Mining truck automatic shifting control method, system, storage medium and computing device

Abstract

This invention relates to an automatic gear shifting control method, system, storage medium, and computing device for mining trucks. The method includes: acquiring offline state data of the mining truck and preprocessing it; sequentially generating an original MTD set; filtering out irrelevant and invalid data to generate an effective MTD set, and randomly dividing it into a training set and a test set; using the training set as input to a pre-established multi-parameter and time-span ResNet-Bi-LSTM-Attention network to train the ResNet-Bi-LSTM-Attention network; inputting the test set into the trained ResNet-Bi-LSTM-Attention network to verify its robustness, obtaining the final ResNet-Bi-LSTM-Attention network model; inputting the acquired offline state data of the mining truck into the model; outputting the final gear position feature; and completing the gear shifting operation. This invention solves the problem in existing technologies that easily overlook the integrity of temporal data context information and cannot effectively integrate the features before and after the data, and can be applied in the field of heavy vehicle gear shifting control.

Description

Technical Field

[0001] This invention relates to the field of heavy vehicle shift control technology, and in particular to an automatic shift control method, system, storage medium and computing device for mining trucks based on ResNet-Bi-LSTM-Attention network and multiple time spans. Background Technology

[0002] Driving mining trucks on rugged mining roads is a complex challenge. Therefore, vehicle design must consider not only driving performance but also economy, power, and driving comfort. Currently, many mining trucks still use manual transmissions, which increases the driver's workload and error rate after long hours of driving. Furthermore, irregular and incorrect gear shifting by the driver increases the probability of transmission damage. However, using an automatic transmission can effectively avoid these problems. Automatic transmissions are controlled by appropriate shift strategies, which reduce operational difficulty and improve shifting efficiency and dynamic performance. Optimizing the shift strategy is crucial in complex road conditions. The shift strategy automatically selects the appropriate gear based on road conditions, fuel economy, and power performance. Therefore, driving comfort is greatly improved, while driving difficulty is significantly reduced.

[0003] In recent years, due to a lack of disclosure from major transmission manufacturers, many scholars have studied the automatic shifting strategies of automatic transmissions. Many automatic shifting strategies in Automated Manual Transmission (AMT) vehicles are based on two or three parameters, such as pedal position, vehicle speed, and engine speed. Other potential factors, such as oil temperature, gear ratio, and output speed, are not used in their shifting strategies. Therefore, they cannot accurately judge the dynamic changes during vehicle operation and can affect vehicle performance under complex and variable road conditions. A comparison of the gear prediction accuracy based on two-parameter and multi-parameter shifting strategies shows that strategies with more vehicle state data can significantly improve gear prediction accuracy.

[0004] However, existing shift control strategies only consider the correlation between information at continuous time points while ignoring the correlation between data at discontinuous time points. This makes it difficult to effectively integrate the characteristics of data before and after the shift, thus failing to meet the requirements of high accuracy in gear prediction and real-time prediction time. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide an automatic gear shifting control method, system, storage medium, and computing device for mining trucks, which solves the problem that existing technologies tend to overlook the integrity of the contextual information of time-series data and cannot effectively integrate the characteristics of data before and after.

[0006] To achieve the above objectives, in a first aspect, the technical solution adopted by the present invention is as follows: an automatic gear shifting control method for mining trucks, comprising: acquiring offline state data of mining trucks and preprocessing it; sequentially generating an original MTD set; filtering out data unrelated to gear shifting and invalid data; generating an effective MTD set; and randomly dividing it into a training set and a test set; using the training set as input to a pre-established multi-parameter and time-span ResNet-Bi-LSTM-Attention network to train the ResNet-Bi-LSTM-Attention network; inputting the test set into the trained ResNet-Bi-LSTM-Attention network to verify the robustness of the ResNet-Bi-LSTM-Attention network; obtaining the final ResNet-Bi-LSTM-Attention network model; and inputting the acquired offline state data of mining trucks into the model to output the final gear position feature and complete the gear shifting operation.

[0007] Furthermore, offline status data of mining trucks is acquired and preprocessed, including:

[0008] The real-time mining truck status data is decoded to generate a raw MTD set with timing. The last column of the set, the gear value CG at the current time, is used as the true value for model training, and the remaining columns are used as potential parameters for gear prediction.

[0009] The H-index of the original MTD set was calculated using the R / S analysis method, and shift parameters with positive correlation were screened out by the H-index.

[0010] Outlier handling and data normalization are performed on the selected shift parameters with positive correlation to generate an effective MTD set composed of valid time data.

[0011] Furthermore, during the R / S analysis, all parameters except CG are used as inputs, and CG is considered the actual value of the model output; the formula for calculating the H-index is:

[0012] H = log n (R / S) n -log n (C)

[0013] in:

[0014]

[0015]

[0016] In the formula, A is the number of consecutive subintervals, M is the total length of the time series, n is the length of each subinterval, and (R / S) aLet C represent the rescaling range of the sub-interval sequence a, where C is a constant.

[0017] Furthermore, the H-index is used to screen for shift parameters with positive correlation, including:

[0018] If the H index is greater than the preset value, then this parameter has a positive correlation during gear shifting;

[0019] If the H index is equal to the preset value, this parameter has no effect on gear shifting;

[0020] If the H index is less than the preset value, this parameter has a negative correlation during gear shifting.

[0021] Furthermore, a pre-built ResNet-Bi-LSTM-Attention network with multiple parameters and time spans includes:

[0022] An improved Residual Network and Bi-LSTM network model is established. After feature fusion, an attention mechanism module is added, followed by full-connection dimensionality reduction, and finally the model features are output.

[0023] Furthermore, the ResNet-Bi-LSTM-Attention network is trained, including:

[0024] The input layer is a set of MTD data based on multiple time spans, and the input features consist of multiple vehicle state data.

[0025] Before inputting into Bi-LSTM and ResNet, a fully connected layer is used for dimensionality reduction. After dimensionality reduction, the features are simultaneously input into Bi-LSTM and ResNet networks for computation.

[0026] In the Bi-LSTM structure, a hidden layer is first input, and various hyperparameters are set. The batch size is manually set before each training iteration. At time step t+n, a bidirectional LSTM layer is input to encode the time series features, and the resulting hidden layer state values ​​h1, h2, ..., h... are then processed. n The results are fed into a hidden layer, and the result is multiplied by the training step size to calculate the value of the current hidden layer state. Random deactivation is added to alleviate the overfitting problem in model training, and the output features of the hidden state at time t+n of Bi-LSTM are output.

[0027] During the training of the ResNet network, a convolutional layer is first input to reduce dimensionality and decrease the overall structural complexity. A 1x1 convolutional kernel is then connected to the first layer, with a normalization function defined for the input channels and ReLU activation set as the default. A 3x3 convolutional kernel is then connected to the second layer, followed by a 1x1 convolutional kernel. A shortcut structure is added, allowing the input layer to be directly connected to the output layer via weighted input. The computation results are then fed into a global average pooling layer for pooling operations, ultimately yielding the computational values ​​of the ResNet layer.

[0028] By using gated residual connections, the results of the hidden layers after dimensionality reduction of Bi-LSTM and ResNet are concatenated, and then dimensionality is reduced through a fully connected network.

[0029] By using a gating mechanism, the initial input features are combined with the total output value of the hidden state as the input to the next layer of the network, and the two hidden layers are connected to fully extract features.

[0030] If the loss calculated by the model trained in the current iteration is less than that of the model trained in the previous epoch, training ends and the current model is saved; otherwise, training is terminated early and the best current model is saved.

[0031] The final gear characteristics are output from the input output layer and the SoftMax layer.

[0032] Furthermore, the robustness of the ResNet-Bi-LSTM-Attention network was verified, including:

[0033] Load the weight file that yielded the best results during training for each network structure, input a test set with the same parameters, and calculate the time to predict a single gear.

[0034] Record the minimum, average, and maximum values ​​of the time spent processing a single gear data point on the test set. Repeat the operation multiple times on the same test set and take the average value as the final result.

[0035] Secondly, the technical solution adopted by the present invention is as follows: an automatic gear shifting control system for mining trucks, comprising: a first processing module, which acquires offline state data of mining trucks and preprocesses it, sequentially generates an original MTD set, filters out data that is irrelevant to gear shifting and invalid data, generates an effective MTD set, and randomly divides it into a training set and a test set; a second processing module, which uses the training set as input to a pre-established multi-parameter and time-span ResNet-Bi-LSTM-Attention network to train the ResNet-Bi-LSTM-Attention network; and an output module, which inputs the test set into the trained ResNet-Bi-LSTM-Attention network to verify the robustness of the ResNet-Bi-LSTM-Attention network, obtains the final ResNet-Bi-LSTM-Attention network model, inputs the acquired offline state data of mining trucks into the model, outputs the final gear position features, and completes the gear shifting operation.

[0036] Thirdly, the technical solution adopted by the present invention is: a computer-readable storage medium for storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0037] Fourthly, the technical solution adopted by the present invention is: a computing device comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described above.

[0038] The present invention has the following advantages due to the adoption of the above technical solutions:

[0039] 1. The present invention is based on an automatic gear shifting system for mining trucks using a ResNet-Bi-LSTM-Attention network with multiple parameters and time span. Compared with RNN, LSTM, Bi-LSTM, ResNet-Bi-LSTM and Bi-LSTM-Attention networks, it can not only consider the correlation between continuous and non-continuous time information at the same time, but also make the training focus more on the feature vector with larger weights, thereby extracting more robust and less noisy features.

[0040] 2. When parameters at nearby times become abnormal due to external disturbances, the accuracy of gear prediction using a single time span calculation will be inaccurate. The multi-time span estimation of this invention can capture numerical information over a longer time span and better prevent deviations in gear prediction due to abnormal single-value readings. Furthermore, this invention also performs training with multiple sets of different batch sizes to find the optimal performance result during network training.

[0041] 3. By comparing the performance metrics of several network models, including RNN, LSTM, Bi-LSTM, ResNet-Bi-LSTM, and Bi-LSTM-Attention, experimental results show that the ResNet-Bi-LSTM-Attention network of this invention, while ensuring that the average time for predicting a single gear meets the actual driving needs of drivers, also exhibits rapid convergence and a minimum loss function, achieving the highest accuracy. Therefore, the performance of this invention is significantly improved compared to other models. Attached Figure Description

[0042] Figure 1 This is a flowchart of the automatic gear shifting control method for mining trucks in an embodiment of the present invention;

[0043] Figure 2 This is a flowchart of the data preprocessing process in an embodiment of the present invention;

[0044] Figure 3 This is a diagram showing the outlier processing results in an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the Bi-LSTM model structure in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0047] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0048] Some gear-shifting strategies, based on complex dynamic mathematical models, can adjust the gear to the appropriate position, but cannot guarantee real-time processing. In recent years, with the rapid development of machine learning technology, many researchers have applied it to the field of gear-shifting strategies, achieving fruitful results. One existing literature-based automatic gear-shifting system uses three dynamic parameters as input: vehicle speed, vehicle acceleration, and pedal position. To handle complex driving conditions, a correction module is added to the network, but it still faces the challenge of processing data with time-varying characteristics. The processing of nonlinear features in time series data has some shortcomings, which can be addressed through deep learning time series modeling. Thus, deep learning technology is widely used to process time series data with nonlinear characteristics, overcoming the shortcomings of traditional neural networks.

[0049] To better handle sequential information, a more efficient method for initializing weights is applied to deep learning. To overcome the limitations of traditional recurrent neural networks (RNNs) in processing temporal historical data and their susceptibility to gradient vanishing or exploding, Long Short-Term Memory (LSTM) networks are employed. A bidirectional LSTM network (Bi-LSTM) is used to develop an automatic shifting strategy based on a mining truck, achieving high prediction accuracy. However, Bi-LSTM neural networks tend to ignore the completeness of contextual information in temporal data. Therefore, many researchers have added convolutional neural networks (CNNs) to compensate for the insufficient feature extraction of Bi-LSTM networks. By fusing Bi-LSTM networks with CNNs, BLSTM-2DPooling and BLSTM-2DCNN models have been proposed, considering not only the dimension of the text feature vector but also the dimension of the time step. Experimental results show that this network can capture richer semantic features. However, CNNs only consider the correlation between information at consecutive time points and ignore the correlation between data at discontinuous time points. Residual Networks (ResNet) are used to train deeper CNN models. ResNet adds an identity mapping structure to CNN, skipping computations in the current or multiple layers. During backpropagation, gradients from the next layer are directly passed to the previous layer. This not only solves the problem of network degradation or gradient explosion caused by simply stacking depth in CNNs, but also avoids adding extra computation to the entire network. Therefore, ResNet is used instead of stacking CNN and Bi-LSTM models. Existing literature presents a network structure combining Bi-LSTM and ResNet for ECG heartbeat classification, achieving good results, reaching 96.2% accuracy on the MIT dataset. Attention mechanisms are widely used in various fields. Attention mechanisms calculate attention probabilities for different features, assigning different weights to different features in the model, thus making training focus more on feature vectors with larger weights and better utilizing the extracted feature information.

[0050] To address the problem that existing technologies often overlook the integrity of temporal data context and fail to effectively integrate features from different periods, this invention proposes an automatic gear shifting control method, system, storage medium, and computing device for mining trucks. This invention utilizes a multi-parameter and time-span ResNet-Bi-LSTM-Attention network for automatic gear shifting in mining trucks. Employing multiple parameters as input features, it assists the driver in predicting gear shifts at future moments and is applicable to other traffic environments. To evaluate the robustness of the ResNet-Bi-LSTM-Attention network model, it was compared with Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Bidirectional Long Short-Term Memory Networks (Bi-LSTMs), Residual Neural Networks (ResNet), and Attention-Bi-LSTM networks. Experimental results show that the multi-parameter and time-span ResNet-Bi-LSTM-Attention network used in this invention outperforms other networks, ensuring high accuracy in gear prediction while meeting the real-time requirements of prediction.

[0051] In one embodiment of the present invention, an automatic gear shifting control method for mining trucks is provided. In this embodiment, the method is implemented based on a multi-parameter and time-span ResNet-Bi-LSTM-Attention architecture, specifically, as follows... Figure 1 As shown, the method includes the following steps:

[0052] 1) Obtain offline data of mining trucks and preprocess it to generate the original MTD (mining truck data) set in sequence. Filter out data that is irrelevant to gear shifting and invalid data to generate an effective MTD set, and randomly divide it into training set and test set.

[0053] In this embodiment, in order to obtain vehicle offline status data, four ZLG CANDTU-200UR CAN bus cards are pre-installed on four mining trucks.

[0054] The MTD (Motor Distribution Data) is preprocessed to create a sequential list of data, filtering out irrelevant and invalid data. The purpose of data preprocessing is to identify parameters positively correlated with shifting, handle invalid data, and reduce the impact of invalid data on model training.

[0055] 2) Train the ResNet-Bi-LSTM-Attention network by using the training set as input to the pre-built ResNet-Bi-LSTM-Attention network with multiple parameters and time spans;

[0056] 3) Input the test set into the trained ResNet-Bi-LSTM-Attention network to verify the robustness of the ResNet-Bi-LSTM-Attention network, obtain the final ResNet-Bi-LSTM-Attention network model, and input the obtained offline state data of the mining truck into the model to output the final gear feature and complete the gear shifting operation.

[0057] In this embodiment, a shift control system for automatic gear shifting is also included between steps 1). This shift control system includes a transmission, a control lever, and a transmission control unit (TCU). The transmission performs a gear shift upon receiving a control signal from the TCU and sends transmission status information to other monitoring devices via a CAN bus. The aforementioned automatic gear shifting control method for mining trucks is integrated within the TCU.

[0058] The TCU that controls gear shifting is installed on the mining truck. It monitors the status data of the mining truck through various sensors and shares it with other relevant electronic control units (ECUs) via the CAN bus, including accelerator pedal position, current gear, engine speed, oil temperature, working pressure, transmission output speed, and system errors.

[0059] When a shift command signal is received from the control lever, the TCU controls the shifting operation. Before shifting, appropriate safety checks are performed, and then the TCU executes upshifts, downshifts, and neutral at the appropriate time to protect the transmission, reduce fuel costs, and maximize vehicle performance.

[0060] In this embodiment, the embedded device on the test vehicle is an NVIDIA Jetson AGX Xavier, and detailed parameter information is shown in Table 1.

[0061] Table 1 Parameter Information

[0062]

[0063] In this embodiment, four ZLG CANDTU-200UR CAN bus cards are installed on four mining trucks to acquire vehicle status data. During manual driving, vehicle data processed and monitored by different ECUs on the mining trucks is transmitted via the CAN bus. Subsequently, the CAN bus cards automatically begin data recording for further development of machine learning models. The raw CAN bus data follows the SAE J1939 protocol, which is widely used in the truck industry. Through a data decoding algorithm, 14 types of vehicle parameters are obtained from the Parameter Group Number (PGN), and each parsed vehicle parameter data also has a timestamp to mark the reception time.

[0064] The experimental mining truck with a manual transmission measures 7.65 meters long, 3.5 meters wide, and 3.8 meters high, and weighs 20 tons. The mining area experiences few road emergencies, has simple traffic conditions, and minimal traffic regulations because it is located in a closed area with low traffic and population density.

[0065] In step 1) above, the offline status data of the mining truck is acquired and preprocessed, such as... Figure 2 As shown, it includes the following steps:

[0066] 1.1) Decode the real-time acquired mining truck status data to generate a raw MTD set with timing. The last column in the set represents the gear value CG at the current time, which is used as the true value for model training. The remaining columns are used as potential parameters for gear prediction and are updated row by row when receiving raw CAN data from the mining truck.

[0067] In one feasible embodiment, a total of 1,020,301 raw CAN data entries are decoded using the SAE J1939 protocol to generate an MTD set with timing; the MTD set consists of 14 columns of data, corresponding to 14 parameter values ​​that can be obtained from the CAN bus.

[0068] 1.2) In order to screen out the parameters with positive correlation, the H index (Hurst index) of the original MTD set is calculated by the R / S analysis method, and the shift parameters with positive correlation are screened out by the H index.

[0069] During the R / S analysis, all parameters except CG are used as inputs, and CG is considered the actual value of the model output; the formula for calculating the H-index is:

[0070]

[0071]

[0072]

[0073] R a =max(X k,a )-min(X k,a (4)

[0074]

[0075]

[0076]

[0077]

[0078] In the formula, A is a time series R with a total length of M.t The number of equal subintervals, where the length of each subinterval is n, (R / S) a I represents the rescaling range of the sub-interval sequence a, where C is a constant; a Let R represent a sub-interval sequence a (a∈[1,2,3,…,A]). k,a Indicate I a The elements in R (k∈[1,2,3,…,n]) represent the average value of the sequence a in the subinterval; i,a Let X represent the k elements in the sub-interval sequence a. k,a R represents the cumulative deviation; a It is X k,a The range, max(X) k,a ) and min(X k,a ) is X k,a Maximum and minimum cumulative deviation; S a It is the standard deviation, (R / S) a Represents the rescaling range of the sub-interval sequence a, (R / S) n It is the average of the rescaled ranges of all sub-intervals; H is the Hearst exponent.

[0079] 1.3) Perform outlier handling and data normalization on the selected shift parameters with positive correlation to generate an effective MTD set composed of valid time data.

[0080] In step 1.2) above, the H index value reflects the importance of certain parameters during gear shifting. Gear shifting parameters with positive correlation are selected using the H index, specifically:

[0081] If the H index is greater than the preset value, then this parameter has a positive correlation during gear shifting;

[0082] If the H index is equal to the preset value, this parameter has no effect on gear shifting;

[0083] If the H index is less than the preset value, this parameter has a negative correlation during gear shifting.

[0084] The experimental results show that the variables with positive correlation are shown in Table 2.

[0085] Table 2 shows the H-index calculation results for gear-related variables.

[0086]

[0087] In step 1.3) above, outlier handling and data normalization are specifically as follows:

[0088] Due to electromagnetic and other environmental factors, data from sensors may be affected and become anomaly-prone during MTD collection. A block diagram for outlier detection is used to identify outliers and ensure data quality. Figure 3 In the diagram, the size of the box is the interquartile range (IQR), and the upper and lower ranges of the box are the upper quartile (Q3) and lower quartile. The upper and lower edges are the UE and LE, which are calculated by equations (9) and (10). Parameter values ​​of UE or LE that are outside the corresponding category are considered outliers and marked with “○”. Figure 3 In the ES (Executable Array) module, since the parameter data from the CAN bus is continuous and the intervals are very short (1 to 10 Hz per second), outliers are replaced with the average data from the previous and next frames to ensure data consistency.

[0089] UE=Q3+1.5IQR (9)

[0090] LE = Q1 - 1.5IQR (10)

[0091] The results show that there are no extreme outliers overall.

[0092] Since the MTD dataset consists of valid time data but is not normalized, to ensure that all data in the MTD dataset is mapped to the same scale and distribution range for subsequent machine learning model training, the original data x in the MTD dataset is subjected to min-max normalization:

[0093]

[0094] In the formula, x′ represents the processed data, min(x) represents the minimum value of x, and max(x) represents the maximum value of x.

[0095] In step 2) above, the pre-established multi-parameter and time-span ResNet-Bi-LSTM-Attention network includes: establishing a Residual Network fused with a Bi-LSTM network model, adding an attention mechanism module after feature fusion, then connecting a fully connected layer for dimensionality reduction, and finally outputting model features. On the one hand, it improves the existing Residual Network by reducing the number of layers in its residual network, and on the other hand, it improves the Loss function to Focal Loss and performs smoothing optimization on the labels, thereby improving the problem of class imbalance.

[0096] 2.1.1) Bi-LSTM Network Structure. The Bi-LSTM structure adds a reverse output and a general output to the LSTM network, enabling the acquisition of bidirectional historical information, such as... Figure 4 As shown. This includes the input gate i for inputting valid information into the structure.t The output gate is used to output the filtered cell information to the next time step. t And the forgetting gate f, which filters out information that is useless for prediction while selecting effective information. t C t It is a cellular memory state, W f W i W o and W c These are the weight matrices for the three gates and the candidate cell states, x t For a total dataset with temporal sequence characteristics, h t b is the state value of the hidden layer at time t. f b i b o and b c This represents the corresponding offset, where σ is the activation function of the three gates, and T is the threshold of the total time series. Let tanh be the activation function for the candidate states, and c be the activation function. t This represents the current state value of the cell. These are the values ​​of neurons in the feedforward hidden layer. y represents the value of the backward hidden layer neuron. t The final eigenvalue is given by the following formula:

[0097] f t =σ(W f ·x t +W f ·h t-1 +b f (12)

[0098]

[0099] i t =σ(W i ·[h t-1 ,x t ]+b i (14)

[0100] o t =σ(W o ·[x t ,h t-1 ]+b o (15)

[0101]

[0102] h t =o t ·tanh(c t (17)

[0103]

[0104]

[0105]

[0106] 2.1.2) ResNet Network Structure. In the residual network, an identity mapping method is used to construct the deep model, where x... (l) For the historical information learned by the current l-level residual block, x (l+1) For the output of layer l, W (l) It is the weight of the l-th layer, F(x) (l) W (l) H(x) represents the fitted residual mapping. (l) The input is mapped to the summed network, and its calculation formula is shown below:

[0107] H(x (l) )=F(x (l) W (l) )+x (l) (twenty one)

[0108] When F(x) (l) W (l) When ) is 0, it constitutes the optimal identity mapping:

[0109] H(x (l) )=x (l) (twenty two)

[0110] The historical information obtained by a residual network of depth L is generated by stacking each layer l shallower than it, where x (l) W represents the input of the l-th residual block. (i) x represents the weight of each layer. (l+1) The output of this residual block is also the input of the (l+1)th residual block. Therefore, the output x of the deeper L residual blocks can be calculated by accumulating the outputs of each residual block. (L) :

[0111]

[0112] As can be seen from the formula, any unit L and l have residual characteristics. The residual network features are formed by the sum of the residual features of each layer. The l+1 layer can incorporate more feature information than the l-th layer.

[0113] During backpropagation, according to the chain rule, the loss function for the l-th residual block x (l) The gradient calculation formula is:

[0114]

[0115] In the formula, This indicates that the gradient of layer L can be directly propagated to any shallower network layer l. This indicates that its value will not be constant at -1 during backpropagation, thus avoiding the gradient vanishing problem. Therefore, the performance degradation problem caused by stacking multiple network layers can be solved by minimizing the residual mapping F(x).

[0116] 2.1.3) Attention network structure. h n This represents the hidden layer state output value at time n, which is the result of concatenating the residual network and the Bi-LSTM network after dimensionality reduction. Let β be the feature representation vector, which is one level higher than the input variables, and let h be the output of the nth hidden layer. n exist The weight of the input at that moment in the overall score is determined by β. A larger β indicates a higher weight of the input in the overall score. The formula for calculating β is as follows:

[0117]

[0118] Where b is the bias term, V, W, and U are all weight matrices, and tanh is a nonlinear activation function;

[0119] ..., This represents the proportion of the total score to the score achieved by matching each feature vector with the output of the hidden layer. The calculation formula is as follows:

[0120]

[0121] Calculate the final feature F based on the attention distribution att :

[0122]

[0123] After obtaining the eigenvector F att Then, the probability distribution y of the classification labels is calculated using the softmax function of the output layer:

[0124]

[0125]

[0126] Where V represents the weight matrix of the model output layer, F′ att This represents the weighted F. att F′ att(i) Represents vector F′ att The i-th component value has a vector length equal to the number of category labels, where T is the number of category labels.

[0127] Finally, to ensure smoother convergence of the model's loss function, cross-entropy is used as the loss function. The cross-entropy loss E(Y,y) is calculated by comparing the probability distribution y with the true class distribution Y:

[0128]

[0129] Step 2) above involves training the ResNet-Bi-LSTM-Attention network, including the following steps:

[0130] 2.2.1) The input layer is a set of MTD data based on multiple time spans, and the input features consist of multiple vehicle state data;

[0131] For example, the input layer is a set of data from x. t-k To x t The MTD data consists of 9 vehicle state data points over an appropriate time span, including ACPP, OSS, AEPT, RPT, ES, AGR, TSOT, TCR, and DRT.

[0132] 2.2.2) Since the features themselves have a large dimension, a fully connected layer is used to reduce the dimension before inputting them into Bi-LSTM and ResNet, so as to obtain an overall parameter size suitable for the model; after the dimension reduction, the features are simultaneously input into Bi-LSTM and ResNet networks for computation.

[0133] 2.2.3) In the Bi-LSTM structure, a hidden layer is first input, and the hyperparameters are set. The batch size is manually set before each training iteration. A bidirectional LSTM layer is input at time step t+n to encode the time series features, mitigating the problem that the LSTM model tends to only capture features from the latest time step. The resulting hidden layer outputs features h1, h2, ..., h... n The results are fed into a hidden layer, and the result is multiplied by the training step size to calculate the value of the current hidden layer state. Random deactivation is added to alleviate the overfitting problem in model training, and the output features of the hidden state at time t+n of Bi-LSTM are output.

[0134] For example, in the Bi-LSTM structure, a hidden layer of size 32 is first input, with a random inactivation rate of 0.22, a learning rate of 0.005, and 500 iterations. The batch size will be manually set before each training session and the optimal value will be sought in the end.

[0135] 2.2.4) During the training of the ResNet network, a convolutional layer is first input to reduce the dimensionality again and decrease the complexity of the overall structure; a 1*1 convolutional kernel is then connected to the first layer, and a normalization function is defined for the input channels, while the activation function is set to ReLU, and other parameters are left as default; a 3*3 convolutional kernel is then connected to the second layer and a 1*1 convolutional kernel is then connected to the third layer in sequence; a shortcut structure is added so that the input layer is directly connected to the output layer through weighted summation, which alleviates the gradient divergence problem in deep networks; to avoid overfitting and improve training speed, the calculation results are input into a global average pooling layer for pooling operations, and finally the calculated values ​​of the ResNet layer are obtained;

[0136] In this embodiment, the parameter settings for the second and third layers are the same as those for the first layer.

[0137] 2.2.5) Using gated residual connections, the results of the hidden layers of Bi-LSTM and ResNet after dimensionality reduction are concatenated, and then dimensionality is reduced through a fully connected network to maintain the original vector dimension of the hidden state. Then, an attention mechanism is used to perform weighted operations on the important features.

[0138] 2.2.6) By using a gating mechanism, the initial input features are combined with the total output value of the hidden state as the input to the next layer of the network. This allows for the full extraction of features from two hidden layers, thus enabling better segmentation of the data features.

[0139] 2.2.7) If the loss calculated by the model trained in the current iteration is less than that of the model trained in the previous epoch, end the training and save the current model; otherwise, terminate the training early and save the best current model.

[0140] 2.2.8) The final gear position feature is output from the input output layer and the SoftMax layer;

[0141] 2.2.9) Repeat the above process, output and save the best result.

[0142] In step 3) above, to verify the robustness of the ResNet-Bi-LSTM-Attention network, experiments were conducted on a server with the same environment configuration as the vehicle embedded system (NVIDIA Jetson AGX Xavier). A total of 900,000 valid MTDs were used to train the machine learning model, and another 100,000 were used for validation. To ensure the reasonableness of the results, each model was trained 10 times under different parameters, and the final result was the average of all results.

[0143] In this embodiment, the robustness of the ResNet-Bi-LSTM-Attention network is verified, including the following steps:

[0144] 3.1) Load the weight file that yielded the best results during training for each network structure, input the test set with the same parameters, and calculate the time to predict a single gear.

[0145] Specifically, the initial time t0 is the time from when the model is loaded, and the final prediction time t1 is the time after the prediction ends. The time difference between the two is the prediction time for a single gear. At the same time, other models are set as a reference.

[0146] 3.2) Record the minimum (MIN), average (AVG), and maximum (MAX) cost of processing a single gear data point on the test set. To obtain more accurate prediction results, repeat the operation multiple times on the same test set and take the average value as the final result.

[0147] Comparative experiments with various models verified that the model of this invention has the highest robustness. Simultaneously, the automatic gear shifting control method was input into the control lever to ultimately complete the gear shifting operation.

[0148] In one embodiment of the present invention, an automatic gear shifting control system for a mining truck is provided, comprising:

[0149] The first processing module acquires offline status data of mining trucks and preprocesses it, sequentially generates the original MTD set, filters out data that is irrelevant to gear shifting and invalid data, generates the effective MTD set, and randomly divides it into training set and test set.

[0150] The second processing module uses the training set as input to a pre-built ResNet-Bi-LSTM-Attention network with multiple parameters and time spans to train the ResNet-Bi-LSTM-Attention network.

[0151] The output module takes the test set as input to the trained ResNet-Bi-LSTM-Attention network to verify the robustness of the ResNet-Bi-LSTM-Attention network, obtains the final ResNet-Bi-LSTM-Attention network model, and inputs the locally stored mining truck state data into the model to output the final gear position feature and complete the gear shifting operation.

[0152] In the first processing module described above, the offline status data of the mining truck is acquired and preprocessed, including:

[0153] The real-time mining truck status data is decoded to generate a raw MTD set with timing. The last column of the set, the gear value CG at the current time, is used as the true value for model training, and the remaining columns are used as potential parameters for gear prediction.

[0154] The H-index of the original MTD set was calculated using the R / S analysis method, and shift parameters with positive correlation were screened out by the H-index.

[0155] Outlier handling and data normalization are performed on the selected shift parameters with positive correlation to generate an effective MTD set composed of valid time data.

[0156] During the R / S analysis, all parameters except CG are used as inputs, and CG is considered the actual value of the model output; the formula for calculating the H-index is:

[0157] H = log n (R / S) n -log n (C)

[0158] in:

[0159]

[0160]

[0161] In the formula, A is the number of consecutive sub-intervals into which a time series of length M is divided, n is the length of each sub-interval, and (R / S) a Let C represent the rescaling range of the sub-interval sequence a, where C is a constant.

[0162] In this embodiment, the shift parameters with positive correlation are selected by using the H-index, specifically as follows:

[0163] If the H index is greater than the preset value, then this parameter has a positive correlation during gear shifting;

[0164] If the H index is equal to the preset value, this parameter has no effect on gear shifting;

[0165] If the H index is less than the preset value, this parameter has a negative correlation during gear shifting.

[0166] In the second processing module of the above embodiment, the pre-established multi-parameter and time-span ResNet-Bi-LSTM-Attention network includes:

[0167] First, a Residual Network fused with a Bi-LSTM network model is established to obtain local features. Long-sequence features can be extracted from the input data. In view of the fact that some categories of the dataset have fewer features, the existing Residual Network is improved by reducing the number of stacked layers of its residual network. At the same time, the loss function is improved to Focal Loss, and the labels are smoothed and optimized to improve the class imbalance problem. Then, an attention mechanism module is added to assign different weights to the features so that the network can focus more on important information. Next, a fully connected layer is added for dimensionality reduction. Finally, the model features are fused and output.

[0168] In the second processing module of the above embodiment, training the ResNet-Bi-LSTM-Attention network includes:

[0169] The input layer is a set of MTD data based on multiple time spans, and the input features consist of multiple vehicle state data.

[0170] Before inputting into Bi-LSTM and ResNet, a fully connected layer is used for dimensionality reduction. After dimensionality reduction, the features are simultaneously input into Bi-LSTM and ResNet networks for computation.

[0171] In the Bi-LSTM structure, a hidden layer is first input, and various hyperparameters are set. The batch size is manually set before each training iteration. At time step t+n, a bidirectional LSTM layer is input to encode the time series features. The resulting hidden layer features h1, h2, ..., h... are then processed. n The results are fed into a hidden layer, and the result is multiplied by the training step size to calculate the value of the current hidden layer state. Random deactivation is added to alleviate the overfitting problem in model training, and the output features of the hidden state at time t+n of Bi-LSTM are output.

[0172] During the training of the ResNet network, a convolutional layer is first input to reduce dimensionality and decrease the overall structural complexity. A 1x1 convolutional kernel is then connected to the first layer, with a normalization function defined for the input channels and ReLU activation set as the default. A 3x3 convolutional kernel is then connected to the second layer, followed by a 1x1 convolutional kernel. A shortcut structure is added, allowing the input layer to be directly connected to the output layer via weighted input. The computation results are then fed into a global average pooling layer for pooling operations, ultimately yielding the computational values ​​of the ResNet layer.

[0173] By using gated residual connections, the results of the hidden layers after dimensionality reduction of Bi-LSTM and ResNet are concatenated, and then dimensionality is reduced through a fully connected network.

[0174] By using a gating mechanism, the initial input features are combined with the total output value of the hidden state as the input to the next layer of the network, and the two hidden layers are connected to fully extract features.

[0175] If the loss calculated by the model trained in the current iteration is less than that of the model trained in the previous epoch, training ends and the current model is saved; otherwise, training is terminated early and the best current model is saved.

[0176] The final gear characteristics are output from the input output layer and the SoftMax layer.

[0177] The above embodiments verify the robustness of the ResNet-Bi-LSTM-Attention network in the output module, including:

[0178] Load the weight file that yielded the best results during training for each network structure, input a test set with the same parameters, and calculate the time to predict a single gear.

[0179] Record the minimum, average, and maximum values ​​of the time spent processing a single gear data point on the test set. Repeat the operation multiple times on the same test set and take the average value as the final result.

[0180] The system provided in this embodiment is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0181] A schematic diagram of a computing device structure is provided in one embodiment of the present invention. The computing device can be a terminal, which may include: a processor, a communication interface, memory, a display screen, and an input device. The processor, communication interface, and memory communicate with each other via a communication bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and a computer program. When the computer program is executed by the processor, it implements an automatic gear shifting control method for mining trucks. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, a management network, NFC (Near Field Communication), or other technologies. The display screen can be a liquid crystal display or an e-ink display. The input device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad mounted on the casing of the computing device, or an external keyboard, touchpad, or mouse, etc. The processor can call logical instructions in memory to execute the following methods: acquire offline state data of mining trucks and preprocess it, sequentially generate the original MTD set, filter out data that is irrelevant to gear shifting and invalid data, generate the effective MTD set, and randomly divide it into training set and test set; use the training set as the input of a pre-established multi-parameter and time-span ResNet-Bi-LSTM-Attention network to train the ResNet-Bi-LSTM-Attention network; input the test set into the trained ResNet-Bi-LSTM-Attention network to verify the robustness of the ResNet-Bi-LSTM-Attention network, obtain the final ResNet-Bi-LSTM-Attention network model, input the locally stored mining truck state data into the model, output the final gear position feature, and complete the gear shifting operation.

[0182] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0183] In one embodiment of the present invention, a computer program product is provided, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, when the program instructions are executed by a computer, the computer can execute the methods provided in the above-described method embodiments, for example including: acquiring offline state data of mining trucks and preprocessing it, sequentially generating an original MTD set, filtering out data that is irrelevant to gear shifting and invalid data, generating an effective MTD set, and randomly dividing it into a training set and a test set; using the training set as input to a pre-established multi-parameter and time-span ResNet-Bi-LSTM-Attention network, training the ResNet-Bi-LSTM-Attention network; inputting the test set into the trained ResNet-Bi-LSTM-Attention network, verifying the robustness of the ResNet-Bi-LSTM-Attention network, obtaining the final ResNet-Bi-LSTM-Attention network model, and inputting locally stored mining truck state data into the model, outputting the final gear position feature, and completing the gear shifting operation.

[0184] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided, which stores server instructions that cause a computer to execute the methods provided in the above embodiments, such as: acquiring offline status data of mining trucks and preprocessing it; sequentially generating an original MTD set; filtering out data that is irrelevant to gear shifting and invalid data; generating an effective MTD set and randomly dividing it into a training set and a test set; using the training set as input to a pre-established ResNet-Bi-LSTM-Attention network with multiple parameters and time spans to train the ResNet-Bi-LSTM-Attention network; inputting the test set into the trained ResNet-Bi-LSTM-Attention network to verify the robustness of the ResNet-Bi-LSTM-Attention network, obtaining the final ResNet-Bi-LSTM-Attention network model, and inputting locally stored mining truck status data into the model to output the final gear position feature and complete the gear shifting operation.

[0185] The computer-readable storage medium provided in the above embodiments has a similar implementation principle and technical effect to the above method embodiments, and will not be described again here.

[0186] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0187] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0188] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0189] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic gear shifting control method for mining trucks, characterized in that, include: Acquire offline status data of mining trucks and preprocess it to generate the original MTD set in sequence. Filter out data that is irrelevant to gear shifting and invalid data to generate the effective MTD set, and randomly divide it into training set and test set. The ResNet-Bi-LSTM-Attention network is trained by using the training set as input to a pre-built ResNet-Bi-LSTM-Attention network with multiple parameters and time spans. The test set is input into the trained ResNet-Bi-LSTM-Attention network to verify the robustness of the ResNet-Bi-LSTM-Attention network and obtain the final ResNet-Bi-LSTM-Attention network model. The obtained offline state data of the mining truck is then input into the model to output the final gear feature and complete the gear shifting operation. Acquire offline status data of mining trucks and perform preprocessing, including: The real-time mining truck status data is decoded to generate a raw MTD set with timing. The last column of the set, the gear value CG at the current time, is used as the true value for model training, and the remaining columns are used as potential parameters for gear prediction. The H-index of the original MTD set was calculated using the R / S analysis method, and shift parameters with positive correlation were screened out by the H-index. Outlier handling and data normalization are performed on the selected shift parameters with positive correlation to generate an effective MTD set composed of valid time data. Training the ResNet-Bi-LSTM-Attention network includes: The input layer is a set of MTD data based on multiple time spans, and the input features consist of multiple vehicle state data. Before inputting into Bi-LSTM and ResNet, a fully connected layer is used for dimensionality reduction. After dimensionality reduction, the features are simultaneously input into Bi-LSTM and ResNet networks for computation. In the Bi-LSTM structure, a hidden layer is first input, and various hyperparameters are set. The batch size is manually set before each training iteration. At time step t+n, a bidirectional LSTM layer is input to encode the time series features, and the resulting hidden layer state values ​​h1, h2, ..., h... are then processed. n The results are fed into a hidden layer, and the result is multiplied by the training step size to calculate the value of the current hidden layer state. Random deactivation is added to alleviate the overfitting problem in model training, and the output features of the hidden state at time t+n of Bi-LSTM are output. During the training of the ResNet network, a convolutional layer is first input to reduce dimensionality and decrease the overall structural complexity. A 1x1 convolutional kernel is then connected to the first layer, with a normalization function defined for the input channels and ReLU activation set as the default. A 3x3 convolutional kernel is then connected to the second layer, followed by a 1x1 convolutional kernel. A shortcut structure is added, allowing the input layer to be directly connected to the output layer via weighted input. The computation results are then fed into a global average pooling layer for pooling operations, ultimately yielding the computational values ​​of the ResNet layer. By using gated residual connections, the results of the hidden layers after dimensionality reduction of Bi-LSTM and ResNet are concatenated, and then dimensionality is reduced through a fully connected network. By using a gating mechanism, the initial input features are combined with the total output value of the hidden state as the input to the next layer of the network, and the two hidden layers are connected to fully extract features. If the loss calculated by the model trained in the current iteration is less than that of the model trained in the previous epoch, training ends and the current model is saved; otherwise, training is terminated early and the best current model is saved. The final gear characteristics are output from the input output layer and the SoftMax layer.

2. The automatic gear shifting control method for mining trucks as described in claim 1, characterized in that, During the R / S analysis, all parameters except CG are used as inputs, and CG is considered the actual value of the model output; the formula for calculating the H-index is: H=log n (R / S) n -log n (C) in: In the formula, A is the number of consecutive subintervals, M is the total length of the time series, n is the length of each subinterval, and (R / S) a Let C represent the rescaling range of the sub-interval sequence a, where C is a constant.

3. The automatic gear shifting control method for mining trucks as described in claim 1, characterized in that, The H-index is used to screen for shift parameters with positive correlation, including: If the H index is greater than the preset value, then this parameter has a positive correlation during gear shifting; If the H index is equal to the preset value, this parameter has no effect on gear shifting; If the H index is less than the preset value, this parameter has a negative correlation during gear shifting.

4. The automatic gear shifting control method for mining trucks as described in claim 1, characterized in that, A pre-built ResNet-Bi-LSTM-Attention network with multiple parameters and time spans includes: An improved Residual Network and Bi-LSTM network model is established. After feature fusion, an attention mechanism module is added, followed by full-connection dimensionality reduction, and finally the model features are output.

5. The automatic gear shifting control method for mining trucks as described in claim 1, characterized in that, Verifying the robustness of the ResNet-Bi-LSTM-Attention network includes: Load the weight file that yielded the best results during training for each network structure, input a test set with the same parameters, and calculate the time to predict a single gear. Record the minimum, average, and maximum values ​​of the time spent processing a single gear data point on the test set. Repeat the operation multiple times on the same test set and take the average value as the final result.

6. An automatic gear shifting control system for mining trucks, used to implement the automatic gear shifting control method for mining trucks as described in any one of claims 1 to 5, characterized in that, include: The first processing module acquires offline status data of mining trucks and preprocesses it, sequentially generates the original MTD set, filters out data that is irrelevant to gear shifting and invalid data, generates the effective MTD set, and randomly divides it into training set and test set. The second processing module uses the training set as input to a pre-built ResNet-Bi-LSTM-Attention network with multiple parameters and time spans to train the ResNet-Bi-LSTM-Attention network. The output module takes the test set as input to the trained ResNet-Bi-LSTM-Attention network to verify the robustness of the ResNet-Bi-LSTM-Attention network, obtains the final ResNet-Bi-LSTM-Attention network model, and inputs the acquired offline state data of the mining truck into the model to output the final gear position feature and complete the gear shifting operation.

7. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described in claims 1 to 5.

8. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described in claims 1 to 5.

Images