Optimization method of loader shovel loading process based on machine learning

Abstract

The present invention aims to provide a method for optimizing the loader shoveling operation process based on machine learning, comprising the following steps: A. The loader performs multiple field shoveling operations and collects 15 bucket position parameters for each operation; B. Collects the three optimization target parameters of shoveling weight, operation energy consumption, and operation time for each operation; C. Collects the 15 bucket position parameters and optimization target parameters for each operation to form multiple bucket position parameter data groups and optimization target parameter data groups, removes abnormal data in the data groups and performs normalization processing; D. Constructs a BP neural network; E. Uses the cost function to train and verify the neural network; F. Uses the trained neural network to input 15 bucket position parameters for prediction, selects the optimal bucket position parameter group after actual operation verification, and realizes the optimization of the loader shoveling operation process. The present invention can optimize the loader shoveling operation process and has good application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of loader design and manufacturing, and in particular to a method for optimizing a loader shoveling operation process based on machine learning. Background Art

[0002] A loader is a kind of earthwork construction machinery widely used in construction projects such as highways, railways, buildings, hydropower, ports, and mines. It is mainly used to shovel and load bulk materials such as soil, sand, lime, coal, etc. It can also do light shoveling operations on ore, hard soil, etc. By replacing different auxiliary working devices, it can also carry out bulldozing, lifting, and loading and unloading operations of other materials such as wood. In the construction of roads, especially high-grade highways, loaders are used for filling and excavation of roadbed projects, asphalt mixtures, and cement concrete material fields. In addition, it can also carry out operations such as pushing soil, leveling the ground, and towing other machinery. Because loaders have the advantages of fast operating speed, high efficiency, good maneuverability, and easy operation, they have become one of the main types of earthwork construction in engineering construction.

[0003] At present, shoveling operations are all performed manually, which is labor-intensive, easy to fatigue operators, and low work efficiency. The existing automatic shoveling technology is still in the emerging stage and the technology is not mature enough. Loader shoveling trajectory planning is the basis for realizing automatic shoveling of loaders. At the same time, reasonable trajectory planning has a great impact on the energy consumption and operating efficiency of the loader operation process. The research on related trajectory optimization is mainly based on theoretical analysis and simulation research, and lacks practical verification. Summary of the invention

[0004] The present invention aims to provide a method for optimizing the loader shoveling operation process based on machine learning. The method can help enterprises and loader users to plan and optimize the loader shoveling trajectory according to the shoveling operation conditions, optimize the loader shoveling operation process, and achieve energy conservation and emission reduction in the shoveling operation process. The method has good application prospects.

[0005] The technical solution of the present invention is as follows:

[0006] The method for optimizing the loader shoveling operation process based on machine learning comprises the following steps:

[0007] A. Set the shoveling material, and the loader performs multiple on-site shoveling operations. Collect the three bucket position parameters of the loader vehicle displacement, boom cylinder displacement, and bucket cylinder displacement during each operation; and record the actual value of each shoveling weight, actual value of operation energy consumption, and actual value of operation time;

[0008] B. Based on the analysis of the three bucket position parameters of the loader vehicle displacement, boom cylinder displacement, and bucket cylinder displacement for each operation, the shoveling operation is divided into four stages: horizontal shoveling stage, bucket flipping shoveling stage, shoveling stage, and bucket lifting stage. 12 bucket position parameters are obtained for each operation;

[0009] From the displacement of the loader vehicle, the displacement of the boom cylinder, and the displacement of the bucket cylinder, three displacement parameters are extracted: the instantaneous speed of the loader shoveling into the pile, the initial length of the boom cylinder, and the initial length of the bucket cylinder;

[0010] C. Collect the 15 displacement parameters of each operation to form multiple bucket position parameter data groups, remove the data groups containing abnormal data, and then perform normalization processing to obtain multiple normalized bucket position parameter data groups, and construct a normalized bucket position parameter data set; respectively perform normalization processing on the actual value of shovel loading weight, actual value of operation energy consumption, and actual value of operation time corresponding to each group of data in the normalized bucket position parameter data set to obtain actual normalized value of shovel loading weight, actual normalized value of operation energy consumption, and actual normalized value of operation time;

[0011] D. Construct a BP neural network, which includes an input layer, four hidden layers and an output layer; the input layer includes 15 nodes; the four hidden layers include 11, 9, 7 and 5 nodes respectively; the output layer includes 3 nodes, corresponding to the shovel weight, operation energy consumption and operation time respectively;

[0012] Among them, the 15 displacement parameters of each bucket position parameter data group in the normalized bucket position parameter data set are respectively input into each node of the first hidden layer, weighted and accumulated by the bias and the 15 displacement parameters, and the calculation result is obtained through the activation function operation; each node of the other hidden layers and the output layer is weighted and accumulated by the bias and all the data output by the previous layer, and the calculation result is obtained through the activation function operation; the initial data weight coefficient and bias are randomly generated;

[0013] E. Construct a cost function, and train and verify the neural network through back propagation according to the cost function to obtain a trained neural network;

[0014] Using the trained neural network, 15 ideal bucket position parameters are set to predict the three optimization target parameters of shoveling weight, operating energy consumption, and operating time. After verification by actual operation, the optimal bucket position parameter group is selected for shoveling trajectory planning, thereby optimizing the loader's shoveling operation process.

[0015] In step C, the process of removing the data group containing abnormal data is as follows:

[0016] The 15 bucket position parameters were sorted by size and their normal distribution was analyzed. A 95% confidence analysis was performed to treat the remaining 5% of the data in each group of parameters as abnormal data and the data groups containing these abnormal data were deleted.

[0017] In step C, the process of removing the array containing abnormal data and then performing normalization processing is as follows:

[0018] The formula for normalization of mapping the middle value of each group of sorting to 0.5, the maximum value to 0.975, and the minimum value to 0.025 is:

[0019] (1).

[0020] In each node in the hidden layer, each input value is weighted and superimposed, and then the bias value is added, and then the sigmoid function is used to operate and output to each node in the next layer; the output value of the last hidden layer is weighted and accumulated, and the bias value is added, and then the sigmoid function is used to operate to obtain the predicted value of shoveling weight, predicted value of operation energy consumption, and predicted value of operation time.

[0021] In the step E, the cost function formula is:

[0022] (2)

[0023] in, is the optimized target parameter value obtained for each test, is the optimized target parameter value predicted by the neural network.

[0024] In step E, the process of training and verifying the neural network using the cost function is as follows:

[0025] a. Randomly select 60%-80% of the normalized displacement data sets from the normalized displacement data sets and input them into the neural network in sequence;

[0026] b. Each time the input is obtained, the shovel weight output value, operation energy consumption output value, and operation time output value of the corresponding operation are compared with the corresponding actual shovel weight value, actual operation energy consumption value, and actual operation time value to obtain the shovel weight error value, operation energy consumption error value, and operation energy consumption error value of this time;

[0027] c. Then input the next normalized displacement data set and repeat step b until all randomly selected data are input. Then optimize the neural network through back propagation according to the cost function. Figure 3 For the optimization process with different learning rates;

[0028] d. Use the remaining 20-40% of the normalized displacement data set for verification, input this part of the normalized displacement data set into the trained neural network respectively, and obtain the predicted value of the shoveling weight, the predicted value of the operation energy consumption, and the predicted value of the operation time for each operation; and normalize and restore the predicted value of each operation, and compare it with the corresponding actual value of the shoveling weight, actual value of the operation energy consumption, and actual value of the operation time; if the error of more than 90% of the data comparison is less than 15%, it means that a better neural network model is obtained, otherwise it is necessary to adjust the BP neural network structure, modify the number of hidden layers and the number of nodes, and then repeat step E.

[0029] In the step C, the data normalization reduction formula is:

[0030] (3).

[0031] The learning rate η of the neural network constructed in step D is 0.1.

[0032] The method of the present invention obtains shoveling operation performance data through test experiments, obtains the correlation mathematical model between the shoveling trajectory and the operation energy efficiency of the loader shoveling operation process by constructing a BP neural network, and uses the model to predict the loader shoveling trajectory, thereby helping enterprises and loaders to use customer shoveling operation conditions to plan and optimize the loader shoveling trajectory, optimize the loader shoveling operation process, and achieve energy conservation and emission reduction in the shoveling operation process, and has good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is an analysis diagram of the shoveling operation process based on the displacement of the entire vehicle and the displacement of the oil cylinder according to Example 1 of the present invention;

[0034] Figure 2 This is a neural network structure diagram of Example 1 of the present invention;

[0035] Figure 3 Graph showing the influence of the learning rate on error convergence in Example 1 of the present invention;

[0036] Figure 4 The error diagram of the training process output value and the test value of Example 1 of the present invention;

[0037] Figure 5 This is a verification error analysis diagram of Example 1 of the present invention. DETAILED DESCRIPTION

[0038] The present invention is described in detail below with reference to the accompanying drawings and embodiments. Example 1

[0039] This embodiment provides a method for optimizing a loader shoveling operation process based on machine learning, comprising the following steps:

[0040] A. Set the shoveling material, and the loader performs 1000 field shoveling operations. Collect the three bucket position parameters of the loader vehicle displacement, boom cylinder displacement, and bucket cylinder displacement during each operation; and record the actual value of the shoveling weight, actual value of the operation energy consumption, and actual value of the operation time;

[0041] B. Based on the analysis of the three bucket position parameters of the loader vehicle displacement, boom cylinder displacement, and bucket cylinder displacement for each operation, the shoveling operation is divided into four stages: horizontal shoveling stage, bucket flipping shoveling stage, shoveling stage, and bucket lifting stage. 12 bucket position parameters are obtained for each operation;

[0042] From the displacement of the loader vehicle, the displacement of the boom cylinder, and the displacement of the bucket cylinder, three displacement parameters are extracted: the instantaneous speed of the loader shoveling into the pile, the initial length of the boom cylinder, and the initial length of the bucket cylinder;

[0043] Figure 1 It reflects the division of four stages: (1) Horizontal shoveling OA stage, the length of the boom cylinder and the bucket cylinder remains unchanged, the bucket does not rotate relative to the whole loader and shovels into the pile as the whole vehicle moves forward; (2) Bucket flipping shoveling AB stage, the length of the boom cylinder does not increase much, the length of the bucket cylinder increases significantly, and the bucket completes the flipping action under the push of the bucket cylinder; (3) Shoveling BC stage, the length of the boom cylinder and the bucket cylinder remains unchanged, and the bucket moves forward a small distance with the inertia of the whole vehicle; (4) Bucket lifting CD stage, the length of the bucket cylinder remains basically unchanged, the length of the boom cylinder increases significantly, and the bucket completes the lifting action under the push of the boom cylinder and leaves the pile;

[0044] C. Collect the 15 displacement parameters of each operation to form 1000 bucket position parameter data groups. Sort the 15 bucket position parameters by size and analyze their normal distribution. Perform 95% confidence analysis to treat the remaining 5% of the data in each parameter group as abnormal data and delete the data groups containing these abnormal data. After normalization, 807 normalized bucket position parameter data groups are obtained to construct a normalized bucket position parameter data set. The actual values ​​of shovel weight, actual values ​​of operating energy consumption, and actual values ​​of operating time corresponding to the 807 normalized bucket position parameter data groups are normalized to obtain actual normalized values ​​of shovel weight, actual normalized values ​​of operating energy consumption, and actual normalized values ​​of operating time.

[0045] The normalization process is:

[0046] Map the middle value of each group of sorts to 0.5, the maximum value to 0.975, and the minimum value to 0.025. The normalization formula is:

[0047] (1);

[0048] D. Construct a BP neural network, which includes 4 hidden layers connected in sequence and an output layer connected to the last hidden layer; the 4 hidden layers include 11, 9, 7, and 5 nodes respectively; the output layer includes 3 nodes, corresponding to the shovel weight, operation energy consumption, and operation time respectively;

[0049] Among them, the 15 displacement parameters of each bucket position parameter data group in the normalized bucket position parameter data set are respectively input into each node of the first hidden layer, weighted and accumulated by the bias and the 15 displacement parameters, and the calculation result is obtained through the activation function operation; each node of the other hidden layers and the output layer is weighted and accumulated by the bias and all the data output by the previous layer, and the calculation result is obtained through the activation function operation; the initial data weight coefficient and bias are randomly generated;

[0050] In each node in the hidden layer, each input value is weighted and superimposed, and then the bias value is added, and then the sigmoid function is used to operate and output to each node in the next layer; the output value of the last hidden layer is weighted and accumulated, and the bias value is added, and then the sigmoid function is used to operate to obtain the predicted value of shoveling weight, predicted value of operation energy consumption, and predicted value of operation time.

[0051] In this embodiment, the neural network optimization test was performed with η=0.2, η=0.05, and η=0.1 respectively. The results are as follows: Figure 3 As shown in the figure, when the learning rate is too large (η=0.2), the error cannot converge well, and when the learning rate is too small (η=0.05), the convergence speed is relatively slow and the number of iterations increases significantly; therefore, after analysis and comparison, the learning rate of the neural network in this step is η=0.1;

[0052] E. Construct a cost function and use the cost function to train and verify the neural network to obtain a trained neural network;

[0053] Among them, the cost function formula is:

[0054] (2)

[0055] in, is the optimized target parameter value obtained for each test, is the optimized target parameter value predicted by the neural network.

[0056] The process of training and verifying the neural network using the cost function is:

[0057] Randomly select 70% of the normalized displacement data sets in the normalized displacement data set and input them into the neural network in turn;

[0058] b. Each time the input is obtained, the shovel weight output value, operation energy consumption output value, and operation time output value of the corresponding operation are compared with the corresponding actual shovel weight value, actual operation energy consumption value, and actual operation time value to obtain the shovel weight error value, operation energy consumption error value, and operation energy consumption error value of this time;

[0059] c. Then input the next normalized displacement data set and repeat step b until all randomly selected data are input. Then optimize the neural network through back propagation according to the cost function. Figure 4 is the optimization process with different learning rates; Figure 4 It can be seen that after calculation by the neural network model, the error between the parameters such as shovel weight, operating energy consumption and operating time output by most of the training data and the actual test parameters is less than 15%, and only a very small number of data have an error greater than 25%.

[0060] d. Use the remaining 30% of the normalized displacement data set for verification. Input this part of the normalized displacement data set into the trained neural network to obtain the predicted value of the shovel weight, the predicted value of the operation energy consumption, and the predicted value of the operation time for each operation; and normalize and restore the predicted value of each operation, and compare it with the corresponding actual value of the shovel weight, actual value of the operation energy consumption, and actual value of the operation time;

[0061] Verification error analysis can be seen Figure 5 .Depend on Figure 5 It can be seen that after verification of the remaining 30% of the data, more than 90% of the data, such as shovel weight, operating energy consumption and operating time, calculated and output by the neural network model, had an error of less than 15% between the actual test values, indicating that the training of the neural network model was completed.

[0062] In the step b, the data normalization reduction formula is:

[0063] (3).

[0064] F. Using the trained neural network, set 15 ideal bucket position parameters to predict the three optimization target parameters of shoveling weight, operating energy consumption, and operating time. After verification by actual operations, select the optimal bucket position parameter group to optimize the loader's shoveling operation process.

Claims

1. A method for optimizing the shovel loading process of a loader based on machine learning, characterized in that The following steps are involved: A. Set the shoveling material, and the loader performs multiple on-site shoveling operations. Collect the three bucket position parameters of the loader vehicle displacement, boom cylinder displacement, and bucket cylinder displacement during each operation; and record the actual value of each shoveling weight, actual value of operation energy consumption, and actual value of operation time; B. Based on the analysis of the three bucket position parameters of the loader vehicle displacement, boom cylinder displacement, and bucket cylinder displacement for each operation, the shoveling operation is divided into four stages: horizontal shoveling stage, bucket flipping shoveling stage, shoveling stage, and bucket lifting stage. 12 bucket position parameters are obtained for each operation; From the displacement of the loader vehicle, the displacement of the boom cylinder, and the displacement of the bucket cylinder, three displacement parameters are extracted: the instantaneous speed of the loader shoveling into the pile, the initial length of the boom cylinder, and the initial length of the bucket cylinder; C. Collect the 15 displacement parameters of each operation to form multiple bucket position parameter data groups, remove the data groups containing abnormal data, and then perform normalization processing to obtain multiple normalized bucket position parameter data groups, and construct a normalized bucket position parameter data set; respectively perform normalization processing on the actual value of shovel loading weight, actual value of operation energy consumption, and actual value of operation time corresponding to each group of data in the normalized bucket position parameter data set to obtain actual normalized value of shovel loading weight, actual normalized value of operation energy consumption, and actual normalized value of operation time; D. Construct a BP neural network, which includes 4 hidden layers connected in sequence and an output layer connected to the last hidden layer; the 4 hidden layers include 11, 9, 7, and 5 nodes respectively; the output layer includes 3 nodes, corresponding to the shovel weight, operation energy consumption, and operation time respectively; Among them, the 15 displacement parameters of each bucket position parameter data group in the normalized bucket position parameter data set are respectively input into each node of the first hidden layer, weighted and accumulated by the bias and the 15 displacement parameters, and the calculation result is obtained through the activation function operation; each node of the other hidden layers and the output layer is weighted and accumulated by the bias and all the data output by the previous layer, and the calculation result is obtained through the activation function operation; the initial data weight coefficient and bias are randomly generated; E. Construct a cost function and use the cost function to train and verify the neural network to obtain a trained neural network; Using the trained neural network, 15 ideal bucket position parameters are set to predict the three optimization target parameters of shoveling weight, operating energy consumption, and operating time. After verification by actual operation, the optimal bucket position parameter group is selected to optimize the loader's shoveling operation process.

2. The method for optimizing the loader shovel loading process based on machine learning according to claim 1, characterized in that: In step C, the process of removing the data group containing abnormal data is as follows: The 15 bucket position parameters were sorted by size and their normal distribution was analyzed. A 95% confidence analysis was performed to treat the remaining 5% of the data in each group of parameters as abnormal data and the data groups containing these abnormal data were deleted.

3. The method for optimizing the shoveling operation process of a loader based on machine learning according to claim 1, characterized in that: In each node in the hidden layer, each input value is weighted and superimposed, and then the bias value is added, and then the sigmoid function is used to operate and output to each node in the next layer; the output value of the last hidden layer is weighted and accumulated, and the bias value is added, and then the sigmoid function is used to operate to obtain the predicted value of shoveling weight, predicted value of operation energy consumption, and predicted value of operation time.

4. The method for optimizing the loader shoveling operation process based on machine learning according to claim 1, characterized in that: In the step E, the cost function formula is: (2) in, is the optimized target parameter value obtained for each test, is the optimized target parameter value predicted by the neural network.

5. The method for optimizing the shoveling operation process of a loader based on machine learning according to claim 4, characterized in that: In step E, the process of training and verifying the neural network using the cost function is as follows: a. Randomly select 60%-80% of the normalized bucket position parameter data sets from the normalized bucket position parameter data sets and input them into the neural network in sequence; b. Each time the input is obtained, the shovel weight output value, operation energy consumption output value, and operation time output value of the corresponding operation are compared with the corresponding actual shovel weight value, actual operation energy consumption value, and actual operation time value to obtain the shovel weight error value, operation energy consumption error value, and operation energy consumption error value of this time; c. Then input the next normalized bucket position parameter data set, repeat step b until all randomly selected data are input, and then optimize the neural network through back propagation according to the cost function; d. Use the remaining 20-40% of the normalized bucket position parameter data set for verification, and input this part of the normalized bucket position parameter data set into the trained neural network to obtain the predicted value of the shovel weight, the predicted value of the operation energy consumption, and the predicted value of the operation time for each operation; After normalizing the predicted value of each operation, compare it with the corresponding actual value of shoveling weight, actual value of operation energy consumption, and actual value of operation time; if the error of more than 90% of the data comparison is less than 15%, it means that the neural network model is successfully established, otherwise it is necessary to adjust the BP neural network structure, modify the number of hidden layers and the number of nodes, and then repeat step E.

6. The method for optimizing the shoveling operation process of a loader based on machine learning according to claim 1, characterized in that: The learning rate η of the neural network constructed in step D is 0.1.

Images