires manual feature extraction. The main problem with this method is that each input data needs to be preprocessed during training or testing, which will reduce the running speed in actual online applications.

[0096] The time series acceleration signal itself already contains many different features. Manual feature extraction cannot utilize all the information of the signal itself and will generate redundancy. Although the acceleration value at any moment cannot reflect the characteristics, the continuous signal fragments can. Therefore, in this case, the classic fully connected neural network with structured input is not applicable. However, the Long Short-Term Memory (LSTM) network is very good at extracting features from time series data with contextual relationships.

[0097] LSTM network has a similar chain structure to recurrent neural network (RNN), but each repeated unit of the former contains more complex layers. Figure 3 As shown, Figure 3 The input of each repeating unit at time t includes the sample time input data x <t>< / t> , the state c of the previous unit <t-1>< / t-1> and the output a of the previous unit <t-1>< / t-1> , the specific unit level is as follows Figure 4 shown.

[0098] exist Figure 4 In, W f , W i , W c , W o are the weights of the forget gate, input gate, and output gate. f , b i , b c , bo is the bias of these gates. Superposition means merging two arrays into one, σ(·) means sigma function, tanh(·) means hyperbolic tangent function, “*” means element-wise multiplication, and “+” means element-wise addition.

[0099] Forget gate Γ f The sigmoid function is used to filter out the part of the state of the previous unit that has not been discarded:

[0100] Γ f =σ(W f [a <t-1>< / t-1> ,x <t>< / t> ]+b f )

[0101] Update gate Γ u Determine what needs to be updated. Optional update data is provided:

[0102]

[0103] Γ u =σ(W i [a <t-1>< / t-1> ,x <t>< / t> ]+b i )

[0104] Updated cell status c <t>< / t> It consists of the selected state at the previous moment and the newly selected candidate variables:

[0105]

[0106] In the output link, tanh(C t ) t The values ​​of are rearranged in the range of (-1,1), and the output gate Γ o Determines the output part and finally obtains the output a of the module <t>< / t> :

[0107] Γ o =σ(W o [a <t-1>< / t-1> ,x <t>< / t> ]+b o )

[0108] a <t>< / t> =Γ o *tanh(c <t>< / t> )

[0109] The architecture of the entire network is as follows Figure 5As shown in Figure 2, the LSTM layer is followed by a fully connected layer (FC), where the number of neurons in the hidden layer is proportional to the number of categories N. c The elements in the fully connected layer are:

[0110]

[0111] In the formula, i represents the subscript of each neuron in the fully connected layer, and w i is the weight of the neuron, b i is the bias. Since this problem is a classification problem, a softmax layer is connected at the end to extract the maximum result in the fully connected layer. The probability y of each category i The calculation method is:

[0112]

[0113] Where e is the natural index. Represents the classification result with the highest probability. This completes the process from the input of the time series acceleration signal to the output of the classification result, and the calculation of each step of the forward propagation has been completed.

[0114] The above network establishment process is forward propagation, and the final loss function L i Defined as the cross entropy form:

[0115]

[0116] The overall objective function for all samples is:

[0117]

[0118] In the formula, Θ refers to the parameters that need to be iterated in the network, such as weight W and bias b, and m is the number of samples. To solve this optimization problem, it is necessary to iterate the parameters by backpropagation, repeat the backpropagation process, and use the calculated output and state variables of the current LSTM module to pass to the previous moment.

[0119] The algorithm used to minimize the objective function is the Adam algorithm (Adaptive optimization algorithm), which is one of the more efficient deep learning algorithms. It combines the advantages of batch gradient descent, momentum method, RMSProp and other algorithms, and can adaptively change the learning rate to quickly find the optimal solution. The variable update iterative algorithm is as follows:

[0120]

[0121]

[0122]

[0123] where v dΘ is the momentum variable, S dΘ is the learning rate adjustment variable, β1 is the gradient decay factor, and β2 is the square gradient decay factor. After several iterations, all variables that need to be updated in the network will be iterated to the value that makes the objective function as small as possible.

[0124] The following is a specific embodiment of an algorithm for identifying road roughness levels from time series acceleration using an LSTM network, wherein the data set used is generated by simulation.

[0125] Considering the various working conditions that may occur in reality, it is difficult to obtain a large amount of training data with rich working conditions using the actual vehicle measurement method. Therefore, a large number of different road roughness signals are generated using filtered white noise, and the wheel vertical acceleration response is calculated through the transfer characteristics of the suspension model. The network is trained based on a large amount of training data, and then the actual measured data is added to the training set later to make the network more suitable for the actual situation. The algorithm flow is as follows: Figure 2 shown.

[0126] The road surface roughness elevation information is usually considered to obey a normal distribution with a mean of zero. In order to describe the characteristics of this random process, the power spectrum density (PSD) is usually used. According to ISO 8601, the power spectrum density of road surface roughness is as follows:

[0127]

[0128] Where: n is the spatial frequency, m is -1 , n0 is the reference spatial frequency, the value is 0.1m -1 , G q (n0) is the road surface power spectral density at the reference spatial frequency, m 3 , also known as the road roughness coefficient, W is the frequency index, which determines the frequency distribution of the road power spectrum density, usually 2. q (n0) and the corresponding variance σ q It varies with different levels of road roughness. The power spectral density is described in the frequency domain as

[0129]

[0130] Considering that road roughness is a limited bandwidth noise, the time domain road roughness with the required road power spectrum density can be generated by first-order filtering of specific white noise:

[0131] G q (ω)=|H(jω)| 2 S ω

[0132] Where S ω is the white noise power spectral density, here it is taken as 1, and the transfer function represented by frequency is:

[0133]

[0134] The state equation of the transfer characteristic is:

[0135]

[0136] Where: q(t) is the road roughness, m, n 00 is the lower cutoff frequency of the space, the value is 0.011m -1 , w(t) is a time domain white noise signal with a mean of zero and a power spectrum density of 1. The E-level road roughness result generated by white noise at a speed of 20m / s is as follows Figure 6 shown.

[0137] The generated road is compared with the standard power spectrum density of Class E road surface. The lower boundary is half of the standard power spectrum density of Class E, while the upper boundary is twice. The result is as follows: Figure 7 As shown in the figure, we can see that the spectrum is in the E-level area. Therefore, the road surface roughness generated by the filtered white noise can be close to the actual collected result and can be used for the training of road surface roughness level recognition.

[0138] In order to demonstrate the transfer characteristics of tires and suspension, a linearized two-degree-of-freedom quarter-vehicle model was selected, such as Figure 8 The figure shows the sprung mass ms, unsprung mass mu, spring stiffness ks, variable damping coefficient cs, tire stiffness kt, and the displacements of the vehicle body, tire, and road surface are represented by xs, xu, and xr respectively.

[0139] The dynamic equations of the quarter vehicle system are:

[0140]

[0141]

[0142] The transfer function from road displacement to unsprung mass acceleration is:

[0143]

[0144]

[0145] The unsprung mass acceleration is obtained from the transfer characteristics of the road surface to the unsprung mass acceleration:

[0146]

[0147] Thus, the vehicle acceleration response that can be used to identify the road surface roughness level is obtained, and then the network training for deep learning can be carried out. Generate vehicle acceleration response signals for five roughness levels including B, C, D, E, and F road surfaces. In order to make the data closer to the data measured by the actual acceleration sensor, a certain power of white noise is superimposed on the acceleration signal to improve the robustness of the algorithm identification. In order to make the generated signal more random and universal and avoid the superposition of certain systematic deviations, a different random number seed is used each time to generate white noise, and a signal ten times the required duration is generated, and then random fragment sampling is performed, which can avoid the existence of certain fixed features in the acceleration signal caused by the generation method, and this feature is not linked to the road surface level. Each sample used in the training set has 1000 acceleration sampling points, which are continuously selected from the 10,000 generated sampling points. Here, the training set can be obtained entirely by experiments, or by both simulation and experiments, and the embodiment of the present invention is not limited. The fragments are selected as follows Fig. 9 shown.

[0148] The training set has 1000 sets of samples, consisting of 5 different levels of road surface stimulus results, each with 200 samples, and each sample has 1000 continuous sampling points. The sampling results of different levels are as follows: Fig.10 shown.

[0149] Considering the amplitude range of different levels of acceleration, in order to make the acceleration signal have a similar distribution, the signal is normalized:

[0150]

[0151] Where μ and σ are the mean and variance of acceleration under a road excitation, respectively.

[0152] The validation set contains 250 samples, has the same distribution as the training set, and is also normalized. After completing the data preparation and network construction, deep learning training can be performed. The hyperparameters used for training are shown in Table 1.

[0153] Table 1

[0154]

[0155] The 1000 samples in the training set are divided into 10 small batches, each with 100 samples. Each time, the entire data set is used for forward and backward propagation to complete 10 iterations, traversing the data set 300 times, for a total of 3000 iterations. Each time the complete data set is traversed, the order of the entire training set samples will be reshuffled, and the learning rate will decay to 0.99 times the previous learning rate. During training, the accuracy of the small batch test set is calculated for each iteration, and the accuracy of the validation set is calculated after traversing the entire data set 5 times to check whether the network is overfitting the training set. If overfitting occurs, the training should be terminated to adjust the sample size or hyperparameter values.

[0156] The training process is as follows Fig.11 As shown in the figure. The training accuracy shows an upward trend in the fluctuation. Since the training samples only contain a time domain signal and no structured data, higher requirements are placed on the network for classification training, requiring a deeper network and more samples. Transferring the computational burden of online processing to the offline training process can facilitate the real-time and rapid performance of the online algorithm. The algorithm uses small batch gradient descent, so there are fluctuations in the process of accuracy increase. The training and validation accuracies are close to 100%, and the loss functions are close to 0. There is no underfitting in the training. The accuracy of the validation set has the same trend as the training set results, which shows that the network has not overfitted on the training set. When the training process reaches 1200 iterations, the accuracy and loss function have obvious fluctuations, which may be due to the underfitting of the small batch samples after the training set is re-shuffled, resulting in a decrease in accuracy.

[0157] The test set has the same distribution and sample size as the validation set. The accuracy of the training set, validation set, and test set are 98%, 95.6%, and 97.2%, respectively. The results show that the trained network has high accuracy for different data sets.

[0158] The algorithm uses the time series acceleration signal for recognition and can be used for online real-time operation. After that, the network's recognition ability for continuous signals is tested. The road surface is imported into the network in the time series order of B, E, C, F and D for testing. The results are as follows: Fig.12 As shown in the figure, it shows that the road roughness level identification algorithm based on LSTM network can process time series acceleration signals in real time with high accuracy.

[0159] The confusion matrix of the test set results is as follows Fig.13 As shown in the figure, most of the predicted categories are consistent with the actual categories. The incorrectly predicted road roughness level is within ±1 of the actual level. Because the characteristics of the F-level road surface are more significant, it is easier for the network to capture the features and classify them, so the accuracy is relatively higher than other categories.

[0160] In addition to accuracy, three common indicators used to evaluate deep learning networks are precision, recall, and F1 score. The three are defined as follows:

[0161]

[0162]

[0163]

[0164] In the formula, TP means that the actual value is true and the prediction is true, FP means that the actual value is false and the prediction is true, FN means that the actual value is true but the prediction is false, and i represents the index of different categories. The results of the test set are as follows Fig.14 As shown in the figure, the precision and recall rates in the same category are both high and close, which makes it obtain a relatively high F1 value. The precision and recall rates of different categories are both high, indicating that they have consistent good effects on different levels of road surfaces. The average F1 value is 0.97, which proves that the network can achieve accurate classification for different levels of road roughness.

[0165] Speed ​​robustness is an important factor that needs to be considered when using the road roughness identification algorithm in practice. In previous methods, the accuracy at different vehicle speeds is usually not mentioned during verification. However, the vehicle speed varies greatly during driving, so this problem cannot be ignored. The training set samples are acceleration signals obtained at a vehicle speed of 20m / s. The normalization process constrains the amplitude range of different accelerations so that the network can adapt to signals in different speed ranges. Another test set selected accelerations of different levels of road excitation at speeds ranging from 5m / s to 50m / s, and was verified through the trained network. The results show that the network is effective over a wide speed range and can guarantee an accuracy rate of more than 95%. The results are shown in the figure below. Fig.15 shown.

[0166] Fig.16 A structural diagram of an online road roughness identification system provided by an embodiment of the present invention is shown in FIG. Fig.16 As shown, it includes: an acquisition module 1601 and a processing module 1602; wherein:

[0167] The acquisition module 1601 is used to acquire the time series signal to be identified of the road surface; the processing module 1602 is used to input the time series signal to be identified into a pre-trained road surface roughness grade classification model to obtain the road surface roughness classification result output by the road surface roughness grade classification model; wherein the road surface roughness grade classification model is based on the training set data of the road surface and the grade labels obtained by classifying according to the standard grade of road surface roughness to train the LSTM network.

[0168] The system provided in the embodiment of the present invention is used to execute the corresponding method mentioned above. Its specific implementation method is consistent with the implementation method of the method, and the algorithm flow involved is the same as the corresponding method algorithm flow, which will not be repeated here.

[0169] Fig.17 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.17 As shown, the electronic device may include: a processor 1710, a communication interface 1720, a memory 1730 and a communication bus 1740, wherein the processor 1710, the communication interface 1720 and the memory 1730 communicate with each other through the communication bus 1740. The processor 1710 may call the logic instructions in the memory 1730 to execute the following method: obtaining a time series signal to be identified of a road surface; inputting the time series signal to be identified into a pre-trained road surface roughness grade classification model to obtain a road surface roughness classification result output by the road surface roughness grade classification model; wherein the road surface roughness grade classification model is obtained by training the LSTM network based on the training set data of the road surface and the grade labels obtained by classifying according to the road surface roughness standard grade.

[0170] In addition, the logic instructions in the above-mentioned memory 1730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0171] On the other hand, an embodiment of the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it is implemented to execute the transmission methods provided in the above-mentioned embodiments, for example, including: obtaining a time series signal to be identified of a road surface; inputting the time series signal to be identified into a pre-trained road surface roughness level classification model to obtain a road surface roughness classification result output by the road surface roughness level classification model; wherein the road surface roughness level classification model is obtained by training an LSTM network based on training set data of the road surface and grade labels obtained by classification according to standard road surface roughness grades.

[0172] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0173] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An online road roughness identification method, characterized in that: include: Acquire the time sequence signal to be identified from the road surface; The time series signal to be identified does not need to be transformed in the frequency domain, does not need to extract features, and does not need to be filtered; Inputting the time series signal to be identified into a pre-trained road surface roughness grade classification model to obtain a road surface roughness classification result output by the road surface roughness grade classification model; wherein the road surface roughness grade classification model is obtained by training an LSTM network based on a training set of road surface data and grade labels obtained by classifying according to standard grades of road surface roughness; The road surface roughness classification model is obtained by the following steps: Acquiring training set data of a road surface, and preprocessing the training set data to obtain preprocessed training set data; Grade the preprocessed training set data according to the road roughness standard grade to obtain the grade label; The training set builds an LSTM network as the initial model; Inputting the preprocessed training set data marked with corresponding grade labels into the initial model for training, and adjusting the network and training parameters of the initial model to obtain the road surface roughness grade classification model; The step of acquiring the training set data of the road surface and preprocessing the training set data to obtain the preprocessed training set data specifically includes: Acquire simulation training set data and experimental training set data; The simulation training set data and the test training set data are intercepted according to a preset time length to obtain training set segments; the simulation training set data is an acceleration signal obtained at a vehicle speed of 5 m / s; the test training set data is an acceleration signal obtained at a vehicle speed of 5 m / s; Normalizing the training set fragments to obtain the preprocessed training set data; The obtaining of simulation training set data and experimental training set data specifically includes: The simulation training set generates a set of excitation signals of several road surface roughness levels through white noise filtering, and calculates a set of wheel vertical accelerations corresponding to the several road surface roughness levels through vehicle suspension transfer characteristics; the vehicle suspension transfer characteristics are displayed using a linearized two-degree-of-freedom quarter vehicle model; The test training set obtains a set of wheel vertical accelerations on the road surfaces of the several roughness levels by installing acceleration sensors at the wheel centers and arranging a data acquisition system; The step of obtaining the time sequence signal to be identified of the road surface specifically includes: By installing an acceleration sensor and a data acquisition system at the wheel center, the vertical acceleration of the wheel is collected during the vehicle's driving process; intercepting the acceleration according to a preset time length to obtain an acceleration segment; The acceleration segments are normalized to obtain the time series signal to be identified; the time series signal to be identified is unstructured data.

2. The online road surface roughness identification method according to claim 1, characterized in that: The construction of the LSTM network as the initial model specifically includes: Using a time series input layer as the first layer of the initial model, the time series input layer is used to input each sampling point in the time series signal into the next layer separately; The LSTM layer is used as the second layer of the initial model, and each LSTM unit is connected to the LSTM units corresponding to the previous and next moments, and the LSTM unit state and output information of the previous moment are transmitted; A fully connected layer is used as the third layer of the initial model, and the fully connected layer performs linear operations on the network output results of the previous layer to obtain weight values ​​of different categories; The Softmax layer is used as the fourth layer of the initial model, the probabilities of different categories are obtained through the Softmax layer, and the one with the largest probability is selected as the classification result.

3. The online road surface roughness identification method according to claim 1, characterized in that: The adjusting of the network and training parameters of the initial model includes adjusting hyperparameters, wherein the hyperparameters include the number of iterations and the learning rate.

4. An online road roughness identification system, characterized in that: include: An acquisition module, used for acquiring a time sequence signal to be identified from a road surface; The time series signal to be identified does not need to be transformed in the frequency domain, does not need to extract features, and does not need to be filtered; A processing module, used for inputting the time series signal to be identified into a pre-trained road surface roughness grade classification model to obtain a road surface roughness classification result output by the road surface roughness grade classification model; wherein the road surface roughness grade classification model is obtained by training an LSTM network based on a training set of road surface data and grade labels obtained by classifying according to standard grades of road surface roughness; The road surface roughness classification model is obtained by the following steps: Acquiring training set data of a road surface, and preprocessing the training set data to obtain preprocessed training set data; Grade the preprocessed training set data according to the road roughness standard grade to obtain the grade label; The training set builds an LSTM network as the initial model; Inputting the preprocessed training set data marked with corresponding grade labels into the initial model for training, and adjusting the network and training parameters of the initial model to obtain the road surface roughness grade classification model; The step of acquiring the training set data of the road surface and preprocessing the training set data to obtain the preprocessed training set data specifically includes: Acquire simulation training set data and experimental training set data; The simulation training set data and the test training set data are intercepted according to a preset time length to obtain training set segments; the simulation training set data is an acceleration signal obtained at a vehicle speed of 5 m / s; the test training set data is an acceleration signal obtained at a vehicle speed of 5 m / s; Normalizing the training set fragments to obtain the preprocessed training set data; The obtaining of simulation training set data and experimental training set data specifically includes: The simulation training set generates a set of excitation signals of several road surface roughness levels through white noise filtering, and calculates a set of wheel vertical accelerations corresponding to the several road surface roughness levels through vehicle suspension transfer characteristics; the vehicle suspension transfer characteristics are displayed using a linearized two-degree-of-freedom quarter vehicle model; The test training set obtains a set of wheel vertical accelerations on the road surfaces of the several roughness levels by installing acceleration sensors at the wheel centers and arranging a data acquisition system; The step of obtaining the time sequence signal to be identified of the road surface specifically includes: By installing an acceleration sensor and a data acquisition system at the wheel center, the vertical acceleration of the wheel is collected during the vehicle's driving process; intercepting the acceleration according to a preset time length to obtain an acceleration segment; The acceleration segments are normalized to obtain the time series signal to be identified; the time series signal to be identified is unstructured data.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the online road surface roughness identification method according to any one of claims 1 to 3 are implemented.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the online road surface roughness identification method as claimed in any one of claims 1 to 3 are implemented.

Images