A method for autonomous repair decision of a wear part

Abstract

The application discloses a kind of self-repairing decision methods of wear component, the historical detection data of wear component maintenance is standardized, and wear component degradation state data is constructed;The principal component analysis is carried out to wear component degradation state data, and wear component degradation feature is screened, and wear component degradation characteristic parameter is extracted;Wear component autonomous maintenance decision model is constructed using multitask learning neural network, and maintenance judgment task and maintenance quantity decision task are simultaneously executed;Wear component autonomous maintenance decision model is trained using wear component degradation characteristic parameter;The neural network parameter of wear component autonomous maintenance decision model is learned;Based on the improved hierarchical grid algorithm, the super parameter of autonomous learning model is learned, and the approximate optimal solution is globally searched.This application uses multitask learning neural network and improved hierarchical grid algorithm to make autonomous decision, which helps the implementation of wear component state repair, improves the economy and safety of wear component maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of repairable equipment maintenance management, and in particular to an autonomous maintenance decision-making method for worn parts. Background Art

[0002] Wear parts wear out over long periods of use. Excessive wear can easily cause significant losses, such as equipment downtime. To ensure safe equipment operation, effective maintenance measures are required to restore worn parts to their standard profile. Currently, equipment maintenance systems are gradually shifting toward condition-based repairs, where maintenance timing decisions are made based on the condition and remaining life of worn parts. Consequently, when worn parts enter the repair shop, repair operations rely on the engineer's experience, often leading to improper maintenance operations, wasted maintenance, or even unscheduled repairs. Therefore, there is an urgent need to objectively implement wear part repair operations, enable autonomous maintenance decisions for wear parts, construct autonomous maintenance decision-making models, and enable autonomous learning of model parameters to improve the economic efficiency of wear part repairs. Summary of the Invention

[0003] The object of the present invention is to provide a method for autonomous maintenance decision-making of worn parts, so as to solve the problem of autonomous decision-making of maintenance operations based on the degradation state of worn parts, thereby supporting the implementation of condition-based maintenance of worn parts.

[0004] The object of the present invention is achieved through the following technical solutions:

[0005] A method for autonomous maintenance decision-making of worn parts, comprising the following steps:

[0006] (1) Standardize the historical inspection data of wear parts maintenance and construct the wear parts degradation status data;

[0007] (2) Perform principal component analysis on the wear parts degradation status data, screen the wear parts degradation characteristics, and extract the wear parts degradation characteristic parameters;

[0008] (3) A multi-task learning neural network is used to construct an autonomous maintenance decision model for worn parts. The autonomous maintenance decision model for worn parts is used to simultaneously execute the maintenance judgment task and the maintenance quantity decision task. The maintenance judgment task is used to determine whether the contour of the worn part needs maintenance; the maintenance quantity decision task is used to determine the maintenance quantity or cutting amount of the worn part and restore the standard contour of the worn part. The maintenance judgment task assists the execution of the maintenance quantity decision task.

[0009] (4) Using degradation characteristic parameters of worn parts to train the autonomous maintenance decision model for worn parts;

[0010] (5) Learning the neural network parameters of the autonomous maintenance decision model for worn parts;

[0011] (6) Based on the improved hierarchical grid algorithm, the model hyperparameters are autonomously learned and the approximate optimal solution is globally searched.

[0012] Preferably, the structure of the autonomous maintenance decision model for worn parts is as follows:

[0013] The first layer group has only one input layer, which contains k input neurons and inputs the degradation characteristic parameters of the worn parts;

[0014] The second layer group shares feature information and contains 3 hidden layers:

[0015] The 2-1 layer, the hidden layer contains 8 neurons, the activation function is tanh, and some nodes are dropped out during training;

[0016] The 2-2 layer, the hidden layer contains 16 neurons, the activation function is tanh, and some nodes are dropped out during training;

[0017] The 2nd and 3rd layers, the hidden layer, contains 8 neurons, the activation function is tanh, and some nodes are dropped out during training;

[0018] The third layer group, the classification task layer, contains 3 hidden layers:

[0019] The 3-1 layer, the hidden layer contains 16 neurons, the activation function is ReLU, and some nodes are dropped out during training;

[0020] The 3-2 layer, the hidden layer contains 8 neurons, the activation function is ReLU, and some nodes are dropped out during training;

[0021] Layer 3-3, classification result output layer, activation function is sigmoid, contains 1 neuron, outputs maintenance judgment vector

[0022] The 4th layer group, which combines the neurons of the 3rd and 3rd layers of maintenance judgment and the 2nd and 3rd layers of shared feature information, contains 9 neurons;

[0023] The fifth layer group, the regression task layer, contains 3 hidden layers:

[0024] The 5-1 layer, the hidden layer contains 16 neurons, the activation function is tanh, and some nodes are dropped out during training;

[0025] The 5-2 layer, the hidden layer contains 8 neurons, the activation function is tanh, and some nodes are dropped out during training;

[0026] Layer 5-3, regression result output layer, contains 1 neuron and outputs the maintenance amount vector

[0027] Preferably, the computational process for training the autonomous maintenance decision model for worn parts is as follows:

[0028] Step S1, the input layer inputs each group of wear parts X in the wear part degradation characteristic parameter matrix;

[0029] Step S21, the weight matrix of the 2-1 layer is w (11) , the bias vector is b (11) , dropout vector r (11) ~Bernoulli(p 11 ), then the output matrix of the 2-1 hidden layer is:

[0030]

[0031] Step S22, the weight matrix of the 2-2 layer is w (12) , the bias vector is b (12) , dropout vector r (12) ~Bernoulli(p 12 ), then the output matrix of the 2-2 hidden layer is:

[0032]

[0033] Step S23, the weight matrix of the 2nd-3rd layer is w (13) , the bias vector is b (13) , dropout vector r (13) ~Bernoulli(p 13 ), then the output matrix of the 2nd-3rd hidden layer is:

[0034]

[0035] Step S31, the weight matrix of the 3-1 layer is w (21) , the bias vector is b (21) , dropout vector r (21) ~Bernoulli(p 21 ), then the output matrix of the 3-1 hidden layer is:

[0036]

[0037] Step S32, the weight matrix of the 3-2 layer is w (22) , the bias vector is b (22) , dropout vector r (22) ~Bernoulli(p 22 ), then the output matrix of the 3-2 hidden layer is:

[0038]

[0039] Step S33, the weight matrix of the 3-3 layer is w (23) , the bias vector is b (23) , then the 3-3 layer maintenance judgment output vector is:

[0040]

[0041] And Round off to the nearest integer,

[0042] Step S4: Merge the maintenance judgment layer 3-3 and the shared feature information layer 2-3

[0043] Step S51, the weight matrix of the 5-1 layer is w (31) , the bias vector is b (31) , dropout vector r (31) ~Bernoulli(p 31 ), then the output matrix of the 5-1 hidden layer is:

[0044]

[0045] Step S52, the weight matrix of the 5-2 layer is w (32) , the bias vector is b (32) , dropout vector r (32) ~Bernoulli(p 32 ), then the output matrix of the 5-2 hidden layer is:

[0046]

[0047] Step S52, the weight matrix of the 5-2 layer is w (33) , the bias vector is b (33) , then the 5-3 maintenance output vector is:

[0048]

[0049] Preferably, the method for learning the neural network parameters of the wear component autonomous maintenance decision model is as follows:

[0050] First, the degradation characteristic parameters of the worn parts are divided into a training data set and a test data set, where the sample ratio of the test data set is r t ; Secondly, cross entropy and mean square error are used to construct the loss function Loss of the classification task layer respectively C And the loss function Loss of the regression task layer R, the joint loss function Loss of multi-task learning; finally, the Adam optimizer is used to optimize the loss function Loss and the learning rate of the optimizer is r l .

[0051] Preferably, the model hyperparameters include the proportion of the test data set r t and the learning rate r of the Adam optimizer l ,The steps of autonomous learning of the hyper-parameters of the autonomous learning model based on the improved hierarchical grid algorithm are as follows:

[0052] Phase 1: Search for extreme value areas:

[0053] Step S11, set the current initial extreme point, construct a square lattice with a side length of 2t+1 with the extreme point as the center, and calculate the (2t+1) square lattice inside the lattice. 2 The loss function value at the point position is used to obtain the extreme point in the lattice;

[0054] Step S12: If the extreme point is located at the edge of the lattice, repeat step S11 with the extreme point as the center; if the extreme point is located inside the lattice, narrow the lattice for searching, set t = t-1, and go to step S21;

[0055] The second stage: Search for extreme points in the extreme area:

[0056] Step S21: construct a square lattice with a side length of 2t+1 with the current extreme point as the center, calculate the loss function value in the lattice, and obtain the extreme point in the lattice;

[0057] Step S22: If the extreme point is located at the edge of the lattice, repeat step S21 with the extreme point as the center; if the extreme point is located inside the lattice, proceed to step S23;

[0058] Step S23, let t = t-1, if t = 1, the iteration is terminated and the current extreme point and extreme value are output; if t> 1, repeat steps S21 to S23.

[0059] The beneficial effects of the present invention are:

[0060] This avoids improper maintenance operations caused by engineer experience, enables autonomous decision-making for wear part repairs, and improves the automation level of maintenance decision-making. Based on actual wear part maintenance data, the autonomous maintenance decision-making model has a maintenance judgment accuracy rate exceeding 97%, and an estimated maintenance error of less than 0.5. These technical indicators meet actual maintenance requirements and can be applied to the maintenance of wear parts on repairable equipment. The maintenance decision-making system not only autonomously determines maintenance operations for wear parts but also autonomously learns maintenance model parameters, making wear part maintenance decisions intelligent, facilitating the implementation of condition-based maintenance for wear parts, and improving the economic and safety of wear part maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of the autonomous maintenance decision-making method for worn parts;

[0062] Figure 2 A diagram of the neural network architecture for multi-task learning maintenance decision-making models;

[0063] Figure 3 Schematic diagram of the extreme value region search in the first stage of the improved hierarchical grid algorithm;

[0064] Figure 4 Schematic diagram of searching for extreme values ​​in extreme value regions in the second stage of the improved hierarchical grid algorithm;

[0065] Figure 5 is the weight matrix of layer 2-1 and the bias vector

[0066] Figure 6 is the weight matrix of layer 2-2 and the bias vector

[0067] Figure 7 is the weight matrix of the 2nd-3rd layer and the bias vector

[0068] Figure 8 is the weight matrix of layer 3-1 and the bias vector

[0069] Figure 9 is the weight matrix of layer 3-2 and the bias vector

[0070] Figure 10 is the weight matrix of layer 3-3 and the bias vector

[0071] Figure 11 is the weight matrix of layer 5-1 and the bias vector

[0072] Figure 12 is the weight matrix of layer 5-2 and the bias vector

[0073] Figure 13 is the weight matrix of layer 5-3 and the bias vector

[0074] exist Figure 3and Figure 4 In the figure, the horizontal and vertical axes represent the two hyperparameters rt and rl to be optimized respectively. DETAILED DESCRIPTION

[0075] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0076] See also Figure 1 As shown, the present embodiment shows a method for autonomous maintenance decision-making of a worn component, comprising the following six steps:

[0077] (1) According to formula (a), the historical inspection data D of the wear parts maintenance m×n Perform standardization to make the wear parts status data dimensionless and construct the wear parts degradation status data S m×n .

[0078]

[0079] S ij For S m×n The normalized value of the jth state parameter of the i-th group of wear parts, i = 1, 2, ..., m; j = 1, 2, ..., n.

[0080] (2) For S m×n Perform principal component analysis to obtain the transformation matrix T n×k Then, according to formula (b), the degradation characteristics of the wear parts with large information content are extracted to obtain the wear parts degradation characteristic parameter matrix I m×k :

[0081] I m×k =S m×n T n×k (b)

[0082] Where 1≤k≤n, k represents the number of features after screening.

[0083] (3) Using a multi-task learning neural network, an autonomous maintenance decision model for wear parts is constructed. The autonomous maintenance decision model for wear parts is used to simultaneously perform the maintenance judgment task (binary classification task) and the maintenance quantity decision task (regression task). Among them, the maintenance judgment task is used to determine whether the contour of the worn part needs to be repaired; the maintenance quantity decision task is used to determine the repair quantity or cutting amount of the worn part and restore the standard contour of the worn part; the maintenance judgment task assists the execution of the maintenance quantity decision task. The structure of the autonomous maintenance decision model for wear parts based on the multi-task learning neural network is as follows:

[0084] The first layer group has only one input layer, which contains k input neurons and inputs the degradation characteristic parameter matrix I of the worn parts. m×k .

[0085] The second layer group shares feature information and contains 3 hidden layers:

[0086] The 2-1 layer, the hidden layer contains 8 neurons, and the activation function is calculated according to formula (c), with probability p 11 = 0.1 to dropout some nodes;

[0087]

[0088] Here, x represents the input data, which will be described in detail in “(4) Training the autonomous maintenance decision model for worn parts”.

[0089] The 2-2 layer, the hidden layer contains 16 neurons, and the activation function is calculated according to formula (c), with probability p 12 = 0.1 to dropout some nodes;

[0090] The 2nd-3rd layer, the hidden layer contains 8 neurons, and the activation function is calculated according to formula (c), with probability p 13 = 0.1 to dropout some nodes.

[0091] The third layer group, the classification task layer, contains 3 hidden layers:

[0092] The 3-1 layer, the hidden layer contains 16 neurons, and the activation function is calculated according to formula (d), with probability p 21 = 0.1 to dropout some nodes;

[0093] f relu (x)=max(0,x) (d)

[0094] The 3-2 layer, the hidden layer contains 8 neurons, and the activation function is calculated according to formula (d), with probability p 22 = 0.1 to dropout some nodes;

[0095] Layer 3-3, classification result output layer, activation function is calculated according to formula (e), contains 1 neuron, and outputs the maintenance judgment vector

[0096]

[0097] The 4th layer group, which combines the neurons of the 3rd to 3rd layers of repair judgment and the 2nd to 3rd layers of shared feature information, contains 9 neurons.

[0098] The fifth layer group, the regression task layer, contains 3 hidden layers:

[0099] The 5-1 layer, the hidden layer contains 16 neurons, and the activation function is calculated according to formula (c), with probability p 31= 0.1 to dropout some nodes;

[0100] The 5-2 layer, the hidden layer contains 8 neurons, and the activation function is calculated according to formula (c), with probability p 32 = 0.1 to dropout some nodes;

[0101] Layer 5-3, regression result output layer, contains 1 neuron and outputs the maintenance amount vector

[0102] (4) Train the autonomous maintenance decision model for worn parts. The calculation steps are as follows:

[0103] Step S1: Input the wear component degradation characteristic parameter matrix of each group of wear components X=I l , 1≤l≤m.

[0104] Step S21, the weight matrix of the 2-1 layer is w (11) , the bias vector is b (11) , dropout vector r (11) ~Bernoulli(p 11 ), then the output matrix of the 2-1 hidden layer is:

[0105]

[0106] Step S22, the weight matrix of the 2-2 layer is w (12) , the bias vector is b (12) , dropout vector r (12) ~Bernoulli(p 12 ), then the output matrix of the 2-2 hidden layer is:

[0107]

[0108] Step S23, the weight matrix of the 2nd-3rd layer is w (13) , the bias vector is b (13) , dropout vector r (13) ~Bernoulli(p 13 ), then the output matrix of the 2nd-3rd hidden layer is:

[0109]

[0110] Step S31, the weight matrix of the 3-1 layer is w (21) , the bias vector is b (21) , dropout vector r (21) ~Bernoulli(p 21 ), then the output matrix of the 3-1 hidden layer is:

[0111]

[0112] Step S32, the weight matrix of the 3-2 layer is w (22) , the bias vector is b (22) , dropout vector r (22) ~Bernoulli(p 22 ), then the output matrix of the 3-2 hidden layer is:

[0113]

[0114] Step S33, the weight matrix of the 3-3 layer is w (23) , the bias vector is b (23) , then the 3-3 layer maintenance judgment output vector is:

[0115]

[0116] And Round off to the nearest integer,

[0117] Step S4: Merge the maintenance judgment layer 3-3 and the shared feature information layer 2-3

[0118] Step S51, the weight matrix of the 5-1 layer is w (31) , the bias vector is b (31) , dropout vector r (31) ~Bernoulli(p 31 ), then the output matrix of the 5-1 hidden layer is:

[0119]

[0120] Step S52, the weight matrix of the 5-2 layer is w (32) , the bias vector is b (32) , dropout vector r (32) ~Bernoulli(p 32 ), then the output matrix of the 5-2 hidden layer is:

[0121]

[0122] Step S52, the weight matrix of the 5-2 layer is w (33) , the bias vector is b (33) , then the 5-3 maintenance output vector is:

[0123]

[0124] (5) The neural network parameters (weight matrix w and bias vector b) of the wear part autonomous maintenance decision model are learned as follows:

[0125] Based on the above neural network structure, the degradation characteristic parameters of the worn parts I m×k Divide into training data set and test data set, let the sample ratio of the test data set be r t ; Secondly, cross entropy and mean square error are used to construct the loss function Loss of the classification task respectively C And the loss function Loss of the regression task R , the cross entropy is calculated according to formula (f), the mean square error is calculated according to formula (g), and the joint loss function Loss of multi-task learning is calculated according to formula (h); finally, the Adam optimizer is used to optimize the loss function, and the learning rate of the Adam optimizer is r l , β1=0.9, β2=0.999, the iteration termination condition is ε adam =1E-7.

[0126]

[0127]

[0128] Loss(w,b)=Loss C (w,b)+Loss R (w,b)(h)

[0129] During the model training process, the small-scale training sample size is set to 100 and the maximum number of iterations is set to 1000.

[0130] Except for the sample proportion r t and the learning rate of the Adam optimizer is r l The two hyperparameters are more sensitive to the neural network model parameters, and the other hyperparameters are set to fixed values. In order to achieve autonomous learning of model parameters, the hyperparameter r is established. t and r l The objective function J(r t ,r l ), calculated according to formula (i), and optimized using the improved hierarchical grid algorithm to minimize the objective function.

[0131]

[0132] in, and Determine the hyperparameter r t ,r l In this case, the weight matrix and bias vector are optimized based on the training data set, and the objective function J(r t ,rl ) is based on the test data set D test The loss function of .

[0133] (6) In order to obtain the best neural network test results, it is necessary to optimize the model hyperparameters. Therefore, based on the improved hierarchical grid algorithm, the model hyperparameters are learned autonomously and the approximate optimal solution is globally searched. The specific steps are as follows:

[0134] Phase 1: Search for extreme value areas:

[0135] Step S11, set the current initial extreme point, construct a square lattice with a side length of 2t+1 with the extreme point as the center, and calculate the (2t+1) square lattice inside the lattice. 2 The loss function value at the point position is used to obtain the extreme point in the lattice.

[0136] Step S12: If the extreme point is located at the edge of the lattice, repeat step S11 with the extreme point as the center; if the extreme point is located inside the lattice, narrow the lattice for searching, set t=t-1, and go to step S21.

[0137] The second stage: Search for extreme points in the extreme area:

[0138] Step S21: construct a square lattice with a side length of 2t+1 with the current extreme point as the center, calculate the loss function value in the lattice, and obtain the extreme point in the lattice.

[0139] Step S22: If the extreme point is located at the edge of the lattice, repeat step S21 with the extreme point as the center; if the extreme point is located inside the lattice, go to step S23.

[0140] Step S23, let t = t-1, if t = 1, the iteration is terminated and the current extreme point and extreme value are output; if t> 1, repeat steps S21 to S23.

[0141] To ensure data confidentiality, the example of a railway freight car in the civilian field is used to describe the embodiment. However, the autonomous maintenance decision-making method for wear parts provided by the present invention is not limited to freight car wheels, but can be applied to maintenance decisions for wear parts of repairable equipment in similar scenarios.

[0142] By collecting historical maintenance data for C80 railway freight car wheelsets and analyzing maintenance documents, we identified maintenance status variables relevant to maintenance decisions. Table 1 shows a sample of wheel maintenance data related to wheelset maintenance decisions. Numbers 1 through 6 represent wheel degradation status variables, while numbers 7 through 8 represent wheel maintenance decision variables (maintenance judgment and maintenance amount).

[0143] Table 1 Partial sample of wheel maintenance data related to wheelset maintenance decision

[0144] Serial number Wheel state variables Example unit 1 <![CDATA[Travel time D m1 > 800 sky 2 <![CDATA[Left wheel tread wear D m2 > 3.2 mm 3 <![CDATA[Right wheel tread wear D m3 > 2 mm 4 <![CDATA[Absolute value D of wheel diameter difference m4 > 0.4 mm 5 <![CDATA[Left wheel rim thickness D m5 > 33.1 mm 6 <![CDATA[Right wheel flange thickness D m6s > 31.5 mm 7 Maintenance judgment C 1 - 8 Maintenance amount R 3.2 mm

[0145] On the basis of obtaining wheel maintenance history data, the wheel degradation status data is standardized, and then feature extraction is performed based on principal component analysis. The transformation matrix T is obtained according to principal component analysis. n×k , as shown in Table 2, and the characteristic parameters are calculated according to formula (b).

[0146] Table 2 Transformation matrix T based on principal component analysis n×k

[0147] 0.339163 0.59887 0.224635 0.335498 0.602258 0.561118 0.019993 -0.13764 0.153852 -0.34102 0.55411 -0.02136 0.154519 0.162839 -0.46544 0.383463 -0.15598 0.388371 -0.78684 0.23499 0.33676 -0.31424 -0.75813 -0.03706 0.418544 0.051242 -0.71932 0.425587 0.465574 0.27195

[0148] Then, a multi-task learning neural network is used to build a wheelset autonomous maintenance decision model. The cross entropy and mean square error are used to construct the loss function of the maintenance judgment classification task and the loss function of the maintenance amount regression task respectively. The joint loss function of the two tasks is constructed in a 1:1 ratio and calculated according to formula (h). Based on the model construction, the r t The data of the model are divided into a test set and the rest are training sets. The model is trained using a batch training method with a batch size of 100. The Adam optimizer is used to optimize the neural network parameters (weight matrix w and bias vector b). The learning rate of the optimizer is set to r. l At the same time, the improved hierarchical grid algorithm is used to autonomously learn and optimize the hyperparameters of the neural network. t and r l , the optimization process is as follows Figure 3 and Figure 4 As shown in Table 3, after two-stage optimization, the optimal hyperparameters are obtained.

[0149] Table 3 Hyperparameters r based on the improved hierarchical grid algorithm learning optimization t and r l

[0150] Serial number Hyperparameters symbol Example 1 Training dataset ratio <![CDATA[r t ]]> 0.5875 2 Adam optimizer learning rate <![CDATA[r l ]]> 0.00175

[0151] In the optimal hyperparameter r t and r l Based on this, the multi-task learning neural network is optimized. The weight matrix w and bias vector b of each layer of the neural network and the calculation steps of each layer are as follows:

[0152] Step S1: Input the wheel degradation characteristic parameter matrix to the input layer, and calculate the wear components X=I. l , 1≤l≤15000;

[0153] Step S21, the weight matrix of the 2-1 layer is w (11) , the bias vector is b (11) , the output matrix of the 2-1 hidden layer is:

[0154]

[0155] Step S22, the weight matrix of the 2-2 layer is w (12) , the bias vector is b (12) , the output matrix of the 2-2 hidden layer is:

[0156]

[0157] Step S23, the weight matrix of the 2nd-3rd layer is w (13) , the bias vector is b (13) , the output matrix of the 2nd-3rd hidden layer is:

[0158]

[0159] Step S31, the weight matrix of the 3-1 layer is w (21) , the bias vector is b (21) , the output matrix of the 3-1 hidden layer is:

[0160]

[0161] Step S32, the weight matrix of the 3-2 layer is w (22) , the bias vector is b (22) , the output matrix of the 3-2 hidden layer is:

[0162]

[0163] Step S33, the weight matrix of the 3-3 layer is w (23) , the bias vector is b (23) , the 3-3 maintenance judgment output vector is:

[0164]

[0165] And Round off to the nearest integer,

[0166] Step S4: Merge the maintenance judgment layer 3-3 and the shared feature information layer 2-3

[0167] Step S51, the weight matrix of the 5-1 layer is w (31) , the bias vector is b (31) , the output matrix of the 5-1th hidden layer is:

[0168]

[0169] Step S52, the weight matrix of the 5-2 layer is w (32) , the bias vector is b (32), the output matrix of the 5-2 hidden layer is:

[0170]

[0171] Step S52, the weight matrix of the 5-2 layer is w (33) , the bias vector is b (33) , the 5-3 maintenance output vector is:

[0172]

[0173] The weight matrix and bias vector of the above neural network are as follows Figures 5 to 13 As shown in Figure 3. On this basis, a complete and optimized autonomous maintenance decision model for wear parts based on multi-task learning is obtained. By following the calculation steps of the autonomous maintenance decision model, the maintenance judgment and maintenance amount of the wheelset can be obtained.

[0174] It is understandable that those skilled in the art can make equivalent substitutions or changes based on the technical solution and inventive concept of the present invention, and all these changes or substitutions should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for autonomous maintenance decision-making of worn parts, characterized in that The following steps are involved: (1) Standardize the historical inspection data of wear parts maintenance and construct the wear parts degradation status data; (2) Perform principal component analysis on the wear parts degradation status data, screen the wear parts degradation characteristics, and extract the wear parts degradation characteristic parameters; (3) A multi-task learning neural network is used to construct an autonomous maintenance decision model for worn parts. The autonomous maintenance decision model for worn parts is used to simultaneously perform the maintenance judgment task and the maintenance quantity decision task. The maintenance judgment task is used to determine whether the contour of the worn part needs to be repaired. The maintenance quantity decision task is used to determine the repair quantity or cutting quantity of the worn part and restore the standard contour of the worn part. The maintenance judgment task assists the execution of the maintenance quantity decision task. The structure of the autonomous maintenance decision model for worn parts is as follows: The first layer group has only one input layer, which contains k input neurons and inputs the degradation characteristic parameters of the worn parts; The second layer group shares feature information and contains 3 hidden layers: The 2-1 layer, the hidden layer contains 8 neurons, the activation function is tanh, and some nodes are dropped out during training; The 2-2 layer, the hidden layer contains 16 neurons, the activation function is tanh, and some nodes are dropped out during training; The 2nd and 3rd layers, the hidden layer, contains 8 neurons, the activation function is tanh, and some nodes are dropped out during training; The third layer group, the classification task layer, contains 3 hidden layers: The 3-1 layer, the hidden layer contains 16 neurons, the activation function is ReLU, and some nodes are dropped out during training; The 3-2 layer, the hidden layer contains 8 neurons, the activation function is ReLU, and some nodes are dropped out during training; Layer 3-3, classification result output layer, activation function is sigmoid, contains 1 neuron, outputs maintenance judgment vector The 4th layer group, which combines the neurons of the 3rd and 3rd layers of maintenance judgment and the 2nd and 3rd layers of shared feature information, contains 9 neurons; The fifth layer group, the regression task layer, contains 3 hidden layers: The 5-1 layer, the hidden layer contains 16 neurons, the activation function is tanh, and some nodes are dropped out during training; The 5-2 layer, the hidden layer, contains 8 neurons, the activation function is tanh, and some nodes are dropped out during training; Layer 5-3, regression result output layer, contains 1 neuron and outputs the maintenance amount vector (4) Using degradation characteristic parameters of worn parts to train the autonomous maintenance decision model for worn parts; (5) Learning the neural network parameters of the autonomous maintenance decision model for worn parts; (6) Based on the improved hierarchical grid algorithm, the model hyperparameters are autonomously learned and the approximate optimal solution is globally searched.

2. The autonomous maintenance decision-making method for wear parts according to claim 1 is characterized in that The computational process for training the autonomous maintenance decision model for worn parts is as follows: Step S1, the input layer inputs each group of wear parts X in the wear part degradation characteristic parameter matrix; Step S21, the weight matrix of the 2-1 layer is w (11) , the bias vector is b (11) , dropout vector r (11) ~Bernoulli(p 11 ), then the output matrix of the 2-1 hidden layer is: Step S22, the weight matrix of the 2-2 layer is w (12) , the bias vector is b (12) , dropout vector r (12) ~Bernoulli(p 12 ), then the output matrix of the 2-2 hidden layer is: Step S23, the weight matrix of the 2nd-3rd layer is w (13) , the bias vector is b (13) , dropout vector r (13) ~Bernoulli(p 13 ), then the output matrix of the 2nd-3rd hidden layer is: Step S31, the weight matrix of the 3-1 layer is w (21) , the bias vector is b (21) , dropout vector r (21) ~Bernoulli(p 21 ), then the output matrix of the 3-1 hidden layer is: Step S32, the weight matrix of the 3-2 layer is w (22) , the bias vector is b (22) , dropout vector r (22) ~Bernoulli(p 22 ), then the output matrix of the 3-2 hidden layer is: Step S33, the weight matrix of the 3-3 layer is w (23) , the bias vector is b (23) , then the 3-3 layer maintenance judgment output vector is: And Round off to the nearest integer, Step S4: Merge the maintenance judgment layer 3-3 and the shared feature information layer 2-3 Step S51, the weight matrix of the 5-1 layer is w (31) , the bias vector is b (31) , dropout vector r (31) ~Bernoulli(p 31 ), then the output matrix of the 5-1 hidden layer is: Step S52, the weight matrix of the 5-2 layer is w (32) , the bias vector is b (32) , dropout vector r (32) ~Bernoulli(p 32 ), then the output matrix of the 5-2 hidden layer is: Step S52, the weight matrix of the 5-2 layer is w (33) , the bias vector is b (33) , then the 5-3 maintenance output vector is:

3. The autonomous maintenance decision-making method for wear parts according to claim 2 is characterized in that The method for learning the neural network parameters of the autonomous maintenance decision model for worn parts is as follows: First, the degradation characteristic parameters of the worn parts are divided into a training data set and a test data set, where the sample ratio of the test data set is r t ; Secondly, cross entropy and mean square error are used to construct the loss function Loss of the classification task layer respectively C And the loss function Loss of the regression task layer R , joint loss function Loss of multi-task learning; Finally, the Adam optimizer is used to optimize the loss function Loss and the learning rate of the optimizer is r l .

4. The autonomous maintenance decision-making method for wear parts according to claim 3 is characterized in that The model hyperparameter contains the proportion of the test data set r t and the learning rate r of the Adam optimizer l ,The steps of autonomous learning of the hyper-parameters of the autonomous learning model based on the improved hierarchical grid algorithm are as follows: Phase 1: Search for extreme value areas: Step S11, set the current initial extreme point, construct a square lattice with a side length of 2t+1 with the extreme point as the center, and calculate the (2t+1) square lattice inside the lattice. 2 The loss function value at the point position is used to obtain the extreme point in the lattice; Step S12: If the extreme point is located at the edge of the lattice, repeat step S11 with the extreme point as the center; if the extreme point is located inside the lattice, narrow the lattice for searching, set t = t-1, and go to step S21; The second stage: Search for extreme points in the extreme area: Step S21: construct a square lattice with a side length of 2t+1 with the current extreme point as the center, calculate the loss function value in the lattice, and obtain the extreme point in the lattice; Step S22: If the extreme point is located at the edge of the lattice, repeat step S21 with the extreme point as the center; if the extreme point is located inside the lattice, proceed to step S23; Step S23, let t = t-1, if t = 1, the iteration is terminated and the current extreme point and extreme value are output; if t> 1, repeat steps S21 to S23.

Images