CN116227540A

CN116227540A - Mining truck automatic gear shifting control method, system, storage medium and computing equipment

Info

Publication number: CN116227540A
Application number: CN202310107752.7A
Authority: CN
Inventors: 王立勇; 苏清华; 许筱毓
Original assignee: Beijing Information Science and Technology University
Current assignee: Beijing Information Science and Technology University
Priority date: 2023-02-14
Filing date: 2023-02-14
Publication date: 2023-06-06

Abstract

The invention relates to a mining truck automatic gear shifting control method, a system, a storage medium and a computing device, which comprise the following steps: acquiring offline state data of a mining truck, preprocessing, sequentially generating an original MTD set, filtering data irrelevant to gear shifting and invalid, generating an effective MTD set, and randomly dividing the effective MTD set into a training set and a testing set; training the ResNet-Bi-LSTM-Attention network by taking the training set as the input of the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans which are established in advance; inputting the test set into the ResNet-Bi-LSTM-Attention network after training, verifying the robustness of the ResNet-Bi-LSTM-Attention network, obtaining a final ResNet-Bi-LSTM-Attention network model, inputting the obtained mining truck offline state data into the model, outputting final gear characteristics, and finishing gear shifting operation. The method solves the problems that the integrity of the context information of the time sequence data is easily ignored and the front and rear characteristics of the data cannot be well integrated in the prior art, and can be applied to the field of gear shifting control of heavy vehicles.

Description

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

Technical Field

The invention relates to the technical field of gear shifting control of heavy vehicles, in particular to an automatic gear shifting control method, system, storage medium and computing equipment of a mining truck based on a ResNet-Bi-LSTM-Attention network and multiple time spans.

Background

Driving mining trucks under rough mining area road conditions is a complex challenge. Therefore, the vehicle is designed in consideration of not only drivability but also economy, power, driving comfort, and the like. Currently, many mining trucks still employ manual shifting transmissions, which increase the workload and the rate of mishandling of the driver after long driving. In addition, irregular shifting and wrong shifting by the driver also increase the probability of transmission damage. However, these problems can be effectively avoided using an automatic shift transmission. The automatic shift transmission is controlled by an appropriate shift strategy, which reduces operational difficulties and improves shift efficiency and dynamic performance. Under complex road conditions, optimizing the gear shifting strategy is important. The shift strategy automatically selects the appropriate gear taking into account road conditions, fuel economy and power performance. Thus, driving comfort is greatly improved, and driving difficulty is remarkably reduced.

Over the past few years, many students have studied automatic shift strategies for automatic transmissions due to the lack of disclosure by the major transmission manufacturers. Many automatic shift strategies on Automated Mechanical Transmission (AMT) automobiles are based on two or three parameters, such as pedal, vehicle speed, and engine speed. Other potential factors, such as oil temperature, speed change ratio, and output speed, are not used in its shift strategy. Therefore, it cannot accurately judge the dynamic change in the running process of the vehicle, and can influence the vehicle performance under complex and changeable road conditions. The comparison of gear prediction accuracy based on the dual-parameter and multi-parameter gear shifting strategies shows that the strategy with more vehicle state data can greatly improve the gear prediction accuracy.

However, in the existing gear shifting control strategy, only the correlation between continuous time information is considered, but the correlation between discontinuous time data is ignored, and the front and rear characteristics of the data cannot be well integrated, so that the requirements of high gear prediction accuracy and real-time prediction time cannot be met.

Disclosure of Invention

Aiming at the problems, the invention aims to provide a mining truck automatic gear shifting control method, a system, a storage medium and a computing device, which solve the problems that the integrity of time sequence data context information is easily ignored and the front and back characteristics of data cannot be well integrated in the prior art.

In order to achieve the above object, according to a first aspect, the present invention adopts the following technical scheme: an automatic shift control method for a mining truck, comprising: acquiring offline state data of a mining truck, preprocessing, sequentially generating an original MTD set, filtering data irrelevant to gear shifting and invalid, generating an effective MTD set, and randomly dividing the effective MTD set into a training set and a testing set; training the ResNet-Bi-LSTM-Attention network by taking the training set as the input of the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans which are established in advance; inputting the test set into the ResNet-Bi-LSTM-Attention network after training, verifying the robustness of the ResNet-Bi-LSTM-Attention network, obtaining a final ResNet-Bi-LSTM-Attention network model, inputting the obtained mining truck offline state data into the model, outputting final gear characteristics, and finishing gear shifting operation.

Further, acquiring offline state data of the mining truck and preprocessing, including:

decoding state data of a mining truck acquired in real time to generate an original MTD set with timing, wherein the last column of gear values CG at the current time in the set is used as a true value of model training, and the other columns are used as potential parameters of gear prediction;

Calculating an H index of an original MTD set by adopting an R/S analysis method, and screening out gear shifting parameters with positive correlation through the H index;

and carrying out outlier processing and data normalization processing on the screened gear shifting parameters with positive correlation to generate an effective MTD set consisting of effective time data.

Further, during R/S analysis, all parameters are used as inputs, except for the parameter CG, which is considered as the actual value of the model output; the calculation formula of the H index is as follows:

H＝log _n (R/S) _n -log _n (C)

wherein:

wherein A is the number of consecutive subintervals, M is the length of the total time series, n is the length of each subinterval, (R/S) _a Representing the rescaling range of the subinterval sequence a, C is a constant.

Further, screening shift parameters with positive correlation through the H index comprises the following steps:

if the H index is greater than a preset value, the parameter has positive correlation in gear shifting;

if the H index is equal to a preset value, the parameter has no influence on gear shifting;

if the H index is less than the preset value, the parameter has a negative correlation in shifting.

Further, a pre-established multi-parameter and time-span ResNet-Bi-LSTM-Attention network, comprising:

and establishing an improved Residual Network and Bi-LSTM Network model, adding an attention mechanism module after feature fusion, then accessing a full-connection dimension reduction module, and finally outputting model features.

Further, training the ResNet-Bi-LSTM-Attention network, comprising:

the input layer is a set of MTD data based on multiple time spans, and the input features are composed of a plurality of vehicle state data;

before Bi-LSTM and ResNet are input, dimension reduction is carried out through a layer of full-connection layer, and after dimension reduction treatment, the characteristics are respectively and simultaneously input into Bi-LSTM and ResNet networks for calculation;

in the Bi-LSTM structure, a hidden layer is firstly input, each super parameter is set, and the batch size is manually set before each training; inputting a bidirectional LSTM layer at the time of t+n time step to encode the time sequence feature, and obtaining the hidden layer state value h ₁ ，h ₂ ,....,h _n Respectively inputting the training steps into a hidden layer, multiplying the obtained result by a training step length, and calculating the value of the current state of the hidden layer; adding random inactivation to relieve the over-fitting problem in model training, and outputting the output characteristics of the hidden state of Bi-LSTM at the final t+n moment;

in the training process of the ResNet network, a layer of convolution layer is firstly input for reducing the dimension again so as to reduce the complexity of the whole structure; accessing a convolution kernel of the first layer 1*1, defining a normalized function for an input channel, and setting an activation function as Relu and other parameters as defaults; sequentially accessing the convolution kernel of the second layer 3*3 and the convolution kernel of the third layer 1*1; adding a Shortcut structure so that the input layer is directly connected to the output layer through weighting; inputting the calculation result into a global average pooling layer for pooling operation, and finally obtaining the calculation value of the ResNet layer;

Splicing the results of the hidden layers after the dimension reduction of the Bi-LSTM and the ResNet layers by using a gate control residual error connection mode, and reducing the dimension through a layer of full-connection network;

combining the initial input characteristics and the total output value of the hidden states through a gating mechanism to serve as the input of a next-layer network, and accessing two hidden layers to fully extract the characteristics;

if the model calculation loss of the current iteration training is smaller than that of the last epoch training, ending the training and saving the current model, otherwise, stopping the training in advance and saving the current best model;

and the access output layer and the softMax layer output the final gear characteristics.

Further, verifying robustness of the ResNet-Bi-LSTM-Attention network, comprising:

loading a weight file with the best result of each network structure during training, inputting a test set with the same parameters, and calculating the time for predicting a single gear;

and recording the minimum value, the average value and the maximum value of the cost for processing the single gear data by the test set, repeating the operation for a plurality of times on the same test set, and taking the average value as a final result.

In a second aspect, the present invention adopts the following technical scheme: an automatic shift control system for a mining truck, comprising: the first processing module acquires offline state data of the mining truck, performs preprocessing, sequentially generates an original MTD set, filters out data irrelevant to gear shifting and invalid, generates an effective MTD set, and randomly divides the effective MTD set into a training set and a testing set; the second processing module takes the training set as the input of the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans which are established in advance, and trains the ResNet-Bi-LSTM-Attention network; and the output module inputs the test set into the ResNet-Bi-LSTM-Attention network after training, verifies the robustness of the ResNet-Bi-LSTM-Attention network, obtains a final ResNet-Bi-LSTM-Attention network model, inputs the acquired mining truck offline state data into the model, outputs final gear characteristics and completes gear shifting operation.

In a third aspect, the present invention adopts the following technical scheme: a computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform any of the methods described above.

In a fourth aspect, the present invention adopts the following technical scheme: a computing apparatus, comprising: one or more processors, 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 comprising instructions for performing any of the methods described above.

Due to the adoption of the technical scheme, the invention has the following advantages:

1. compared with RNN, LSTM, bi-LSTM, resNet-Bi-LSTM and Bi-LSTM-Attention networks, the mining truck automatic shifting method based on the multi-parameter and time span mining truck automatic shifting method not only can consider the correlation between continuous and discontinuous time information, but also can enable training to be focused on feature vectors with larger weights, so that the characteristics of robustness and less noise are extracted.

2. When the parameters at adjacent moments generate gear abnormality along with external interference, the accuracy can generate errors when the single time span calculation is adopted to finish the up-down gear prediction. The multi-time span estimation of the invention can capture numerical information in a longer time span and can better prevent the phenomenon that the expected shift position is deviated due to the abnormality of a single time value. Furthermore, the invention also performs a plurality of groups of training with different batch sizes so as to find the best performance result in the process of training the network.

3. Compared with performance indexes of network models of RNN, LSTM, bi-LSTM, resNet-Bi-LSTM and Bi-LSTM-Attention, experimental results show that the ResNet-Bi-LSTM-Attention network provided by the invention can ensure that the average time of predicting a single gear can meet the actual driving requirement of a driver, and meanwhile, the loss function is quickly converged and minimized, and the accuracy is highest. Therefore, the performance of the invention is obviously improved compared with other models.

Drawings

FIG. 1 is a flow chart of an automatic shift control method for a mining truck in an embodiment of the present invention;

FIG. 2 is a flow chart of data preprocessing in an embodiment of the invention;

FIG. 3 is a graph of outlier handling results in an embodiment of the present invention;

FIG. 4 is a schematic diagram of a Bi-LSTM model structure in accordance with an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. It will be apparent that the described embodiments are some, but not all, embodiments of the invention. All other embodiments, which are obtained by a person skilled in the art based on the described embodiments of the invention, fall within the scope of protection of the invention.

It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular is also intended to include the plural unless the context clearly indicates otherwise, and furthermore, it is to be understood that the terms "comprises" and/or "comprising" when used in this specification are taken to specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof.

Some shift strategies are based on complex kinetic mathematical models, which allow gear to be adjusted to the proper position, but do not guarantee real-time processing. In recent years, with the rapid development of machine learning technology, many researchers have also applied it to the field of gear shifting strategies, and have achieved great achievements. An automatic gear shifting system based on a neural network in the prior literature inputs three dynamic parameters: vehicle speed, vehicle acceleration, and pedal position. To handle complex driving conditions, a correction module is added in the network, but it still faces the challenge of handling data with temporal characteristics. Nonlinear feature processing of time series has some drawbacks that can be addressed by deep learning time series modeling. Thus, the deep learning technology is widely applied to processing time series data with nonlinear characteristics, and can overcome the defects of the traditional neural network.

In order to better process the sequence information, a more efficient method of initializing weights is applied to the field of deep learning. In order to solve the problems that the traditional Recurrent Neural Network (RNN) is insufficient in processing time sequence historical data and gradient disappearance or explosion is easy to occur, a long-short-term memory network (LSTM) is adopted, and a Bi-directional long-short-term memory network (Bi-LSTM) is utilized to develop an automatic gear shifting strategy based on a mine car and obtain higher prediction accuracy. However, the Bi-LSTM neural network easily ignores the integrity of the context information of the time series data, so that many people adopt the Convolutional Neural Network (CNN) to make up for the problem that the Bi-LSTM network is insufficient in extraction characteristics, the Bi-LSTM network is fused with the CNN, BLSTM-2DPooling and BLSTM-2DCNN models are provided, not only the dimension on text feature vectors but also the dimension of time steps are considered, and experimental results show that the network can capture more abundant semantic features. But CNN ignores the correlation between non-continuous time data considering only the correlation between continuous time information. A residual network (ResNet) is adopted to train a deeper CNN model, an identity mapping structure is added on the basis of CNN in the ResNe network, the operation of the layer or layers is skipped, and meanwhile, in the backward propagation process, the gradient of the network of the lower layer can be directly transferred to the upper layer, so that the problem that the CNN is degraded or exploded due to the fact that the depth is simply overlapped is solved, and additional operation amount cannot be generated on the whole network. Therefore, the ResNe is adopted to replace the CNN model to be overlapped with the Bi-LSTM model, the network structure of the combination of the Bi-LSTM and the ResNet disclosed in the prior document is adopted to classify the ECG, and finally, a good result is obtained, and the accuracy rate of 96.2% on the MIT data set can be achieved. Attention mechanisms (Attention) are widely applied to various fields, and the Attention mechanisms give different weights to different features in a model by calculating Attention probabilities of the different features, so that training is focused on feature vectors with larger weights, and extracted feature information is better utilized.

The invention provides an automatic gear shifting control method, system, storage medium and computing equipment for a mining truck, which are used for solving the problem that the integrity of time sequence data context information is easily ignored and the front and rear characteristics of data cannot be well integrated in the prior art. The mining truck based on the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans automatically shifts gears, adopts the multiple parameters as input characteristics, assists a driver to switch gears at future moments, and has applicability in other traffic environments. To evaluate the robustness of the ResNet-Bi-LSTM-Attention network model of the present invention, it is compared with Recurrent Neural Networks (RNNs), long-short-term memory networks (LSTMs), two-way long-short-term memory networks (Bi-LSTMs), residual networks (Residual Neural Network, resNet) ResNet-Bi-LSTMs, and Attention-mechanism networks (Attention-mechanism) Attention-Bi-LSTM networks; from the experimental results, the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans adopted by the invention is superior to other networks, and meets the requirement of real-time prediction time while ensuring high accuracy of gear prediction.

In one embodiment of the invention, a mining truck automatic shift control method is provided. In this embodiment, the method is implemented based on multi-parameter and time span ResNet-Bi-LSTM-Attention, specifically, as shown in FIG. 1, the method includes the following steps:

1) Acquiring offline data of a mining truck, preprocessing the offline data, sequentially generating an original MTD (mining truck data) set, filtering data irrelevant to gear shifting and invalid, generating an effective MTD set, and randomly dividing the effective MTD set into a training set and a testing set;

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

The MTD is preprocessed to build up a sequential data list, filtering data that is shift independent and invalid. The purpose of the data preprocessing is to find parameters positively correlated with the shift, process the invalid data, and reduce the impact of the invalid data on model training.

2) Training the ResNet-Bi-LSTM-Attention network by taking the training set as the input of the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans which are established in advance;

3) Inputting the test set into the ResNet-Bi-LSTM-Attention network after training, verifying the robustness of the ResNet-Bi-LSTM-Attention network, obtaining a final ResNet-Bi-LSTM-Attention network model, inputting the obtained mining truck offline state data into the model, outputting final gear characteristics, and finishing gear shifting operation.

In this embodiment, a shift control system for performing an automatic shift is further included between executing step 1). The shift control system includes: a transmission, a control handle, and a Transmission Control Unit (TCU). The transmission performs a shift upon receiving the TCU control signal and sends transmission status information to other monitoring devices via the CAN bus. The mining truck automatic gear shifting control method is arranged in the TCU.

The TCU controlling the gear shift is mounted on the mining truck, monitors mining truck status data via various sensors, and is shared with other associated Electronic Control Units (ECU) via CAN bus, including accelerator pedal position, current gear, engine speed, oil temperature, operating pressure, transmission output speed, and system errors.

The TCU controls a shift operation when a shift command signal from the control handle is received. Before shifting, appropriate safety checks are made, and then the TCU performs upshifts, downshifts, and neutral at appropriate times to protect the gearbox, reduce fuel costs, and maximize vehicle performance.

In this example, the embedded device on the test vehicle was NVIDIA Jetson AGX Xavier and the detailed parameter information is shown in table 1.

TABLE 1 parameter information

In the embodiment, in order to acquire vehicle state data, four ZLG CANDTU-200UR CAN buses are connected to four mining trucks; during manual driving, vehicle data processed and monitored by different ECUs on the mine car will be sent over the CAN bus. The CAN bus card then automatically starts data recording to further develop the machine learning model. The raw CAN bus data follows the SAEJ1939 protocol, which is widely used in the truck industry. 14 types of vehicle parameters are obtained from a Parameter Group Number (PGN) by a data decoding algorithm, and each parsed vehicle parameter data will also have a time stamp for marking the time of receipt.

The size of the experimental mine car of the manual transmission was 7.65 meters long, 3.5 meters wide, 3.8 meters high and a weight of 20 tons. There are few road emergencies in the mine, the traffic conditions are simple, and traffic regulations are less restricted because the mine is located in a closed area where traffic and population density are low.

In the step 1), offline state data of the mining truck is obtained and preprocessed, as shown in fig. 2, and the method comprises the following steps:

1.1 Decoding state data of the mining truck obtained in real time to generate an original MTD set with timing, wherein the last column in the set represents a gear value CG of the current time as a true value of model training, the other columns serve as potential parameters of gear prediction, and the data are updated row by row when CAN original data are received from the mining truck;

in a possible embodiment, the 1020301 pieces of CAN raw data acquired in total are decoded by the SAEJ1939 protocol and an MTD set with timing is generated; the MTD set consists of 14 columns of data, corresponding to 14 parameter values available from the CAN bus.

1.2 In order to screen out the parameter of the positive correlation, an R/S analysis method is adopted to calculate the H index (Hersteter index) of the original MTD set, and the gear shift parameter with the positive correlation is screened out through the H index;

During R/S analysis, all parameters are used as inputs, except for the parameter CG, which is considered as the actual value of the model output; the calculation formula of the H index is as follows:

/>

R _a ＝max(X _k,a )-min(X _k,a ) (4)

wherein A is a time series R with a total length of M _t Number of equal-divided subintervals, wherein each subinterval has a length of n, (R/S) _a Representing the rescaling range of the subinterval sequence a, C being a constant; i _a Representing subinterval sequence a (a.epsilon.1, 2,3, …, A]),R _k,a Representation I _a Element (k.epsilon.1, 2,3, …, n)]) And represents the average value of the sequence a of subintervals; r is R _i,a Represents k elements, X in the subinterval sequence a _k,a Representing the cumulative deviation; r is R _a Is X _k,a Is within a range of max (X _k,a ) And min (X) _k,a ) Is X _k,a Maximum and minimum cumulative deviation of (2); s is S _a Is the standard deviation, (R/S) _a Representing the rescaling range of subinterval sequence a, (R/S) _n Is the average of the rescaled ranges for all subintervals; h is the hurst index.

1.3 Abnormal value processing and data normalization processing are carried out on the screened gear shifting parameters with positive correlation, and an effective MTD set composed of effective time data is generated.

In the step 1.2), the H index value reflects the importance of certain parameters in the gear shifting process, and the gear shifting parameters with positive correlation are screened out through the H index, which specifically comprises the following steps:

The experimental results show that the variables of positive correlation are shown in table 2.

Table 2H index calculation results of shift related variables

In the step 1.3), the outlier processing and the data normalization processing specifically include:

data from the sensors may be affected and become anomalous in the collection of the MTD due to electromagnetic and other environmental factors. The block diagram for outlier detection is used to find outliers and to guarantee data quality. In fig. 3, the size of the block is the inter-quartile range (IQR), and the upper and lower ranges of the block are the upper quartile (Q3) and the lower quartile. The upper and lower edges are UE and LE, which are calculated by equations (9) and (10). Parameter values exceeding the UEs or LEs in the respective categories are regarded as outliers, marked as "o" in the ES in fig. 3. Since the parameter data from the CAN bus is continuous and the interval is very short (1 to 10Hz per second), the outliers will be replaced with the average data of the previous and next frames to ensure data consistency.

UE＝Q3+1.5IQR (9)

LE＝Q1-1.5IQR (10)

According to the generated result, the overall electrodeless end outlier is known.

Since the MTD consists of valid time data, it is not normalized. To ensure that all data in the MTD map to the same scale and distribution range for subsequent machine learning model training, the raw data x in the MTD dataset is subjected to a maximum-minimum normalization process:

where x' represents the processed data, min (x) represents the minimum value of x, and max (x) represents the maximum value of x.

In the step 2), the pre-established multi-parameter and time span ResNet-Bi-LSTM-Attention network comprises: the method comprises the steps of establishing a Residual Network fusion Bi-LSTM Network model, adding an attention mechanism module after feature fusion, then accessing full-connection dimensionality reduction, and finally outputting model features, improving the existing Residual Network, reducing the number of layers of Residual Network superposition, improving a Loss function to be Focal Loss, and performing smooth optimization processing on labels, thereby improving the problem of class unbalance.

2.1.1 Bi-LSTM network structure. The Bi-LSTM structure is that a reverse output and a general output are added on the basis of the LSTM network, so that bidirectional history information can be obtained, as shown in fig. 4. Comprising an input gate i for inputting valid information into the structure _t Output gate o for outputting the unit information after screening to the next time _t Forgetting door f for screening effective information and filtering useless information for prediction _t ，C _t Is the memory state of the cell, W _f 、W _i 、W _o And W is _c The weights matrix of three gates and cell candidate states, x, respectively _t For the total data set with time sequence, h _t Is the state value of the hidden layer at time t, b _f 、b _i 、b _o And b _c Is the corresponding offset, where σ is the activation function of the three gates, T is the threshold of the total time series,

an activation function for which tanh is a candidate state, c _t For the state value of the current cell, +.>

For the forward hidden layer neuron value,

for the backward hidden layer neuron value, y _t For the final eigenvalue, the specific formula is as follows:

f _t ＝σ(W _f ·x _t +W _f ·h _t-1 +b _f ) (12)

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

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

h _t ＝o _t ·tanh(c _t ) (17)

2.1.2 ResNet network architecture. In the residual network, an identity mapping method is adopted to construct a deep model, wherein x is as follows ^(l) History information, x, learned for current l-layer residual block ^(l+1) For the output of layer I, W ^(l) Is the weight of the first layer, F (x ^(l) ,W ^(l) ) To fit the residual map, H (x ^(l) ) Is the network map input to the summation, whose calculation formula is as follows:

H(x ^(l) )＝F(x ^(l) ,W ^(l) )+x ^(l) (21)

when F (x) ^(l) ,W ^(l) ) When 0, an optimal identity map is formed:

H(x ^(l) )＝x ^(l) (22)

the history information obtained by the residual network with depth L is generated by overlapping each layer with depth shallower than the residual network, wherein x is ^(l) Representing the input of the first residual block, W ⁽ⁱ⁾ For the weight of each layer, x ^(l+1) For the output of this residual block, which is also the input of the 1+1th residual block, the output x of the deeper L residual block can be calculated by accumulating each residual block ^(L) ：

It can be seen from the equation that any unit L and L has the characteristics of a residual network of residual characteristics, and the characteristics of the unit L are accumulated by the residual characteristics of each layer, wherein the layer l+1 can be combined with more characteristic information than the layer L.

In the back propagation process, the loss function is for the first residual block x according to the chain derivative rule ^(l) The gradient calculation formula of (2) is:

in the method, in the process of the invention,

indicating that the gradient of the L layer can be transferred directly to any network layer L shallower than it,

indicating that the value is not constant equal to-1 during back propagation, thus avoiding the problem of gradient extinction. The performance degradation problem that arises from stacking multiple layers of network layers can thus be solved by minimizing the residual mapping F (x).

2.1.3 An Attention network structure). h is a _n Representing the hidden layer output value of the residual network and the Bi-LSTM network at the n time after the dimension reduction,

for feature representation vectors that are one level higher than the input variables, β represents the nth hidden layer output h _n At->

The larger β is, the higher the weight of the input representing the moment in the whole is, and the following formula is calculated:

wherein b is a bias term, V, W, U is a weight matrix, and tanh is a nonlinear activation function;

...,/>

outputting the proportion of the matching score of each feature vector to the total score for the hidden layer, +.>

The calculation formula is as follows:

calculating a final feature F based on the attention distribution _att ：

In the process of obtaining the feature vector F _att Then, the probability distribution y of the classification labels is calculated through a softmax function of the output layer:

wherein V represents a weight matrix of the model output layer, F' _att Representing weighted F _att ，F′ _att(i) Representing vector F' _att The vector length is equal to the number of class labels, and T is the number of class labels.

Finally, to make the model loss function convergence smoother, cross entropy is used as the loss function. The probability distribution Y and the true category distribution Y are subjected to cross entropy loss E (Y, Y):

/>

in the step 2), training is carried out on the ResNet-Bi-LSTM-Attention network, and the method comprises the following steps:

2.2.1 The input layer is a set of MTD data based on multiple time spans, and the input features are made up of a plurality of vehicle state data;

for example, the input layer is a set of slave x _t-k To x _t Each feature x consists of 9 vehicle state data, including ACPP, OSS, AEPT, RPT, ES, AGR, TSOT, TCR, DRT, over an appropriate time span.

2.2.2 Because the feature is larger in dimension, dimension reduction is carried out through a layer of full-connection layer before Bi-LSTM and ResNet are input, so that the whole parameter size suitable for a model can be obtained; after the dimension reduction treatment, respectively inputting the characteristics into Bi-LSTM and ResNet networks at the same time for calculation;

2.2.3 In Bi-LSTM structure, a hidden layer is first input and arrangedSetting each super parameter, and manually setting the batch size before each training; inputting a bidirectional LSTM layer at the time of t+n to code the time sequence features, so that the problem that the LSTM model is more biased to acquire only the latest time features can be alleviated; outputting the obtained hidden layer output characteristics h ₁ ，h ₂ ,....,h _n Respectively inputting the training steps into a hidden layer, multiplying the obtained result by a training step length, and calculating the value of the current state of the hidden layer; adding random inactivation to relieve the over-fitting problem in model training, and outputting the output characteristics of the hidden state of Bi-LSTM at the final t+n moment;

for example, in the Bi-LSTM structure, a hidden layer of size 32 is first input, a random deactivation rate of 0.22 is set, a learning rate of 0.005, the number of iterations is 500, and the batch size is manually set before each training and the most appropriate value is finally sought.

2.2.4 During the training process of the ResNet network, firstly, a layer of convolution layer is input for reducing the dimension again so as to reduce the complexity of the whole structure; accessing a convolution kernel of the first layer 1*1, defining a normalized function for an input channel, and setting an activation function as Relu and other parameters as defaults; sequentially accessing the convolution kernel of the second layer 3*3 and the convolution kernel of the third layer 1*1; the Shortcut structure is added, so that an input layer is directly connected to an output layer through weighting, and the gradient divergence problem in a depth network is relieved; in order to avoid overfitting and improve training speed, inputting a calculation result into a global average pooling layer (Global Average Pooling) for pooling operation, and finally obtaining a calculation value of a ResNet layer;

in this embodiment, the parameter settings of the second layer and the third layer are the same as those of the first layer.

2.2.5 The method comprises the steps of) splicing the results of the hidden layers after the dimension reduction of the Bi-LSTM layer and the ResNet layer by using a gate residual connection mode, reducing the dimension by a layer of fully connected network so as to maintain the vector dimension of the original hidden state, and then weighting important features by using an attention mechanism;

2.2.6 Combining the initial input characteristics and the total output value of the hidden states through a gating mechanism to serve as the input of a next-layer network, and accessing two hidden layers to fully extract the characteristics, so that the characteristics of data can be better divided;

2.2.7 If the model calculation loss of the current iteration training is smaller than that of the last epoch training, ending the training and saving the current model, otherwise, stopping the training in advance and saving the current best model;

2.2.8 The output layer is accessed and the SoftMax layer outputs the final gear characteristics;

2.2.9 Repeating the above process, and outputting and storing the best result.

In the step 3), in order to verify the robustness of the ResNet-Bi-LSTM-Attention network, experiments were performed on a server configured in the same environment as the vehicle embedded system (NVIDIA Jetson AGX Xavier). Of which there are 900000 total effective MTDs for training the machine learning model, and another 100000 for validation. Meanwhile, in order to ensure the rationality of the results, each model was trained 10 times under different parameters, and the final results were the average of all results.

In this embodiment, the robustness of the ResNet-Bi-LSTM-Attention network is verified, comprising the steps of:

3.1 Loading a weight file with the best result when each network structure is trained, inputting a test set with the same parameters, and calculating the time for predicting a single gear;

specifically, the initial time t from the loading of the model ₀ The single time t is finally predicted after the prediction is finished ₁ The time difference between the two is the predicted time of a single gear, and other models are set as comparison.

3.2 Recording the Minimum (MIN), average (AVG) and Maximum (MAX) values of the cost of the test set for processing the single gear data, and taking the average value as a final result after repeating the operation for a plurality of times for the same test set in order to obtain more accurate prediction results.

And comparing experiments of the models prove that the robustness of the model is highest. Meanwhile, the automatic gear shifting control method is input to the control handle, and finally the gear shifting operation is completed.

In one embodiment of the present invention, there is provided a mining truck automatic shift control system comprising:

the first processing module acquires offline state data of the mining truck, performs preprocessing, sequentially generates an original MTD set, filters out data irrelevant to gear shifting and invalid, generates an effective MTD set, and randomly divides the effective MTD set into a training set and a testing set;

the second processing module takes the training set as the input of the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans which are established in advance, and trains the ResNet-Bi-LSTM-Attention network;

and the output module inputs the test set into the ResNet-Bi-LSTM-Attention network after training, verifies the robustness of the ResNet-Bi-LSTM-Attention network, obtains a final ResNet-Bi-LSTM-Attention network model, inputs the state data of the mining truck stored locally into the model, outputs the final gear characteristic and completes the gear shifting operation.

In the first processing module, the method for acquiring and preprocessing the offline state data of the mining truck includes:

Wherein during the R/S analysis, all parameters are used as inputs except for the parameter CG, which is considered as the actual value of the model output; the calculation formula of the H index is as follows:

H＝log _n (R/S) _n -log _n (C)

wherein:

wherein A is the number of consecutive subintervals divided into a time series of length M, n is the length of each subinterval, (R/S) _a Representing the rescaling range of the subinterval sequence a, C is a constant.

In this embodiment, a shift parameter with positive correlation is screened out through an H index, which specifically includes:

The above embodiment in the second processing module, the pre-established multi-parameter and time span ResNet-Bi-LSTM-Attention network includes:

firstly, a Residual Network fusion Bi-LSTM Network model is established to acquire local features, the input data can be subjected to long-sequence feature extraction, aiming at the condition that the number of data sets is small in some types, the existing Residual Network is improved on one hand, the number of layers of overlapping Residual networks is reduced, meanwhile, a Loss function of Loss is improved to be Focal Loss, a label is subjected to smooth optimization processing, the problem of unbalanced types is solved, then an attention mechanism module is increased, different weights are given to the features to enable the Network to be more focused on important information, then all-connection-layer dimension reduction is carried out, and finally the model features are fused and then output.

The above embodiment trains the ResNet-Bi-LSTM-Attention network in the second processing module, including:

in the Bi-LSTM structure, a hidden layer is firstly input, each super parameter is set, and the batch size is manually set before each training; inputting a bidirectional LSTM layer at the time of t+n time step to encode the time sequence characteristics, and obtaining the hidden layer characteristics h ₁ ，h ₂ ,....,h _n Respectively inputting the training steps into a hidden layer, multiplying the obtained result by a training step length, and calculating the value of the current state of the hidden layer; adding random inactivation to relieve the over-fitting problem in model training, and outputting the output characteristics of the hidden state of Bi-LSTM at the final t+n moment;

The above embodiment verifies the robustness of the ResNet-Bi-LSTM-Attention network in the output module, comprising:

The system provided in this embodiment is used to execute the above method embodiments, and specific flow and details refer to the above embodiments, which are not described herein.

A schematic structural diagram of a computing device provided in an embodiment of the present invention, where the computing device may be a terminal, may include: a processor (processor), a communication interface (Communications Interface), a memory (memory), a display screen, and an input device. The processor, the communication interface and the memory complete communication with each other through a communication bus. The processor is configured to provide computing and control capabilities. The memory comprises a nonvolatile storage medium and an internal memory, wherein the nonvolatile storage medium stores an operating system and a computer program, and the computer program is executed by a processor to realize a mining truck automatic gear shifting control method; the internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface is used for carrying out wired or wireless communication with an external terminal, and the wireless mode can be realized through WIFI, a manager network, NFC (near field communication) or other technologies. The display screen can be a liquid crystal display screen or an electronic ink display screen, the input device can be a touch layer covered on the display screen, can also be a key, a track ball or a touch pad arranged on the shell of the computing equipment, and can also be an external keyboard, a touch pad or a mouse and the like. The processor may call logic instructions in memory to perform the following method: acquiring offline state data of a mining truck, preprocessing, sequentially generating an original MTD set, filtering data irrelevant to gear shifting and invalid, generating an effective MTD set, and randomly dividing the effective MTD set into a training set and a testing set; training the ResNet-Bi-LSTM-Attention network by taking the training set as the input of the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans which are established in advance; inputting the test set into the ResNet-Bi-LSTM-Attention network after training, verifying the robustness of the ResNet-Bi-LSTM-Attention network, obtaining a final ResNet-Bi-LSTM-Attention network model, inputting the state data of the mining truck stored locally into the model, outputting the final gear characteristics, and finishing the gear shifting operation.

Further, the logic instructions in the memory described above may be implemented in the form of software functional units and stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

In one embodiment of the present invention, there is provided a computer program product comprising a computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions which, when executed by a computer, are capable of performing the methods provided by the method embodiments described above, for example comprising: acquiring offline state data of a mining truck, preprocessing, sequentially generating an original MTD set, filtering data irrelevant to gear shifting and invalid, generating an effective MTD set, and randomly dividing the effective MTD set into a training set and a testing set; training the ResNet-Bi-LSTM-Attention network by taking the training set as the input of the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans which are established in advance; inputting the test set into the ResNet-Bi-LSTM-Attention network after training, verifying the robustness of the ResNet-Bi-LSTM-Attention network, obtaining a final ResNet-Bi-LSTM-Attention network model, inputting the state data of the mining truck stored locally into the model, outputting the final gear characteristics, and finishing the gear shifting operation.

In one embodiment of the present invention, there is provided a non-transitory computer-readable storage medium storing server instructions that cause a computer to perform the methods provided by the above embodiments, for example, including: acquiring offline state data of a mining truck, preprocessing, sequentially generating an original MTD set, filtering data irrelevant to gear shifting and invalid, generating an effective MTD set, and randomly dividing the effective MTD set into a training set and a testing set; training the ResNet-Bi-LSTM-Attention network by taking the training set as the input of the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans which are established in advance; inputting the test set into the ResNet-Bi-LSTM-Attention network after training, verifying the robustness of the ResNet-Bi-LSTM-Attention network, obtaining a final ResNet-Bi-LSTM-Attention network model, inputting the state data of the mining truck stored locally into the model, outputting the final gear characteristics, and finishing the gear shifting operation.

The foregoing embodiment provides a computer readable storage medium, which has similar principles and technical effects to those of the foregoing method embodiment, and will not be described herein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may 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, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic gear shift control method for a mining truck is characterized by comprising the following steps:

acquiring offline state data of a mining truck, preprocessing, sequentially generating an original MTD set, filtering data irrelevant to gear shifting and invalid, generating an effective MTD set, and randomly dividing the effective MTD set into a training set and a testing set;

training the ResNet-Bi-LSTM-Attention network by taking the training set as the input of the ResNet-Bi-LSTM-Attention network with multiple parameters and time spans which are established in advance;

inputting the test set into the ResNet-Bi-LSTM-Attention network after training, verifying the robustness of the ResNet-Bi-LSTM-Attention network, obtaining a final ResNet-Bi-LSTM-Attention network model, inputting the obtained mining truck offline state data into the model, outputting final gear characteristics, and finishing gear shifting operation.

2. The mining truck automatic shift control method of claim 1, wherein acquiring and preprocessing offline state data of the mining truck comprises:

3. The mining truck automatic shift control method of claim 2, wherein during the R/S analysis, all parameters are used as inputs except for a parameter CG, which is considered as an actual value of the model output; the calculation formula of the H index is as follows:

H＝log _n (R/S) _n -log _n (C)

wherein:

4. The mining truck automatic shift control method according to claim 2, wherein selecting shift parameters having positive correlation by H index includes:

5. The mining truck automatic shift control method of claim 1, wherein the pre-established multi-parameter and time span res net-Bi-LSTM-Attention network comprises:

6. The mining truck automatic shift control method of claim 1, wherein training the res net-Bi-LSTM-Attention network comprises:

7. The mining truck automatic shift control method of claim 1, wherein verifying robustness of the res net-Bi-LSTM-Attention network comprises:

8. An automatic shift control system for a mining truck, comprising:

and the output module inputs the test set into the ResNet-Bi-LSTM-Attention network after training, verifies the robustness of the ResNet-Bi-LSTM-Attention network, obtains a final ResNet-Bi-LSTM-Attention network model, inputs the acquired mining truck offline state data into the model, outputs final gear characteristics and completes gear shifting operation.

9. A computer readable storage medium storing one or more programs, wherein the one or more programs comprise instructions, which when executed by a computing device, cause the computing device to perform any of the methods of claims 1-7.

10. A computing device, comprising: one or more processors, 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 comprising instructions for performing any of the methods of claims 1-7.