Loading vehicle driving deviation correction method and system based on intelligent classification of road surface perception data

Abstract

The present invention discloses a method and system for correcting the deviation of a loading vehicle based on intelligent classification of road surface perception data, the method comprising: obtaining data to be classified from multiple high-frequency sensors on a full-scale road surface; constructing an adaptive threshold according to the data characteristics of different high-frequency sensors on the full-scale road surface, and extracting sensor data in the loading process; visually converting the obtained multiple pressure perception data fragments to obtain corresponding image data to be classified; constructing a DCNN6 convolutional neural network and training the convolutional neural network; using the trained convolutional neural network model to obtain the classification results of the image data to be classified; and using the classification results to prompt the driving trajectory of the vehicle. The present invention can effectively solve the problem of automatic classification of dynamic high-frequency sensor data in a full-scale loop, provide strong data support for the force analysis of the entire road area and the grasp of the evolution law of rutting, and has fast classification speed and high accuracy, saving manpower and material resources.

Description

Technical Field

[0001] The present invention belongs to the technical field of road engineering, and in particular relates to a method and system for correcting the deviation of a loader based on intelligent classification of road surface perception data. Background Art

[0002] At present, as the use time of highways gradually increases, the road surface and its internal structure will be damaged to varying degrees, the life of the road is shortened, and the safety of vehicle driving is also threatened to a certain extent. In 2015, China's full-scale pavement test loop engineering system was officially completed to carry out the study of the evolution law of multi-service behavior of the full-scale pavement structure throughout its life cycle. The test loop is referred to as RIOHTRACK. Each layer of the pavement structure is equipped with a large number of sensors, which mainly measure various properties of asphalt pavement. In order to accurately control the distribution of the wheel track of the loading vehicle in the area above the pre-set sensors through data, so as to more accurately analyze the different pavement stress conditions caused by the load, and thus conduct full-area stress analysis and rutting evolution law research, it is necessary to conduct intelligent classification research on such sensor data in a timely manner, quickly detect vehicle deviations from the preset track, and correct the driver's driving deviation.

[0003] Early high-frequency sensor data classification and anomaly detection were done manually, and data screening efficiency was extremely low and easily affected by subjective factors, resulting in an inability to quickly and effectively analyze data objectively and accurately, which led to deviations in subsequent pavement performance and stress analysis. In terms of automated data detection, existing methods have problems such as complex processes, long time consumption, and lack of pertinence.

[0004] In addition, the existing high-frequency sensor data classification and anomaly detection methods usually directly process data. Since a complete loading situation probably includes at least 8,000 data points, traditional methods cannot identify abnormal situations in a complete loading process. The current data anomaly detection method cannot classify normal and abnormal situations that require a large amount of data to jointly determine the data status. Therefore, it is urgent to propose a targeted automatic detection method for high-frequency loop sensor data. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a method and system for correcting the deviation of a loader based on intelligent classification of road surface perception data. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] One aspect of the present invention provides a loader driving deviation correction system based on intelligent classification of road surface perception data, comprising:

[0007] Obtain raw data from multiple high-frequency sensors on full-scale road surfaces;

[0008] An adaptive threshold is constructed based on the data characteristics of different high-frequency sensors on the full-scale road surface, and the sensor data in the loading process is extracted to obtain multiple pressure sensing data fragments;

[0009] Performing visual conversion on the multiple pressure sensing data segments to obtain corresponding image data to be classified;

[0010] Constructing a convolutional neural network, wherein the convolutional neural network includes a feature extraction unit and a feature mapping unit, wherein the feature extraction unit is used to extract various features of the input image data, including six convolutional layers and one pooling layer connected in sequence, the six convolutional layers respectively use a relu function to extract features from the image data to obtain a feature image, and the pooling layer is used to compress and transmit the feature image; the feature mapping unit is used to obtain a classification result of the feature image by using a mapping relationship between the feature image and image category transition, and the feature mapping unit includes a flat layer and two fully connected layers connected in sequence;

[0011] Processing the image data to be classified using a trained convolutional neural network model to obtain a classification result of the image data to be classified;

[0012] The classification result is used to obtain the time and space position corresponding to the abnormal data, and to provide a prompt for the driving trajectory of the vehicle.

[0013] In one embodiment of the present invention, an adaptive threshold is constructed according to different high-frequency sensor data features of a full-scale road surface, and sensor data in the loading process is extracted, including:

[0014] Storing relevant parameters of the plurality of high-frequency sensors, wherein the relevant parameters include a data source format and parameter information of the sensors;

[0015] Adaptively set the data threshold required for data extraction according to the relevant parameters of the current high-frequency sensor;

[0016] The raw data to be classified are extracted by using the relevant parameters and the data threshold, and the pressure sensing data in the loading process is extracted.

[0017] In one embodiment of the present invention, the data threshold required for data extraction is adaptively set according to relevant parameters of the current high-frequency sensor, including:

[0018] Extract descriptive indicators of different high-frequency sensor data, including at least mean, median, mode, maximum, minimum, 25th percentile, 75th percentile, range, quartile range, and dispersion coefficient;

[0019] The data distribution characteristics of different high-frequency sensors are comprehensively analyzed through the descriptive indicators, and the data thresholds corresponding to different sensors are adaptively adjusted according to the data distribution characteristics.

[0020] In one embodiment of the present invention, the data threshold includes an upper extreme value threshold, a lower extreme value threshold, a maximum interval extreme difference threshold between the upper and lower extreme value thresholds, and a minimum interval extreme difference threshold between the upper and lower extreme value thresholds.

[0021] In one embodiment of the present invention, a convolutional neural network is constructed, including:

[0022] Constructing a convolutional neural network for classifying the image data;

[0023] Acquire a large amount of raw data from multiple high-frequency sensors on a full-scale road surface, extract and visualize the pressure sensing data in the loading process in the raw data, and obtain a large amount of image data;

[0024] Annotating the large amount of image data, adding normal labels and abnormal labels, and obtaining a large amount of image data with normal labels or abnormal labels to form the training data set;

[0025] The convolutional neural network is trained using the training data set to obtain a convolutional neural network model that has good classification capabilities for the data.

[0026] In one embodiment of the present invention, the large amount of image data is labeled, and normal labels and abnormal labels are added, including:

[0027] The constructed annotation network is pre-trained using multiple manually annotated high-frequency sensor data images, and the high-frequency sensor data image features are gradually learned through multiple layers of intermediate convolutional layers and the network parameters are initialized; the annotation error is back-propagated to optimize the node weights to minimize the network error; the weights of each node are updated, and the node weights of each layer are updated in turn using back-propagation to optimize the parameters of the annotation network and complete the training of the annotation network;

[0028] The large amount of image data to be labeled is input into a trained network model with image classification capability to complete semi-automatic labeling of the image.

[0029] Another aspect of the present invention provides a loader driving correction system based on intelligent classification of road surface perception data, including a data acquisition module, a data extraction module, a data visualization module, a data classification module and a prompt module, wherein:

[0030] The data acquisition module is connected to a plurality of high-frequency sensors located on a full-scale road surface, and is used to obtain raw data of the plurality of high-frequency sensors;

[0031] The data extraction module is used to construct an adaptive threshold according to different high-frequency sensor data characteristics of the full-scale road surface, extract sensor data in the loading process, and obtain multiple pressure perception data fragments;

[0032] The data visualization module is used to perform visualization conversion on the multiple pressure sensing data segments to obtain corresponding image data to be classified;

[0033] The data classification module is provided with a convolutional neural network, and the convolutional neural network is used to classify the image data to be classified after training and output the classification result.

[0034] In one embodiment of the present invention, the data extraction module includes a storage unit, a threshold setting unit and a data processing unit, wherein:

[0035] The storage unit is used to store relevant parameters of the multiple high-frequency sensors, wherein the relevant parameters include data source format, sensor information, and associated data of the sensors;

[0036] The threshold setting unit is used to adaptively set the data threshold required for data extraction according to the relevant parameters of the current high-frequency sensor, wherein the data threshold includes the upper extreme value threshold, the lower extreme value threshold, the maximum interval extreme difference threshold between the upper and lower extreme value thresholds, and the minimum interval extreme difference threshold between the upper and lower extreme value thresholds;

[0037] The data processing unit is used to extract the raw data to be classified by using the relevant parameters and the data threshold, and extract the pressure sensing data in the loading process;

[0038] The prompt module is used to obtain the time and space position corresponding to the abnormal data using the classification result, and to prompt the driving trajectory of the vehicle.

[0039] In one embodiment of the present invention, the convolutional neural network includes a feature extraction unit and a feature mapping unit, wherein:

[0040] The feature extraction unit is used to extract various features of the input image data, wherein the feature extraction unit includes six convolutional layers and one pooling layer connected in sequence, the six convolutional layers respectively use the relu function to extract features from the image data to obtain a feature image, and the pooling layer is used to compress and transmit the feature image;

[0041] The feature mapping unit is used to obtain the classification result of the feature image by utilizing the mapping relationship between the feature image and the image category transition. The feature mapping unit includes a flat layer and two fully connected layers connected in sequence.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The present invention proposes a method for correcting the deviation of a loader based on intelligent classification of road surface perception data. The method directly uses time domain data to visualize image data, which simplifies the complex processing process of other existing methods that require time-frequency domain conversion of data or two-dimensional arrangement of one-dimensional time series data.

[0044] 2. The present invention extracts effective information fragments in the time domain of high-frequency sensors through adaptive thresholds and related rule settings, discards a large amount of invalid data that is not loaded, saves nearly 80% of high-frequency data storage space, and effectively improves the overall processing speed of the method.

[0045] 3. According to the characteristics of visualized data, the present invention designs a lightweight deep convolutional neural network model. Compared with the commonly used classification network, the network parameters of this network are greatly reduced, the convergence speed is fast, and it is not easy to fall into overfitting. It realizes the rapid classification and anomaly detection of road pressure high-frequency sensor monitoring data, and provides strong data support for the stress analysis of the entire road area and the grasp of the evolution law of rutting.

[0046] 4. The present invention processes a large amount of raw data from multiple high-frequency sensors on a full-scale road surface into images, which can make the data distribution and classification characteristics more obvious, intuitive and easy to distinguish. A lightweight classification network is constructed in the later stage to realize the recognition of the classification network.

[0047] 5. Through intelligent classification of high-frequency road sensor monitoring data, the deviation correction of the loading vehicle's driving trajectory can be completed, which can assist in controlling the trajectory of the loading vehicle. It is highly targeted and more sensitive and accurate than the trajectory correction based on the loading vehicle's GPS system. At the same time, it is of great significance to improve the loading effectiveness in rutting research.

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of a method for correcting the deviation of a loading vehicle based on intelligent classification of road surface perception data provided by an embodiment of the present invention;

[0050] Figure 2 is image data obtained by using the visualization method of an embodiment of the present invention;

[0051] Figure 3 It is a structural schematic diagram of a DCNN6 convolutional neural network provided by an embodiment of the present invention;

[0052] Figure 4 It is a parameter diagram of a DCNN6 convolutional neural network provided by an embodiment of the present invention;

[0053] Figure 5 It is a process schematic diagram of a semi-automatic annotation method for high-frequency sensor data images provided by an embodiment of the present invention;

[0054] Figure 6 is a structural diagram of a labeling network provided by an embodiment of the present invention;

[0055] Figure 7 It is a schematic diagram of the steps of providing vehicle driving trajectory prompts using the splitting results of an embodiment of the present invention;

[0056] Figure 8 This is a flow chart of a verification experiment of the intelligent classification method of an embodiment of the present invention;

[0057] Fig. 9 It is a partial convolution feature map obtained by using the DCNN6 convolutional neural network of an embodiment of the present invention;

[0058] Fig.10 is a curve diagram of the classification accuracy change of the DCNN6 convolutional neural network in the training process of an embodiment of the present invention;

[0059] Fig.11 It is a loss change curve diagram of the DCNN6 convolutional neural network in the training process of an embodiment of the present invention;

[0060] Fig.12 is a confusion matrix of the classification result using the DCNN6 convolutional neural network of an embodiment of the present invention;

[0061] Fig.13 This is a curve chart of the classification accuracy change of the classification model Resnet50 during the training process;

[0062] Fig.14 This is the loss change curve of the classification model Resnet50 during the training process;

[0063] Fig.15 It is the confusion matrix of the classification results using the classification model Resnet50;

[0064] Fig.16 It is a schematic diagram of the structure of the Ghost module in the existing GhostNet model;

[0065] Fig.17 It is a curve chart of the classification accuracy change of the GhostNet model during the training process;

[0066] Fig.18 It is the loss change curve of the GhostNet model during training;

[0067] Fig.19It is the confusion matrix of the classification results using the GhostNet model. DETAILED DESCRIPTION

[0068] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description of a method and system for correcting the driving of a loader based on intelligent classification of road surface perception data proposed by the present invention, in combination with the accompanying drawings and specific implementation methods.

[0069] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0070] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants are intended to cover non-exclusive inclusion, so that an article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the statement "including one..." do not exclude the existence of other identical elements in the article or device including the elements.

[0071] Embodiment 1

[0072] See also Figure 1 , Figure 1 This is a flow chart of a method for correcting the deviation of a loader based on intelligent classification of road surface perception data provided by an embodiment of the present invention. The intelligent classification method includes:

[0073] S1: Obtain raw data from multiple high-frequency sensors on a full-scale road surface.

[0074] During the full-scale road test, a large number of high-frequency sensors are arranged on each layer of the road structure, so that the road stress caused by the distribution of the vehicle wheel track can be accurately grasped through data. This embodiment can obtain the original data when the vehicle passes by through the data acquisition equipment connected to each high-frequency sensor.

[0075] S2: An adaptive threshold is constructed based on the data characteristics of different high-frequency sensors on the full-scale road surface, and the sensor data in the loading process is extracted to obtain multiple pressure perception data fragments.

[0076] It should be noted that the high-frequency sensor can collect thousands of data within 1s. The raw data sampling frequency is high and contains a large amount of noise information in the non-loading time period. After obtaining the raw data, this embodiment first extracts the pressure sensing data in the raw data during the loading process. By analyzing the sensor technical parameters and the descriptive characteristics of the sensor acquisition data and other information, the extraction of effective loading data can be automatically completed. The technical parameters of the sensor mainly include: sensor type, standard range, resolution, accuracy, sampling interval and sampling frequency, etc.

[0077] Specifically, step S2 includes:

[0078] S21: storing relevant parameters of the plurality of high-frequency sensors, wherein the relevant parameters include a data source format and parameter information of the sensors.

[0079] In this embodiment, the relevant parameters also include data type, acquisition device status, signal feature points, etc. In addition, this embodiment can obtain the road section to which the data belongs, hourly data volume, and daily data volume in real time; and can set different benchmark data according to different road section data.

[0080] S22: Adaptively set the data threshold and extraction time required for data extraction according to the relevant parameters of the current high-frequency sensor. The data threshold includes four types of thresholds: the upper extreme value threshold, the lower extreme value threshold, the maximum interval extreme value threshold between the upper and lower extreme value thresholds, and the minimum interval extreme value threshold between the upper and lower extreme value thresholds. For example, by setting the maximum interval extreme value threshold between the upper and lower extreme value thresholds, that is, the maximum interval between the highest point and the lowest point of the data, abnormal data with large changes or strong interference can be removed to a certain extent; by setting the minimum interval threshold between the maximum and minimum values ​​of the data, that is, the minimum extreme value interval, valid data after vehicle loading can be obtained, and a large amount of invalid data without loading can be filtered out. The setting of the threshold is based on the specific information parameters of the sensor. During the same operation, the extraction time is fixed and can be flexibly set according to actual conditions.

[0081] In this embodiment, a python program is written to extract descriptive indicators of different sensor data, including: mean, median, mode, maximum, minimum, 25th percentile, 75th percentile, range, quartile range, dispersion coefficient, etc. Through the above indicators, the data distribution characteristics of different sensors are comprehensively analyzed, and then the data thresholds corresponding to different sensors are adaptively adjusted according to the characteristics. Finally, the sensor data of different types of high-frequency sensors (earth pressure gauge, dynamic pressure gauge, static pressure gauge, asphalt strain gauge, vertical strain gauge, etc.) in the original data during the loading process are extracted through the adaptive threshold.

[0082] Through the above threshold setting process, the program can automatically analyze the data distribution characteristics of multiple dynamic high-frequency sensors and automatically complete the adjustment of the adaptive threshold. There is no need to rely on a lot of manual experience and subjective consciousness to interfere in the threshold adjustment process every time. After the system automatically analyzes the data distribution characteristics and obtains the adaptive threshold, it only needs to combine a small amount of manual intervention (people look at the system analysis results and the different thresholds obtained, and manually confirm multiple groups of thresholds automatically analyzed and set by the system at one time without too much deviation) to complete the confirmation of multiple groups of sensor thresholds at one time.

[0083] Furthermore, multiple sets of data are input at the same time, and the system outputs the threshold setting results after analysis. With a small amount of manual intervention, the confirmation of multiple sets of sensor thresholds can be completed at one time. If there is no manual confirmation, the system will use the adaptive threshold automatically set by the system to extract the corresponding sensor data by default when extracting data.

[0084] S23: extracting the raw data to be classified by using the relevant parameters and the data threshold, and extracting the pressure perception data in the loading process.

[0085] Specifically, if the difference between the maximum and minimum values ​​of the data in the current extraction time period exceeds the range of the preset threshold, the data segment is considered to be abnormal data or invalid data, and the extraction is abandoned. If the difference between the maximum and minimum values ​​of the data in the current extraction time period is within the range of the preset threshold, the data segment is extracted and saved. According to the extraction principle, the raw data obtained from each high-frequency sensor can be extracted in sequence, and the abnormal data or invalid data can be removed in advance to obtain multiple extracted pressure perception data segments. In this embodiment, the extracted pressure perception data segments can be stored in chronological order according to different sensors for subsequent visualization and classification.

[0086] S3: Visually transform the multiple pressure perception data segments to obtain corresponding image data to be classified.

[0087] It should be noted that, considering that the raw data collected by the full-scale loop high-frequency sensor is large in volume, with complex and unclear features, it is difficult to directly analyze the data. Therefore, the data can be converted into images for batch processing, and the extracted pressure sensing data segments can be drawn into visual images. This embodiment can use any appropriate software or equipment to convert data into images, and there is no restriction on the conversion method. Please refer to Figure 2 , Figure 2It is the image data obtained by using the visualization method of the embodiment of the present invention, wherein the first column is normal data with relatively regular curves, and the second column is abnormal data with irregular curves. It should be noted that during the full-scale circuit test of this embodiment, a six-axle heavy-loaded trailer was used, with a maximum single-axle weight of 20t and a vehicle running speed of 40-60km / h. Therefore, it can be seen that in the normal data, each peak combination will include six different individual peaks, corresponding to the six axles of the trailer, that is, the six pairs of parallel tires of the trailer apply pressure to the sensor. In addition, depending on the specific setting of the time interval for extracting data, it may happen that vehicles pass one after another within the same time interval, such as Figure 2 As shown in , multiple peak combinations appear in one visualization image, usually including two peak combinations and three peak combinations.

[0088] This embodiment processes a large amount of raw data from multiple high-frequency sensors on a full-scale road surface into images, which can make the data distribution and classification characteristics more obvious, intuitive and easy to distinguish. A lightweight classification network is constructed later to realize the recognition of the classification network.

[0089] S4: Construct a convolutional neural network.

[0090] Specifically, step S4 includes:

[0091] S41: Construct a convolutional neural network for classifying the image data.

[0092] Specifically, based on the network pruning optimization theory and combined with the data characteristics of high-frequency sensors, this embodiment needs to perform some structural pruning on the general complex deep learning classification network when constructing a convolutional neural network.

[0093] First, remove the excessive feature extraction layers in the existing deep learning classification network. This is because in this embodiment, the dynamic perception data after data type conversion only needs to retain its shape features. The data is mainly classified based on the overall fluctuation trend of the data and the number of peak points. Therefore, it is sufficient to choose to retain an appropriate number of convolutional layers and pooling layers.

[0094] Secondly, since this embodiment needs to quickly classify the monitoring data in real time, when abnormal data is generated due to the vehicle deviating from the buried area of ​​the road sensor, the driver can be contacted in time to correct the direction of the loading vehicle and ensure that its wheel track is distributed in a normal position. It is necessary to increase the network training speed and pay attention to the design of the overall network parameters. In the training process of the convolutional neural network, the parameters are transmitted layer by layer, and the size of the overall network parameters mainly depends on the format of the input data and the design of the network structure. Since the size of the high-frequency sensor image after visualization is fixed, when specifying the number of channels, this embodiment designs a single-channel input to ignore the color characteristics of the data, thereby reducing the number of network parameters. At the same time, when more attention is paid to the shape characteristics of the data, it will also be more conducive to the network itself to extract more effective features.

[0095] In addition, the choice of pooling method will also affect the final classification results. Average pooling can reduce the error caused by the limited neighborhood size, and retain more image background information. Maximum pooling can reduce the offset error of the estimated mean caused by the convolution layer parameter error, and retain more texture information. Since the visualized high-frequency sensor data image does not have too much background information, the maximum pooling layer is designed to be added to the convolutional neural network of this embodiment.

[0096] In summary, in order to reduce the number of parameters and improve the network classification efficiency, the embodiment of the present invention proposes a deep convolutional neural network (DCNN6) with few parameters and fast training speed to classify the visualized perception data. Figure 3 , Figure 3 is a structural diagram of a DCNN6 convolutional neural network provided by an embodiment of the present invention, Figure 4 It is a parameter diagram of a DCNN6 convolutional neural network provided in an embodiment of the present invention. The DCNN6 convolutional neural network includes a feature extraction unit and a feature mapping unit, wherein the feature extraction unit is used to extract various features of the input image data, and the feature extraction unit includes six convolutional layers and one pooling layer connected in sequence, and the six convolutional layers respectively use the relu function as the activation function, and combine the pooling layer to extract features of the image data to obtain a feature image; the feature mapping unit is used to use the mapping relationship between the feature image and the image category transition to obtain the classification result of the feature image, and the feature mapping unit includes a flat layer and two fully connected layers connected in sequence. Please refer to Table 1, which is the parameters of the six convolutional layers in the embodiment of the present invention.

[0097] Table 1 Parameters of each convolutional layer

[0098]

[0099] Specifically, in this embodiment, the convolutional neural network input is an image of size (650*202), that is, the extracted pressure perception data fragment is visualized as image data of size (650*202) and single color. The convolution layer design of the feature extraction unit is 6 layers in total, each layer uses the relu function as the activation function, and uses a sliding window similar to a sliding filter to extract features from the image. The input of each neuron is connected to the local receptive field of the previous layer, and the local features are extracted, and the local weights are shared. Weight sharing reduces the amount of network parameter calculations, reduces the complexity of the model, and does not need to consider the location where the features appear in the image. The convolution layer is usually used to extract features from the input data, and the implicit correlation in the original data is abstracted through the convolution kernel matrix. The convolution operation is actually a mathematical operation of dot multiplication and summation of two pixel matrices, one of which is the input data matrix, and the other is the convolution kernel (filter or feature matrix). The result obtained is represented by a specific local feature extracted from the original image. The activation function is responsible for activating the features extracted by the convolution layer. Since the convolution operation is a linear change relationship between the input matrix and the convolution kernel matrix, the activation function is required to map it nonlinearly. The activation function is to embed a nonlinear function based on the output result of the convolution layer, so that the output feature map has a nonlinear relationship. ReLU is usually used as an activation function in convolutional networks (including tanh and sigmoid, etc.). The ReLU function can limit values ​​less than 0 to 0, while values ​​greater than or equal to 0 remain unchanged.

[0100] The pooling layer is located behind multiple convolutional layers and is used to compress and transmit data and parameters to avoid overfitting during the classification process. The role of the pooling layer is to screen the features within the receptive field and extract the most representative features in the area, which can effectively reduce the output feature scale and thus reduce the number of parameters required for the model. Pooling methods are generally divided into average pooling and maximum pooling. The embodiment uses the maximum pooling method to compress and transmit images, and the specific parameters include the window size of the maximum pooling and the step size of the pooling window movement. The window size is (2,2) and the step size is (2,2). In the convolutional neural network, the role of the pooling layer is to identify images that have been displaced, scaled, and otherwise distorted without changing their properties.

[0101] The feature mapping layer includes a flat layer and two fully connected layers, which are composed of the mapping relationship between image features and image category changes. The activation function of the feature mapping layer of this embodiment adopts the Softmax function. The output range of the Softmax function is between (0,1) and has the smoothness of an exponential function. In the classification output result, the closer to 1, the greater the possibility of the class.

[0102] Specifically, the flattening layer is used to "flatten" the data input to the flattening layer, that is, to convert the multi-dimensional input into one dimension, for the transition from the convolutional layer to the fully connected layer. The flattening layer does not affect the size of the step size. It is reflected in the program as the Flatten function.

[0103] The core operation of the fully connected layer is the matrix-vector product, which is essentially a linear transformation from one feature space to another feature space. Therefore, the purpose of the fully connected layer is to transform the previously extracted features nonlinearly, extract the associations between these features, and finally map them to the output space. The effect achieved is to integrate all the features extracted by the feature extraction layer together, and then use the activation function equivalent to the classifier to output the final classification result.

[0104] S42: Obtain a training data set using multiple high-frequency sensors from a full-scale road surface;

[0105] After the convolutional neural network is constructed, the image data to be classified cannot be directly classified. The convolutional neural network needs to be trained using a training data set.

[0106] Specifically, step S42 of this embodiment includes:

[0107] First, a large amount of raw data from multiple high-frequency sensors on a full-scale road surface is obtained. The specific process is the same as step 1 and will not be repeated here. Alternatively, the raw data of existing experiments on the same full-scale road surface can also be collected.

[0108] Next, the pressure perception data in the loading process from the large amount of raw data is extracted using the above-mentioned adaptive threshold. Specifically, if the difference between the maximum data value and the minimum data value in the current extraction time period exceeds the range of the preset threshold, the segment of data is considered to be abnormal data or invalid data, and the extraction is abandoned. If the difference between the maximum data value and the minimum data value in the current extraction time period is within the range of the preset threshold, the segment of data is extracted and saved. According to the extraction principle, the raw data obtained from each high-frequency sensor can be extracted in turn, and the abnormal data or invalid data can be removed in advance to obtain a plurality of extracted pressure perception data segments. In this embodiment, the extracted pressure perception data segments can be stored in chronological order according to different sensors for subsequent visualization and classification. Subsequently, the extracted pressure perception data is visualized and converted to obtain a large amount of image data.

[0109] Finally, the large amount of image data is annotated, and normal labels and abnormal labels are added, so as to obtain a large amount of image data with normal labels or abnormal labels to form the training data set. Specifically, after obtaining a large amount of image data, a normal label or an abnormal label is set for each image data through annotation, that is, the image data is manually annotated as normal data or abnormal data.

[0110] In this embodiment, a semi-automatic labeling method for high-frequency sensor data images based on a high similarity criterion is used, which is characterized by combining automatic image labeling with manual review, thereby saving a lot of labeling time and energy.

[0111] Fast and accurate image feature annotation of the extracted and visualized high-frequency sensor data image set can provide accurate loading information for later network training, and on this basis, a data intelligent classification network model with strong generalization ability can be trained. The high similarity annotation criterion is an automated image annotation method, which is manually reviewed after the annotation is completed. Figure 5 and Figure 6 , Figure 5 It is a process schematic diagram of a semi-automatic annotation method for high-frequency sensor data images provided by an embodiment of the present invention; Figure 6 Schematic diagram of a labeling network provided by an embodiment of the present invention. The labeling network of this embodiment includes an input layer, four convolutional layers, a fully connected layer and an output layer connected in sequence.

[0112] In this embodiment, the training process of the annotation network model based on the high similarity criterion is mainly divided into three stages.

[0113] (I) Pre-training process: Use 20 manually annotated high-frequency sensor data images to pre-train the constructed annotation network, realize the visual representation of the intermediate convolution layer of the high-frequency sensor data image target, and initialize the network parameters. Specifically, it includes the following processes: training the node weight W1 between the input layer and the first convolution layer; using the conditional probability of the first convolution layer node as the input of the second convolution layer; combining the first convolution layer and the second convolution layer to train the node weight W2; using the same method to train the node weights of the remaining 2 convolution layers and 2 fully connected layers.

[0114] (II) Parameter fine-tuning stage: using back propagation to minimize network error

[0115] The sum of the labeling errors of each node is calculated as the overall labeling error. The cross entropy loss function is selected as the error here. The node weights are fine-tuned and optimized by back-propagating the labeling error to minimize the KL divergence (Kullback–Leibler divergence) between the predicted probability of the labeling network and the reference probability, also known as relative entropy. This indicator is used to measure the difference between two probability distributions.

[0116] (III) Update the weight of each node

[0117] Back propagation is used to update the node weights of each layer to optimize the parameters of the annotation network and complete the training of the annotation network. Subsequently, the large amount of image data to be annotated is input into the trained annotation network to complete the semi-automatic annotation of the image. In order to obtain better accuracy, a simple manual review is performed on the high-frequency sensor monitoring data images after the network automatic annotation. Here, only the image sets of different categories need to be confirmed as a whole, without the need to judge the category of a single image. Compared with manual annotation, it can greatly save time and manpower.

[0118] Of course, in other embodiments, the high-frequency sensor data image set may also be directly manually annotated to form a training data set.

[0119] In this embodiment, as described above, according to the specific setting of the time interval for extracting data, vehicles may pass by one after another in the same time interval, so that Figure 2 As shown in the normal data in the first column, multiple peak combinations appear in a visualized image, usually including two peak combinations and three peak combinations. Therefore, in this step, the normal data can also be divided into image data with one peak combination, image data with two peak combinations, and image data with three peak combinations, so that the normal data is annotated into three categories, and the abnormal data is added, and all the image data are annotated into 4 categories of manually annotated image data to form the final training data set.

[0120] It should be noted that after the data is processed according to the corresponding rules of extraction and visualization, the normal data fragments obtained include the above three types of image data. Because these three types of image data have significantly different data characteristics, the network can classify the images more easily and quickly by classifying each type of image data separately. Therefore, the three types of images are first labeled into three categories, namely, image data with one peak combination, image data with two peak combinations, and image data with three peak combinations. Then input the classification network to complete the four-category classification. Finally, the abnormal data fragments are extracted from the four categories, and the three types of normal images are generally counted.

[0121] The classification speed of this classification method for a single image is higher than the classification speed when the three types of images are not distinguished and directly regarded as normal images. At the same time, because the characteristics of each type of image are clear and distinct after classification, the classification network can better learn its characteristics, so the classification accuracy is also higher than the classification accuracy when the three types of images are not distinguished and directly regarded as normal images.

[0122] S43: Using the training data set to train the neural network model to obtain a trained convolutional neural network model.

[0123] Specifically, the images in the training data set are input into the constructed convolutional neural network for continuous training until the loss curve of the convolutional neural network basically stops changing, and the training is stopped. At this time, the accuracy reaches the optimal level, the error is minimized, and a trained convolutional neural network model is obtained. It should be noted that the training process of the convolutional neural network is divided into two stages. The first stage is the stage where data is propagated from a low level to a high level, that is, the forward propagation stage. The other stage is when the result obtained by the forward propagation does not meet the expectation, the error is propagated from the high level to the low level for training, that is, the back propagation stage. The specific training process includes:

[0124] (1) Initializing the weights of the convolutional neural network;

[0125] (2) inputting the image data in the training data set, and forward propagating through the convolutional layer, the pooling layer, and the fully connected layer to obtain the output value;

[0126] (3) Calculate the error between the network output value and the target value;

[0127] (4) When the error is greater than the expected value, the error is passed back to the network, and the errors of the fully connected layer, pooling layer, and convolutional layer are calculated in turn. The error of each layer can be understood as how much of the total error of the network should be borne by the network; when the error is equal to or less than our expected value, the training ends.

[0128] S5: Using the trained convolutional neural network model to process the image data to be classified, to obtain a classification result of the image data to be classified.

[0129] Specifically, after the convolutional neural network model is trained, the acquired high-frequency sensor raw data is extracted and visualized, and the image data is input into the trained convolutional neural network model to obtain the category of each image data. It should be noted that if the convolutional neural network is trained using a data set with two types of labels, the trained neural network model can classify the image data into two categories and determine whether the image belongs to normal data or abnormal data; if the convolutional neural network is trained using a data set with four types of labels, the trained neural network model can classify the image data into four categories and determine whether the image belongs to normal data with one peak combination, normal data with two peak combinations, normal data with three peak combinations, or abnormal data.

[0130] According to the characteristics of visualized data, the embodiment of the present invention designs a lightweight deep convolutional neural network model DCNN6. Compared with the commonly used classification network, the network parameters of this network are greatly reduced, the convergence speed is fast, and it is not easy to fall into overfitting. It realizes the rapid classification and anomaly detection of the road pressure high-frequency sensor monitoring data, and provides strong data support for the force analysis of the entire road area and the grasp of the evolution law of rutting. This method directly uses the method of visualizing image data using time domain data, which simplifies the complex processing process of other existing methods that require data to be converted into time and frequency domains or one-dimensional time series data to be arranged in two dimensions.

[0131] S6: Using the classification results, the time and space positions corresponding to the abnormal data are obtained, and the driving trajectory of the vehicle is prompted.

[0132] See also Figure 7 , Figure 7 1 is a schematic diagram of the steps of using the splitting result of an embodiment of the present invention to provide a vehicle driving trajectory prompt. Specifically, using the classification result:

[0133] 1) If it is the classification result of real-time high-frequency sensor monitoring data:

[0134] The spatiotemporal information of the images classified as abnormal is extracted. Specifically, the time and spatial position information corresponding to the abnormal images is extracted. The specific loading vehicle is found based on the location information of the sensor abnormal data and the GPS tracking data of the loading vehicle. The driver is prompted to correct the driving trajectory of the vehicle in time and effectively load the full-scale road surface according to the buried position of the high-frequency pressure gauge.

[0135] 2) If it is the classification result of historical high-frequency sensor monitoring data:

[0136] The ratio of normal data to abnormal data is calculated. If the ratio of abnormal data exceeds 0.5%, it is necessary to extract the time and space information corresponding to this period of historical data, and find the specific loading vehicle based on the location information of the sensor abnormal data and the GPS tracking data of the loading vehicle, and prompt the driver's driving trajectory, that is, remind the driver to effectively load according to the location where the high-frequency pressure gauge is buried.

[0137] The loader driving correction method based on intelligent classification of road surface perception data in this embodiment can effectively solve the problem of automatic classification of dynamic high-frequency sensor data in a full-scale loop, and provide strong data support for the force analysis of the entire road area and the grasp of the evolution law of rutting. It has fast classification speed and high accuracy, saves manpower and material resources, and has great engineering significance.

[0138] Embodiment 2

[0139] On the basis of Example 1, this embodiment provides a loading vehicle driving correction system based on intelligent classification of road surface perception data, the system comprising a data acquisition module, a data extraction module, a data visualization module, a data classification module and a prompt module, wherein the data acquisition module is connected to a plurality of high-frequency sensors located on a full-scale road surface, for acquiring the raw data of the plurality of high-frequency sensors; the data extraction module is used to construct an adaptive threshold according to the data characteristics of different high-frequency sensors on the full-scale road surface, extract the sensor data in the loading process, and obtain a plurality of pressure perception data fragments; the data visualization module is used to perform visual conversion on the plurality of pressure perception data fragments, and obtain corresponding image data to be classified; a DCNN6 convolutional neural network is provided in the data classification module, and the DCNN6 convolutional neural network is used to classify the image data to be classified after training, and output the classification results; the prompt module is used to obtain the time and space position corresponding to the abnormal data using the classification results, and to prompt the driving trajectory of the vehicle.

[0140] During the full-scale road test, a large number of high-frequency sensors are arranged on each layer of the road structure, so that the road stress caused by the distribution of the vehicle wheel track can be accurately grasped through data. In this embodiment, the raw data of the vehicle passing by is obtained through the data acquisition module 1 connected to each high-frequency sensor.

[0141] Furthermore, the data extraction module of this embodiment includes a storage unit, a threshold setting unit and a data processing unit, wherein the storage unit is used to store relevant parameters of the multiple high-frequency sensors, and the relevant parameters include data source format, sensor information and associated data of the sensors; the threshold setting unit is used to adaptively set the data threshold and extraction time required for data extraction according to the relevant parameters of the current high-frequency sensor, wherein the data threshold includes the upper extreme value threshold, the lower extreme value threshold of the data, the maximum interval extreme difference threshold between the upper and lower extreme value thresholds, and the minimum interval extreme difference threshold between the upper and lower extreme value thresholds; the data processing unit is used to extract the original data to be classified using the relevant parameters and the data threshold, and extract the pressure perception data in the loading process.

[0142] The raw data sampling frequency is high and contains a lot of noise information in the unloaded time period. After obtaining the raw data, it is necessary to extract the pressure sensing data under loading. The data extraction module of this embodiment can automatically complete the extraction of effective loading data by setting information such as sensor technical parameters. For the specific data extraction process, please refer to Example 1, which will not be repeated here.

[0143] For further information, see Figure 3 and Figure 4 The DCNN6 convolutional neural network of this embodiment includes a feature extraction unit and a feature mapping unit, wherein the feature extraction unit is used to extract various features of the input image data, wherein the feature extraction unit includes six convolutional layers and one pooling layer connected in sequence, the six convolutional layers respectively use the relu function as the activation function, and combine the pooling layer to extract features of the image data to obtain a feature image; the feature mapping unit is used to obtain the classification result of the feature image by utilizing the mapping relationship between the feature image and the image category transition, and the feature mapping unit includes a flat layer and two fully connected layers connected in sequence.

[0144] It should also be noted that the DCNN6 convolutional neural network is trained using an annotated image data set, which includes a large number of annotated normal images and a large number of abnormal data. In this embodiment, the image data set is semi-automatically annotated using a trained annotation network, wherein the specific structure of the annotation network and the annotation process are described in Example 1, which will not be repeated here.

[0145] In this embodiment, as described above, according to the specific setting of the time interval for extracting data, vehicles may pass by one after another in the same time interval, so that Figure 2As shown in , multiple peak combinations appear in a visualized image, usually including two peak combinations and three peak combinations. Therefore, in this step, normal data can also be divided into image data with one peak combination, image data with two peak combinations, and image data with three peak combinations, so that normal data can be annotated into three categories, and all image data can be annotated into 4 categories of manually annotated image data to form the final training data set.

[0146] The images in the training data set are input into the constructed convolutional neural network model for continuous training until the loss curve of the convolutional neural network model basically stops changing, and the training is stopped. At this time, the accuracy reaches the optimal level and the error is minimized, thereby obtaining a trained convolutional neural network model.

[0147] After the convolutional neural network model is trained, the high-frequency sensor raw data is obtained, and the image data is input into the trained convolutional neural network model through the above data extraction and visualization conversion to obtain the category of each image data. It should be noted that if the convolutional neural network is trained using a data set with two types of labels, the trained neural network model can classify the image data into two categories and determine whether the image belongs to normal data or abnormal data; if the convolutional neural network is trained using a data set with four types of labels, the trained neural network model can classify the image data into four categories and determine whether the image belongs to normal data with one peak combination, normal data with two peak combinations, normal data with three peak combinations, or abnormal data.

[0148] Furthermore, in order to further verify the effectiveness of the intelligent classification method proposed in the embodiment of the present invention, the classification method based on the DCNN6 neural network of the present embodiment, the classification method based on the existing Resnet50 neural network, and the classification method based on the existing GhostNet model are used to classify the same data set. Figure 8 , Figure 8 This is a flow chart of the verification experiment of the intelligent classification method of the embodiment of the present invention. In this embodiment, the training data set contains a total of 2404 images as the experimental data set. The experiment belongs to a four-category task, and the normal data is divided into three categories according to unimodal, bimodal, and trimodal distribution, and one category of abnormal data. The data set is divided into a training set and a test set, and the ratio is about 5:1, that is, 2000 training sets and 404 test sets. Among them, there are 500 images in each category of the training set, and the number of unimodal, bimodal, and trimodal abnormalities in the test set are 99, 99, 92, and 112 respectively. The designed DCNN6 neural network model is trained, and after the convolution layer, the following is obtained. Fig. 9 The characteristic diagram shown. Fig. 9It can be seen that the DCNN6 neural network proposed in this embodiment focuses more on the shape features of the image, which meets our expectations. Furthermore, the change curves of data classification accuracy and loss based on the DCNN6 model are as follows: Fig.10 and Fig.11 As shown in the figure, the confusion matrix and classification evaluation indicators are as follows: Fig.12 As shown in Table 2, the classification accuracy of the first three categories of normal data can reach 100%, and the accuracy of the fourth category of abnormal data is about 93%.

[0149] Table 2 Classification results of DCNN6 model

[0150]

[0151] In this example, in order to demonstrate the accuracy of the present invention, a comparative experiment is conducted on the abnormal data using a classification method based on the Resnet50 neural network. Fig.13 and Fig.14 , Fig.13 and Fig.14 These are curves showing the change in classification accuracy and loss of the classification model Resnet50 during training. Analysis shows that after the second iteration, the accuracy and loss value of the classification model Resnet50 basically do not change. At the same time, the classification accuracy obtained on the training set is extremely high (basically 100%), but the accuracy is extremely low when using the test set. Printing the network training parameters shows that most of the remaining parameters of the network have not been updated, which means that the structure of the network is too complex for the training set data, which makes it very easy to overfit during training. For more information, please refer to Fig.15 As shown in Table 3, after testing the test set, the average accuracy of the classification model Resnet50 is only about 20.5%.

[0152] Table 3 Classification results of traditional Resnet50 model

[0153]

[0154] For further information, see Fig.16 , Fig.16 This is a schematic diagram of the structure of the Ghost module in the existing GhostNet model. The Ghost module in the GhostNet model can use fewer parameters while generating the same number of feature maps. The main principle is to divide the ordinary convolution layer into two parts for training. The first part is a part of the convolution with a strictly controlled total number, which is used to generate an inherent number of feature maps. Then the second part needs to apply some simple linear operations to generate more feature maps. When the output feature map is the same size, the number of parameters and computational complexity are greatly reduced.

[0155] According to the test results of the Resnet50 model and DCNN model after designing the parameters, try to use the lighter GhostNet model to do the binary classification experiment. Figures 17 to 19 , Fig.17 It is a curve chart of the classification accuracy change of the GhostNet model during the training process; Fig.18 It is the loss change curve of the GhostNet model during training;

[0156] Fig.19 It is the confusion matrix of the classification results using the GhostNet model.

[0157] The recall rate, precision rate and F1-score of the four-category classification using the GhostNet model are shown in Table 4.

[0158] Table 4 Classification results of GhostNet model

[0159]

[0160] As can be seen from the above table, the average accuracy of the four-category classification using the GhostNet model is above 90%, which proves that the model has good classification performance on the above dataset.

[0161] Table 5 Comparison of classification accuracy evaluation indicators of different classification network models

[0162]

[0163] Table 6 Comparison of different classification network model parameters and classification speed evaluation indicators

[0164]

[0165] As shown in Tables 5 and 6, the Resnet50 model with adjusted parameters performs poorly on the data set. Its classification accuracy, speed, and model parameter quantity are all at a poor level. The parameter quantity is more than 33M, and the training set test time exceeds 10 minutes. Although the acc and loss change trends of the Resnet50 model during training are good, its performance on the validation set is too poor, with the average recall rate and average precision rate both below 0.3, and the F1-score is even only 0.12. Therefore, it is inferred that overfitting is very likely to occur. The lightweight model GhostNet can achieve good classification accuracy, with the average recall rate, average precision rate, and F1-score all above 90%. At the same time, the parameter quantity is greatly reduced to only 2.59M, but the classification speed is still insufficient, and the training set test time exceeds 10 minutes. The DCNN6 model proposed in this embodiment performs well in classification accuracy and classification speed. The average recall rate, average precision rate and F1-score can reach more than 95%, and the training set test time is also within 10 minutes, but the model has a large number of parameters, exceeding 22M, and there is still room for improvement. Among the three indicators, classification accuracy and classification speed can best reflect the performance of the model on the data set of this embodiment. Therefore, the proposed DCNN6 model can efficiently and accurately identify full-scale loop dynamic perception data.

[0166] According to the characteristics of visualized data, this embodiment proposes a lightweight deep convolutional neural network model DCNN6. Compared with the commonly used classification network, the network parameters of this network are greatly reduced, the convergence speed is fast, and it is not easy to fall into overfitting. It can realize the rapid classification and anomaly detection of the road pressure high-frequency sensor monitoring data, monitor the driving trajectory of the loading vehicle, quickly discover the vehicle deviation from the preset trajectory, and correct the driver's driving.

[0167] This provides strong data support for the stress analysis of the entire road area and the understanding of the evolution law of rutting. This embodiment processes a large amount of raw data from multiple high-frequency sensors on the full-scale road surface into images, which can make the data distribution and classification characteristics more obvious, intuitive and easy to distinguish. In the later stage, a lightweight classification network is constructed to realize the recognition of the classification network.

[0168] In the several embodiments provided by the present invention, it should be understood that the apparatus and method disclosed by the present invention can be implemented in other ways. For example, the apparatus embodiments described above are only illustrative, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation, for example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0169] In addition, each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0170] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A method for correcting the deviation of a loading vehicle based on intelligent classification of road surface perception data, characterized in that: include: Obtain raw data from multiple high-frequency sensors on full-scale road surfaces; An adaptive threshold is constructed based on the data characteristics of different high-frequency sensors on the full-scale road surface, and the sensor data in the loading process is extracted to obtain multiple pressure sensing data fragments; Performing visual conversion on the multiple pressure sensing data segments to obtain corresponding image data to be classified; Constructing a convolutional neural network, wherein the convolutional neural network includes a feature extraction unit and a feature mapping unit, wherein the feature extraction unit is used to extract various features of the input image data, including six convolutional layers and one pooling layer connected in sequence, the six convolutional layers respectively use a relu function to extract features from the image data to obtain a feature image, and the pooling layer is used to compress and transmit the feature image; the feature mapping unit is used to obtain a classification result of the feature image by using a mapping relationship between the feature image and image category transition, and the feature mapping unit includes a flat layer and two fully connected layers connected in sequence; Processing the image data to be classified using a trained convolutional neural network model to obtain a classification result of the image data to be classified; The classification result is used to obtain the time and space position corresponding to the abnormal data, and to provide a prompt for the driving trajectory of the vehicle.

2. The method for correcting the deviation of a loading vehicle based on intelligent classification of road surface perception data according to claim 1 is characterized in that: An adaptive threshold is constructed based on the data characteristics of different high-frequency sensors on the full-scale road surface to extract the sensor data in the loading process, including: Storing relevant parameters of the plurality of high-frequency sensors, wherein the relevant parameters include a data source format and parameter information of the sensors; Adaptively set the data threshold required for data extraction according to the relevant parameters of the current high-frequency sensor; The raw data to be classified are extracted by using the relevant parameters and the data threshold, and the pressure sensing data in the loading process is extracted.

3. The method for correcting the deviation of a loading vehicle based on intelligent classification of road surface perception data according to claim 2 is characterized in that: The data threshold required for data extraction is adaptively set according to the relevant parameters of the current high-frequency sensor, including: Extract descriptive indicators of different high-frequency sensor data, including at least mean, median, mode, maximum, minimum, 25th percentile, 75th percentile, range, quartile range, and dispersion coefficient; The data distribution characteristics of different high-frequency sensors are comprehensively analyzed through the descriptive indicators, and the data thresholds corresponding to different sensors are adaptively adjusted according to the data distribution characteristics.

4. The method for correcting the deviation of a loading vehicle based on intelligent classification of road surface perception data according to claim 3 is characterized in that: The data thresholds include an upper extreme value threshold and a lower extreme value threshold of the data, a maximum interval extreme difference threshold between the upper and lower extreme value thresholds, and a minimum interval extreme difference threshold between the upper and lower extreme value thresholds.

5. The method for correcting the deviation of a loading vehicle based on intelligent classification of road surface perception data according to claim 1 is characterized in that: Construct a convolutional neural network, including: Constructing a convolutional neural network for classifying the image data; Acquire a large amount of raw data from multiple high-frequency sensors on a full-scale road surface, extract and visualize the pressure sensing data in the loading process in the raw data, and obtain a large amount of image data; Annotating the large amount of image data, adding normal labels and abnormal labels, and obtaining a large amount of image data with normal labels or abnormal labels to form the training data set; The convolutional neural network is trained using the training data set to obtain a trained convolutional neural network model.

6. The method for correcting the deviation of a loading vehicle based on intelligent classification of road surface perception data according to claim 5 is characterized in that: The large amount of image data is labeled, and normal labels and abnormal labels are added, including: The constructed annotation network is pre-trained using multiple manually annotated high-frequency sensor data images to achieve visual representation of the intermediate convolutional layer of the high-frequency sensor data image target and initialize the network parameters; the annotation error is back-propagated to optimize the node weights to minimize the network error; the weights of each node are updated, and the node weights of each layer are updated in turn using back-propagation to achieve parameter optimization of the annotation network and complete the training of the annotation network; The large amount of image data to be annotated is input into a trained annotation network to complete semi-automatic annotation of the image.

7. A loader driving correction system based on intelligent classification of road surface perception data, characterized in that: It includes data collection module, data extraction module, data visualization module, data classification module and prompt module, among which, The data acquisition module is connected to a plurality of high-frequency sensors located on a full-scale road surface, and is used to obtain raw data of the plurality of high-frequency sensors; The data extraction module is used to construct an adaptive threshold according to different high-frequency sensor data characteristics of the full-scale road surface, extract sensor data in the loading process, and obtain multiple pressure perception data fragments; The data visualization module is used to perform visualization conversion on the multiple pressure sensing data segments to obtain corresponding image data to be classified; The data classification module is provided with a convolutional neural network, and the convolutional neural network is used to classify the image data to be classified after training and output the classification result; The prompt module is used to obtain the time and space position corresponding to the abnormal data using the classification result, and to prompt the driving trajectory of the vehicle.

8. The loader driving deviation correction system based on intelligent classification of road surface perception data according to claim 7 is characterized in that: The data extraction module includes a storage unit, a threshold setting unit and a data processing unit, wherein: The storage unit is used to store relevant parameters of the multiple high-frequency sensors, wherein the relevant parameters include data source format, sensor information, and associated data of the sensors; The threshold setting unit is used to adaptively set the data threshold required for data extraction according to the relevant parameters of the current high-frequency sensor, wherein the data threshold includes the upper extreme value threshold, the lower extreme value threshold, the maximum interval extreme difference threshold between the upper and lower extreme value thresholds, and the minimum interval extreme difference threshold between the upper and lower extreme value thresholds; The data processing unit is used to extract the raw data to be classified by using the relevant parameters and the data threshold, and extract the pressure perception data in the loading process.

9. The loader driving deviation correction system based on intelligent classification of road surface perception data according to claim 7 or 8, characterized in that: The convolutional neural network includes a feature extraction unit and a feature mapping unit, wherein: The feature extraction unit is used to extract various features of the input image data, wherein the feature extraction unit includes six convolutional layers and one pooling layer connected in sequence, the six convolutional layers respectively use the relu function to extract features from the image data to obtain a feature image, and the pooling layer is used to compress and transmit the feature image; The feature mapping unit is used to obtain the classification result of the feature image by utilizing the mapping relationship between the feature image and the image category transition. The feature mapping unit includes a flat layer and two fully connected layers connected in sequence.

Images