Engineering vehicle load state identification method

Abstract

The invention discloses an engineering vehicle load state identification method, which comprises the following operation steps of: acquiring original transportation data of an engineering vehicle, and performing data preprocessing on the original transportation data, including data cleaning and normalization; performing differential feature construction and feature selection on the preprocessed data, extracting a feature vector set, and segmenting the data by using a sliding window to obtain an input sequence with a fixed length; extracting local features and global features containing a feature channel dependency relationship from the input sequence through a GTCN-STransform embedded with an SE mechanism to obtain deep feature representation of the original input sequence; and carrying out classification identification on the load state of the engineering vehicle on the extracted deep features by using a full connection network and a Softmax function to obtain the load state of the engineering vehicle. The method has the advantages that the accuracy of engineering vehicle load state recognition is improved, and the problem of monitoring of transported goods is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of time series data classification, in particular to an engineering vehicle load state recognition method. BACKGROUND

[0002] In recent years, with the continuous acceleration of urban development, the volume of goods transportation is increasing day by day, and goods transportation has gradually become an important link in urban construction. However, due to the lack of effective real-time monitoring means for the state of the vehicle, the engineering vehicle has behaviors such as dumping at will, private pulling and disorderly transportation during transportation, which on the one hand causes environmental pollution, and on the other hand, easily increases the risk of traffic accidents due to overload and overlimit. Therefore, it is urgent to monitor the load state of the engineering vehicle in real time.

[0003] At present, the method for monitoring the load and load state of the engineering vehicle mainly adopts weighing, and the weighing methods mainly include static weighing method and dynamic weighing method. The static weighing method is mostly realized by setting weighing equipment such as a loadometer on a fixed road section, and has high measurement accuracy, but the weighing equipment is easy to be damaged and the measurement affects traffic safety because the vehicle stops running; the dynamic weighing method collects vehicle data in real time by means of a vehicle-mounted sensor, but due to the influence of the actual running environment and the insensitivity of the sensor, the data collected by the sensor has noise and fluctuation, and there is a certain gap with the real load, so it is difficult to accurately judge the load state of the vehicle only by relying on the change of the sensor data without the help of algorithm analysis.

[0004] The data generated during the transportation process of the engineering vehicle has time correlation and is a typical time series data, which can be studied by referring to the analysis method related to time series. The engineering vehicle load state recognition is a subdivision task in time series classification, and the purpose is to judge the load state (loading, unloading, non-loading and unloading state) of the vehicle according to the information collected by the sensor installed on the vehicle during the transportation process of the vehicle. For the analysis of time series data, the main analysis methods at present include recurrent neural network, long short-term memory network and the like. These methods can process variable-length time series and have high flexibility, but when capturing the time dependence of the time series, only the current input and its historical state are concerned at each time step, which limits the understanding of the global structure of the entire sequence, and the serial processing of data affects the training efficiency.

[0005] Therefore, how to provide an engineering vehicle load state recognition method with improved accuracy and reduced computational complexity has become a technical problem to be solved by those skilled in the art. SUMMARY

[0006] The purpose of the present application is to provide an engineering vehicle load state recognition method, which has the advantages of improving the accuracy of engineering vehicle load state recognition and effectively solving the monitoring problem of transported goods.

[0007] The above-mentioned technical objective of the present invention is achieved through the following technical solution: A method for identifying the load status of engineering vehicles includes the following steps: S1. Obtain the raw transportation data of the engineering vehicles and perform data preprocessing on the raw transportation data, including data cleaning and normalization. S2. Perform differential feature construction and feature selection on the preprocessed data, extract the feature vector set, and use a sliding window to segment the data to obtain a fixed-length input sequence. S3. By embedding the SE mechanism, the GTCN-STransformer extracts local and global features containing feature channel dependencies from the input sequence to obtain a deep feature representation of the original input sequence. S4. The extracted deep features are used to classify and identify the load status of engineering vehicles using a fully connected network and the Softmax function, so as to obtain the load status of engineering vehicles.

[0008] The preferred option is as follows: Preferred: In step S1, Acquire raw transportation data of engineering vehicles, including load, sensor deformation coefficient, direction, speed, vehicle voltage, longitude, latitude, altitude and data acquisition time; Time series data records with missing or duplicate entries in the original data are discarded, i.e., data cleaning. The Z-score normalization method is used to transform different feature data into dimensionless data to avoid large-scale features dominating the results of load status classification and recognition, thereby eliminating the influence of differences in the dimensions between features on load status classification and recognition.

[0009] Preferred: In step S2, S21. Construct differential features for the load, direction, and speed features in the preprocessed data, and add the differential features to the original features to form an initial feature set; S22. Select features from the initial feature set using the maximum information coefficient (MIC) to obtain a feature vector set; S23. For the feature vector set, a sliding window is used to segment the data to obtain a fixed-length input sequence.

[0010] Preferably: In step S22, Each feature and its corresponding label in the initial feature set are obtained sequentially to form a two-dimensional data point set; Map the set of two-dimensional data points onto a two-dimensional grid and calculate the corresponding mutual information values; Exhaustively enumerate all two-dimensional grid sizes and select the maximum mutual information under different two-dimensional grid sizes as the MIC score of the feature to measure the correlation between the feature and the load state; Repeat the above steps to obtain the MIC scores of all features in the initial feature set; Features with a MIC score greater than 0.4 are selected to form a feature vector set.

[0011] Preferred: In step S3, S31. Extract local temporal features from the input sequence using grouped temporal convolution GTCN; S32. By using the squeeze excitation mechanism SE embedded in the temporal convolution GTCN, the interdependencies between local temporal feature channels extracted by the temporal convolution GTCN are captured. By assigning different weights to the feature channels, more attention is paid to the features that contribute more to the load state recognition task. Features of GTCN are extracted by integrating temporal convolutions through grouped shuffling temporal convolutions. S33. Global modeling of multidimensional temporal local features containing channel dependencies is performed using the Sparse Transformer. The sparsity mechanism in the Sparse Transformer reduces the amount of computation and reduces the impact of redundant information.

[0012] Preferably: In step S31, When the extracted temporal local feature channels reach their maximum, i.e. the neck of the network, the temporal local features containing channel dependencies are further integrated through grouped shuffling temporal convolution in Temporal Convolution GTCN, while reducing computational complexity.

[0013] Preferably: In step S32, The squeezing stage aggregates the feature information of each channel through global average pooling and generates a channel-by-channel local feature descriptor for each channel; The activation phase learns the relationship between local feature descriptors of different channels through two layers of nonlinear mapping, generates attention weights for each channel with global information, and enhances important feature channels by multiplying them with the original feature channels.

[0014] Preferably: In the global modeling of step S33, The multidimensional local features of the input sparse transformer STransformer are positionally encoded using a sine-cosine function to generate a unique representation for each position in the local features. By inputting the multi-dimensional local feature representation containing location information into the multi-head sparse attention layer in the STransformer, the contextual dependencies of local features are learned to obtain the temporal global feature representation. At the same time, the computational complexity is reduced through sparse connections between self-attentions. By applying nonlinear mapping at each location through the feedforward layer in the Sparse Transformer, the temporal global feature representations learned by the multi-head sparse attention layer are further fused, thereby enhancing the expressive power of the features.

[0015] Preferably, in step S4, the deep features extracted by the GTCN-STransformer with embedded SE mechanism are transformed into probabilities of three load states: loading, unloading, and non-loading / unloading, through a fully connected network and the Softmax function, so as to obtain the load state identification result.

[0016] In summary, this invention utilizes temporal convolution within the temporal convolutional blocks of the Grouped Temporal Convolutional Network (GTCN) to extract local features from the input temporal data, expanding the receptive field. It leverages the SE mechanism within the temporal convolutional blocks to model the interdependencies between different channels. When the extracted local temporal features reach their maximum, grouped shuffling temporal convolution within the GTCN further integrates these channel-dependent features while reducing computational complexity. Furthermore, it employs a sparse transformer (STransformer) to model the global relationships among the multidimensional local temporal features containing channel dependencies, and further reduces computational complexity through sparse connections in the self-attention layer. This effectively improves the recognition accuracy in the scenario of identifying the load status of engineering vehicles. Attached Figure Description

[0017] Figure 1 This is a flowchart of an embodiment; Figure 2 This is a schematic diagram of the GTCN-STransformer network model with SE mechanism in an embodiment; Figure 3 This is a diagram of the temporal convolution block structure in an embodiment. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings.

[0019] A method for identifying the load status of engineering vehicles, such as Figure 1 As shown, the operation steps are as follows: S1. Obtain the raw transportation data of the engineering vehicles and perform data preprocessing on the raw transportation data, including data cleaning and normalization. S2. Perform differential feature construction and feature selection on the preprocessed data, extract the feature vector set, and use a sliding window to segment the data to obtain a fixed-length input sequence. S3. By embedding the SE mechanism, the GTCN-STransformer extracts local and global features containing feature channel dependencies from the input sequence to obtain a deep feature representation of the original input sequence. S4. The extracted deep features are used to classify and identify the load status of engineering vehicles using a fully connected network and the Softmax function, so as to obtain the load status of engineering vehicles.

[0020] In step S1, Raw transportation data of engineering vehicles is acquired by sensors installed on the vehicles. The raw data includes load, sensor deformation coefficient, direction, speed, vehicle voltage, longitude, latitude, altitude and data acquisition time. Time series data records with missing or duplicate entries in the original data are discarded, i.e., data cleaning. The Z-score normalization method is used to transform different feature data into dimensionless data to avoid the large-scale features dominating the results of load status classification and recognition, so as to eliminate the influence of the difference in the dimensions between features on load status classification and recognition. Dimensionless data is a measure of the degree of deviation of the original data from the mean (in units of standard deviation). Taking vehicle load as an example, the load normalization process at time t is as follows:

[0021] in, for The load of time, and These are all the times in the column containing the load. Mean and standard deviation of the load: In step S2, differential feature construction and feature selection are performed on the preprocessed data to extract the feature vector set, and a sliding window is used to tap a fixed-length input sequence at time t. ; S21. Construct differential features for the load, direction, and speed features in the preprocessed data, and add the differential features to the original features to form an initial feature set; To reflect the changing trends of various features in vehicle operation data over time and to access more historical data without increasing the sliding window length, forward differencing is used to further extract features from the preprocessed data. Taking vehicle load as an example... time Forward step difference is divided into: Since speed difference can reflect acceleration or deceleration behavior, direction difference can capture vehicle steering, and load difference can specifically reflect whether the vehicle load increases or decreases, preliminary analysis constructs 5-step and 10-step difference features for the load, direction, and speed features in the preprocessed data, and adds the difference features to the original features to form an initial feature set.

[0022] S22. Select features from the initial feature set using the maximum information coefficient (MIC) to obtain a feature vector set. Sequentially obtain each feature in the initial feature set and its corresponding label to form a two-dimensional data point set. Map the set of two-dimensional data points onto a two-dimensional grid and calculate the corresponding mutual information values; Exhaustively enumerate all two-dimensional grid sizes and select the maximum mutual information under different two-dimensional grid sizes as the MIC score of the feature to measure the correlation between the feature and the load state; Repeat the above steps to obtain the MIC scores of all features in the initial feature set; Features with a MIC score greater than 0.4 are selected to form a feature vector set.

[0023] IC calculates the correlation between each feature variable and the label variable in the initial feature set, i.e., the loading, unloading, and non-loading / unloading states, to capture the linear or non-linear relationship between the features and the load state, and selects the features with high correlation for subsequent learning. The specific steps of MIC are as follows: Step 1: Obtain each feature and its corresponding label from the initial feature set sequentially to form a two-dimensional data point set. For the first feature... Features and tags , Represented as length is set , Represented as length is set The resulting set of two-dimensional data points is ; Step two, set the data points Map the data onto a two-dimensional mesh and calculate the corresponding mutual information values: and They are based on and The number of equally wide grids to divide the value range. yes and In the Line number Joint probabilities in column grid cells and They are and The marginal probability; Step 3: Enumerate all 2D mesh sizes and The maximum mutual information value under different 2D grid sizes is selected as the MIC score of the feature, and the MIC scores are normalized to make the MIC scores of different features comparable. This is the upper limit of the grid size (usually) ), It is a normalization factor that scales the MIC score to the range of [0,1]. Step 4: Repeat the above steps to obtain the MIC scores of all features in the initial feature set; Step 5: Select features with a MIC score greater than 0.4, i.e., load capacity. ), sensor deformation coefficient ( ),speed( ),longitude( ),latitude( ), load 5-step differential ( ), load 10-step differential ( ), speed 5-step difference ( ), speed 10-step difference ( ) constitute the feature vector set .

[0024] S23. For the feature vector set, a sliding window is used to segment the data to obtain a fixed-length input sequence; To provide richer feature representations, the feature vector set... A fixed input sequence is obtained at each time step using a sliding window. Specifically, in Always centering on yourself, capture both forward and backward segments. The records form a length of (Pick The number of feature variables is (Pick The input sequence of ) : exist At any given moment, slide the sliding window down one line to capture the content. The input sequence at each time step is obtained by using a sliding window, and this sequence is used as input for the network to learn and determine the load status at the current time step.

[0025] In step S3, GTCN consists of temporal convolutional blocks with embedded SE mechanisms and grouped shuffling temporal convolutions, extracting local features containing nonlinear relationships between channels. S31. Extract local temporal features from the input sequence using grouped temporal convolution GTCN; When the extracted temporal local feature channels reach their maximum, i.e. the neck of the network, the temporal local features containing channel dependencies are further integrated through grouped shuffling temporal convolution in Temporal Convolution GTCN, while reducing computational complexity.

[0026] Temporal Convolutional Networks (TCNs) evolved from Convolutional Neural Networks (CNNs) and are deep network architectures specifically designed for processing time series data. TCNs introduce causal convolution, allowing convolutions to capture only the local dependencies between the current and historical moments, preserving the sequential nature of the time series. Causal convolution is achieved by modifying the index offset. in, It is a convolution kernel The weight parameters, It is the input sequence exist The values ​​at each time point are used to obtain local features by applying causal convolution along the time axis. , It is the number of convolution kernels, which is also the number of output channels.

[0027] Temporal convolutional networks also introduce dilated convolutions, which expand the receptive field without increasing the number of parameters by increasing the spacing between convolutional kernels, thus capturing long-range dependencies. Since the network learns by truncating short input sequences to identify the load state at each time step in real time, only causal convolutions within the temporal convolution framework are used, without dilated convolutions, to better utilize temporal convolutions to capture local dependencies in the input sequence while preventing overfitting. To reduce the risks of vanishing and exploding gradients, temporal convolutional networks use residual connections to sum the features after causal convolution, weight normalization, and ReLU activation to form temporal convolution blocks. In the residual branch, a 1×1 convolution is used to keep the input channels consistent with the feature channels after causal convolution for summation, such as... Figure 3 As shown. The temporal convolutional block containing residual connections is shown in the following formula: in, It is a convolutional kernel with weights normalized. It is the input sequence. It is the input after a 1×1 convolution. It is the output of the time-convolutional block. It is a squeeze excitation calculation.

[0028] S32. The squeeze excitation mechanism SE embedded in the temporal convolution GTCN captures the interdependencies between local temporal feature channels extracted by the temporal convolution GTCN, such as... Figure 3As shown, the SE mechanism enhances the representational power of features extracted by temporal convolution by assigning different weights to feature channels to pay more attention to features that contribute more to the load state recognition task. The SE mechanism consists of two phases: squeezing and activation. The squeezing phase aggregates the feature information of each channel through global average pooling and generates a channel-wise local feature descriptor for each channel. During the activation phase, the relationship between local feature descriptors of different channels is learned through two layers of nonlinear mapping, generating attention weights for each channel with global information, and enhancing important feature channels by multiplying them with the original feature channels. The squeezing stage aggregates the feature information of each channel through global average pooling and generates a channel-by-channel local feature descriptor for each channel: in, These are local features obtained after causal convolution. It is a local feature descriptor obtained by extrusion.

[0029] The activation phase learns the nonlinear relationship between local feature descriptors of different channels through two layers of nonlinear mapping, generating attention weights with channel context information. : in, and These are the weight matrices of two fully connected layers. The first fully connected layer compresses information through dimensionality reduction, allowing the network to extract the most important parts from all channel information. The second fully connected layer maps neurons to their original number of channels. The sigmoid activation function transforms the output of the second fully connected layer into a weight vector in the range [0,1]. The specific formula for the sigmoid activation function is shown below: Temporal local features containing channel dependencies are obtained by multiplying with the original feature channels. :

[0030] Features extracted by temporal convolution blocks are integrated through grouped shuffling temporal convolution; Grouped shuffling convolution (GSConv) is a convolutional method that combines ordinary convolution and depthwise separable convolution. In object detection tasks, GSConv is used at the neck of the network to further enhance the features extracted by the backbone network, reducing computational complexity while maintaining accuracy. Applying grouped shuffling convolution to a temporal convolutional network forms grouped shuffling temporal convolution, which is used in the task of recognizing the load status of engineering vehicles. When the feature channels extracted by the temporal convolution block reach their maximum (i.e., the neck of the network), grouped shuffling temporal convolution further integrates temporal local features containing channel dependencies, such as... Figure 2 As shown. Grouped shuffling temporal convolution first extracts features with half the original number of channels using temporal convolution, preserving the connectivity between the original feature channels. Then, depthwise temporal convolution is used, applying convolution operations to each feature channel to further extract features. and These are the convolution kernels for temporal convolution and depthwise temporal convolution, respectively. It is time convolution in Time-based output, application Time convolution to obtain , It is the depthwise temporal convolution at the 1st... One channel The output at each time step is obtained by applying depthwise temporal convolution to each channel. .

[0031] Features extracted by temporal convolution and depthwise temporal convolution are concatenated and then fused using channel shuffling. Channel shuffling is achieved by reshaping the concatenated features into two sets of features, transposing them, and then reshaping them back to their original shapes. This method has low computational cost and effectively promotes the fusion of different sets of features. in, These are the characteristics after splicing, Mapped to Space, through Mapped to After Map back space.

[0032] The computational complexity of grouped shuffling time convolution The computational complexity of standard time convolution is... When the number of channels At that time, the ratio of the two Therefore, the computational complexity of grouped shuffling time convolution is reduced by half.

[0033] S33. Global modeling of multidimensional temporal local features containing channel dependencies is performed using the Sparse Transformer. The sparsity mechanism in the Sparse Transformer reduces the amount of computation and reduces the impact of redundant information.

[0034] The multidimensional local features of the input sparse transformer STransformer are positionally encoded using a sine-cosine function to generate a unique representation for each position in the local features. By inputting the multi-dimensional local feature representation containing location information into the multi-head sparse attention layer in the STransformer, the contextual dependencies of local features are learned to obtain the temporal global feature representation. At the same time, the computational complexity is reduced through sparse connections between self-attentions. By applying nonlinear mapping at each location through the feedforward layer in the Sparse Transformer, the temporal global feature representations learned by the multi-head sparse attention layer are further fused, thereby enhancing the expressive power of the features.

[0035] To capture global dependencies, the Transformer uses a self-attention mechanism at its core, generating attention scores by calculating the similarity between positions in the input sequence through dot products. Since attention scores exhibit a long-tailed distribution, calculating self-attention for all positions in the Transformer leads to information redundancy. The Sparse Transformer (STransformer) adds a sparsity mechanism to the Transformer, reducing computational complexity while modeling global dependencies for local features containing channel dependencies.

[0036] The sparse transformer STransformer consists of three parts: a location embedding layer, a multi-head sparse attention layer, and a feedforward layer. For example... Figure 2 As shown, since the computation of self-attention is order-independent, a sine-cosine positional encoding is introduced to preserve the positional relationships of the input sequence: in, and They are the locations The , 3D position encoding, It is the feature dimension. Positional information is embedded into the feature input through summation operations to obtain input with positional information. : By decoupling the feature input containing location information through linear mapping, a query (Q), a key (K), and a value (V) are generated as inputs to the sparse self-attention, as shown in the following formula: To achieve sparsity, KL divergence is used to measure the attention probability distribution of the query at position i. With uniform probability distribution The difference between them is shown in the following formula: in, It is the weight matrix of the linear mapping. It is the feature dimension of each subspace. Under a long-tailed distribution, random sampling Calculate the dot product pairs To avoid numerical stability issues, the max operation is used instead of the log-sum-exp operation.

[0037] The larger the KL divergence, the more information the attention contains, and the more important the query. The top u queries with large KL divergences are selected as the dominant queries to calculate sparse self-attention, as shown in the following formula: To extract features from input data from different perspectives, multi-head sparse attention will... Split into multiple subspaces along the feature dimension and compute sparse attention in parallel. The features from multiple subspaces are then concatenated and further fused using a linear mapping to obtain the output of the multi-head sparse attention. To alleviate the gradient vanishing problem, residual connections are added to the output of the multi-head sparse attention, and layer normalization is performed. By incorporating a sparsity mechanism into multi-head self-attention, the computational complexity is reduced from that of the original multi-head self-attention. Reduced to .

[0038] By applying a nonlinear mapping at each position through a feedforward layer, the temporal global feature representations learned by the multi-head sparse attention layer are further fused to enhance the expressive power of the features. in, It is the weight matrix and bias of the two linear mappings of the feedforward layer. , , , , It is the input of the feedforward layer, and finally the STransformer is calculated through residual connections and layer normalization.

[0039] In step S4, the deep features extracted by the GTCN-STransformer with embedded SE mechanism are transformed into probabilities of three load states: loading, unloading, and non-loading / unloading, through a fully connected network and the Softmax function, so as to obtain the load state identification result.

[0040] The features extracted by the GTCN-STransformer network with SE mechanism are reduced in dimensionality by global average pooling and then input into the fully connected layer. The extracted features are mapped to the class space, and the output of the fully connected layer is converted into a probability representation of three load states through the Softmax function. in, It is the fully connected layer. The classification output of each neuron. This indicates that the input sample is classified into the first category. The probability of a class satisfies .

[0041] The MIC scores of other features besides data acquisition time, as well as load, direction, and speed difference features in the engineering vehicle operation data obtained by the MIC method in this invention are shown in Table 1. Features with MIC scores higher than 0.4 are used to construct the feature vector set. After screening, the load capacity was selected. ), sensor deformation coefficient ( ),speed( ),longitude( ),latitude( ), load 5-step differential ( ), load 10-step differential ( ), speed 5-step difference ( ), speed 10-step difference ( ) is a characteristic term.

[0042] Table 1. MIC scores of candidate features The engineering vehicle acquires raw vehicle transportation data through onboard sensors, and then undergoes multiple steps to classify and identify the vehicle's load status, including data preprocessing and differential feature construction, feature selection, sliding window truncation of the input sequence, GTCN local feature extraction with embedded SE mechanism, STransformer global feature extraction, and classification of vehicle load status using fully connected layers and Softmax functions. The accuracy of the model's recognition is evaluated using four performance metrics: accuracy (Acc), macro precision (Macro_P), macro recall (Macro_R), and F1 score (Macro_F1). Comparison with benchmark models demonstrates the effectiveness of the engineering vehicle load status recognition method using GTCN-STransformer with SE mechanism; specific results are shown in Table 2.

[0043] Table 2. Recognition Performance Indicators of Different Models Analysis of Table 2 shows that the GTCN-SE-STransformer model has a significantly higher recognition accuracy than both GTCN-STransformer and STransformer. The Macro_F1 score of GTCN-SE-STransformer is 1.28% and 2.86% higher than that of GTCN-STransformer and STransformer, respectively. Therefore, the GTCN-SE-STransformer model has better classification accuracy and is more accurate in its recognition.

[0044] This invention addresses the problems of arbitrary dumping and unauthorized transport by engineering vehicles during transportation. It proposes a GTCN-STransformer-based engineering vehicle load status recognition model with a SE (Search Engine Optimization) mechanism to identify vehicle load status. First, sensor data is cleaned and normalized to obtain preprocessed data. Then, differential features are constructed and selected from the preprocessed data, and the input time series is extracted using a sliding window. The time series is then input into the GTCN-STransformer-based engineering vehicle load status recognition model for deep feature extraction. GTCN extracts local features, the SE mechanism models the inter-channel dependencies, and STransformer further extracts the global dependencies of local features containing channel dependencies. Finally, a fully connected network and a Softmax function are used to classify and identify the extracted deep features. Experimental results show that the GTCN-STransformer-based engineering vehicle load status recognition model with SE mechanism achieves higher accuracy, macro precision, macro recall, and F1 score in classifying and identifying the load status of engineering vehicles, improving the accuracy of vehicle status recognition and effectively monitoring problems such as arbitrary dumping and unauthorized transport during transportation.

[0045] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A method for identifying the load status of engineering vehicles, characterized in that, The following steps are included: S1. Obtain the raw transportation data of the engineering vehicles and perform data preprocessing on the raw transportation data, including data cleaning and normalization. S2. Perform differential feature construction and feature selection on the preprocessed data, extract the feature vector set, and use a sliding window to segment the data to obtain a fixed-length input sequence. S3. By embedding the SE mechanism, the GTCN-STransformer extracts local and global features containing feature channel dependencies from the input sequence to obtain a deep feature representation of the original input sequence. S4. The extracted deep features are used to classify and identify the load status of engineering vehicles using a fully connected network and the Softmax function, so as to obtain the load status of engineering vehicles.

2. The method for identifying the load status of engineering vehicles according to claim 1, characterized in that: In step S1, Acquire raw transportation data of engineering vehicles, including load, sensor deformation coefficient, direction, speed, vehicle voltage, longitude, latitude, altitude and data acquisition time; Time series data records with missing or duplicate entries in the original data are discarded, i.e., data cleaning. The Z-score normalization method is used to transform different feature data into dimensionless data to avoid large-scale features dominating the results of load status classification and recognition, thereby eliminating the influence of differences in the dimensions between features on load status classification and recognition.

3. The method for identifying the load status of engineering vehicles according to claim 2, characterized in that: In step S2, S21. Construct differential features for the load, direction, and speed features in the preprocessed data, and add the differential features to the original features to form an initial feature set; S22. Select features from the initial feature set using the maximum information coefficient (MIC) to obtain a feature vector set; S23. For the feature vector set, a sliding window is used to segment the data to obtain a fixed-length input sequence.

4. The method for identifying the load status of engineering vehicles according to claim 3, characterized in that: In step S22, Each feature and its corresponding label in the initial feature set are obtained sequentially to form a two-dimensional data point set; Map the set of two-dimensional data points onto a two-dimensional grid and calculate the corresponding mutual information values; Exhaustively enumerate all two-dimensional grid sizes and select the maximum mutual information under different two-dimensional grid sizes as the MIC score of the feature to measure the correlation between the feature and the load state; Repeat the above steps to obtain the MIC scores of all features in the initial feature set; Features with a MIC score greater than 0.4 are selected to form a feature vector set.

5. The method for identifying the load status of engineering vehicles according to claim 4, characterized in that: In step S3, S31. Extract local temporal features from the input sequence using grouped temporal convolution GTCN; S32. By using the squeeze excitation mechanism SE embedded in the temporal convolution GTCN, the interdependencies between local temporal feature channels extracted by the temporal convolution GTCN are captured. By assigning different weights to the feature channels, more attention is paid to the features that contribute more to the load state recognition task. Features of GTCN are extracted by integrating temporal convolutions through grouped shuffling temporal convolutions. S33. Global modeling of multidimensional temporal local features containing channel dependencies is performed using the Sparse Transformer. The sparsity mechanism in the Sparse Transformer reduces the amount of computation and reduces the impact of redundant information.

6. The method for identifying the load status of engineering vehicles according to claim 5, characterized in that: In step S31, When the extracted temporal local feature channels reach their maximum, i.e. the neck of the network, the temporal local features containing channel dependencies are further integrated through grouped shuffling temporal convolution in Temporal Convolution GTCN, while reducing computational complexity.

7. The method for identifying the load status of engineering vehicles according to claim 6, characterized in that: In step S32, The squeezing stage aggregates the feature information of each channel through global average pooling and generates a channel-by-channel local feature descriptor for each channel; The activation phase learns the relationship between local feature descriptors of different channels through two layers of nonlinear mapping, generates attention weights for each channel with global information, and enhances important feature channels by multiplying them with the original feature channels.

8. The method for identifying the load status of engineering vehicles according to claim 7, characterized in that: In the global modeling of step S33, The multidimensional local features of the input sparse transformer STransformer are positionally encoded using a sine-cosine function to generate a unique representation for each position in the local features. By inputting the multi-dimensional local feature representation containing location information into the multi-head sparse attention layer in the STransformer, the contextual dependencies of local features are learned to obtain the temporal global feature representation. At the same time, the computational complexity is reduced through sparse connections between self-attentions. The temporal global feature representation learned by the multi-head sparse attention layer is further fused at each location by applying a nonlinear mapping through the feedforward layer in the sparse transformer STransformer. Enhance the expressive power of features.

9. The method for identifying the load status of engineering vehicles according to claim 8, characterized in that: In step S4, the deep features extracted by the GTCN-STransformer with embedded SE mechanism are transformed into probabilities of three load states: loading, unloading, and non-loading / unloading, through a fully connected network and the Softmax function, so as to obtain the load state identification result.

Images