Electric shovel fault diagnosis method based on sparse evolutionary algorithm and neural network

Abstract

The invention discloses an electric shovel fault diagnosis method based on a sparse evolutionary algorithm and a neural network. The method comprises the steps of 1, acquiring electric shovel data; 2, establishing a neural network; 3, generating a guide vector and selecting a parent by using a mating selection strategy in SPEA2; 4, clustering the decision variables into groups, and calculating the importance of each group according to the guide vector of the current population; 5, generating filial generations by using crossover mutation operators; and step 6, performing continuous iteration, selecting N solutions by using an environment selection strategy of SPEA2, and finally obtaining a group of solutions which meet the Pareto leading edge surface and have relatively high classification precision and relatively low neural network sparsity. Multiple solutions can be obtained through one-time operation to meet different electric shovel fault diagnosis requirements, and therefore the electric shovel fault diagnosis efficiency and precision can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent diagnosis of electric shovel faults, and particularly relates to an electric shovel fault diagnosis method based on a sparse evolutionary algorithm and a neural network. Background Art

[0002] In open-pit mine production operations, electric shovels are important mining equipment. The bucket of the electric shovel interacts with the ore and needs to withstand impact loads, as well as shear and extrusion forces. The components fixed together on the bucket are prone to fracture, loosening or detachment, resulting in various faults. Existing neural network-based fault detection methods have achieved remarkable results in electric shovel fault diagnosis, but there are still some challenges and problems.

[0003] First, traditional neural network training methods mainly rely on gradient optimization algorithms such as gradient descent. These algorithms have limitations in dealing with complex non-convex objective functions, discrete optimization problems, or neural network structure designs involving non-differentiable operations. For example, the automatic design of neural network architectures and the global optimization of weights usually involve high-dimensional search spaces and local optimal traps, and gradient methods are difficult to effectively handle.

[0004] Second, existing neural network optimization methods based on evolutionary algorithms mainly focus on improving network performance and searching for the best network weights through global search. However, these methods often cannot generate lightweight neural networks because most evolutionary processes result in most elements in the weight matrix not being zero, leading to high network complexity and large resource consumption. Moreover, existing evolutionary algorithms for optimizing neural networks are less efficient and have a slow convergence rate when dealing with a large number of variables. This limits their application in industrial scenarios sensitive to computing resources. In addition, the parameter settings of evolutionary algorithms have a greater impact on the performance of the algorithms and need to be adjusted according to specific problems, increasing the complexity and difficulty of using the algorithms. Therefore, existing neural network-based fault detection methods still have some problems, such as low optimization efficiency, difficulty in generating lightweight networks, and large computing resource consumption. These problems limit the popularization and application of neural network-based fault detection methods in actual industrial applications. Summary of the Invention

[0005] Object of the Invention: Aiming at the above existing problems and deficiencies, the object of the present invention is to provide an electric shovel fault diagnosis method based on a sparse evolutionary algorithm and a neural network. This method is a large-scale sparse multi-objective evolutionary algorithm based on dynamic variable clustering (DVCEA), which can improve the efficiency and accuracy of electric shovel fault diagnosis, thereby more effectively protecting the equipment and providing more accurate and reliable decision-making support.

[0006] Technical solution: To achieve the above-mentioned invention objective, the present invention adopts the following technical solution: An electric shovel fault diagnosis method based on a sparse evolutionary algorithm and a neural network, comprising the following steps:

[0007] S1. Obtain electric shovel data, collect vibration signal data of vibration sensors on key components of the electric shovel, and label the state type for the vibration signal data according to fault characteristics. The state type includes different fault types or normal states, and is used as input data;

[0008] S2. Initialize a population P with N randomly generated solutions, obtain and establish an N×D single-hidden-layer feedforward neural network with D decision variables, and then construct a solution combination matrix X of the single-hidden-layer neural network;

[0009] S3. Based on the guiding vector gv of the solution population P i , generate 2N parent solutions from the solution population P through the mating selection strategy in SPEA2 to form an intermediate population P';

[0010] S4. Based on the guiding vector gv of the solution population P i , perform a descending order sorting and clustering grouping on the decision variables;

[0011] S5. Based on the parent solutions selected in step S3, combined with the decision variable clustering grouping in step S4, generate offspring solutions through a crossover mutation operator. During the process of generating offspring solutions, the decision variables in the same group cross and mutate simultaneously;

[0012] S6. After the iteration of the preset number FE max is completed in step S5, adopt the environmental selection strategy of SPEA2 to select a population of N solutions, and obtain an optimized neural network model with higher classification accuracy and lower sparsity.

[0013] Furthermore, the construction process of the solution combination matrix X of the single-hidden-layer neural network in step S2 is as follows:

[0014] S2.1. Obtain the neural network parameter combination data dec in the N×D single-hidden-layer feedforward neural network, dec = [dec1, dec2,..., dec i ,..., dec D , where dec i represents the decision variable of the i-th neural network, D represents the number of decision variables, and represents the j-th decision variable in the i-th neural network;

[0015] S2.2. Let the binary variable mask of the j-th decision variable in the i-th neural network be Indicate, initialization The rest Thus, the j-th decision variable gene in the i-th neural network is constructed Thus, the i-th neural network individual x is obtained i , Furthermore, the neural network solution combination matrix X is obtained, X = [x1, x2, …, x i , …, x D .

[0016] Furthermore, the formation process of the intermediate population P′ in step S3 is as follows:

[0017] S3.1, Calculate the guiding vector gv of the i-th neural network through the following formula i ,

[0018]

[0019] In the formula, gv i Represents the guiding vector of the i-th neural network, reflecting the proportion of each decision variable in the current population set to 0, R represents all neural networks to be optimized, |R| represents the number of all neural networks to be optimized, and mask i Represents the binary variable mask matrix of the parameters in the i-th neural network;

[0020] S3.2, Generate 2N parent solutions from the population P of solutions through the mating selection strategy in SPEA2 to form the intermediate population P′.

[0021] Furthermore, the process of descending sorting and clustering grouping of decision variables in step S4 is as follows:

[0022] S4.1, Calculate the population sparsity Sparsity and the decision variable grouping size GroupSize of the population P of solutions through the following formula respectively

[0023]

[0024] In the formula, |w i |0 represents the number of non-zero elements of the weight vector w i in the i-th neural network;

[0025] S4.2, Determine the size GroupSize of the grouping by calculating the magnitude of the guiding vector gv i of the neural network decision variables, sorting the decision variables, and then dividing the decision variables into Groups groups based on the value of Sparsity.

[0026] Further, the process of generating offspring solutions through the crossover mutation operator in step S5 includes:

[0027] S5.1, Initialize an empty offspring solution set O and determine the value of the total number of groups MaxGroup variable according to the number of input groups of the decision variable Groups;

[0028] S5.2, In each iteration, randomly select two neural network combination individuals p iter from the population P of the iter-th generation iter and q iter as parents, and select a random group groups from the decision variable grouping Groups;

[0029] Step 5.3, The bitwise XOR operation is applied to the binary masks of the parents to generate the indices index of different variables. Index is used to flip a group at a time during mutation crossover operations to accelerate the generation speed of the optimal solution;

[0030] S5.4, Delete the selected parent solutions p iter and q iter from P iter , and assign the binary vector p iter .mask of the parent P iter to the binary vector o iter of an offspring individual in the iter-th generation; iter .mask;

[0031] S5.5, Perform crossover mutation operations on the binary vector o iter .mask of the offspring individual with the same probability according to step S5.5.1 or step S5.5.2;

[0032] S5.5.1, With a probability of 50%, randomly select half of the decision variables from the intersection of index ∩ group, and set the binary values of the variables to 1. Otherwise, set their binary values to 0;

[0033] S5.5.2, With a probability of 50%, flip the binary values of the elements randomly selected from o iter .mask that are 0 to 1. Otherwise, flip the binary values of the elements that are 1 to 0;

[0034] S5.6, Perform simulated binary crossover and polynomial mutation operations on the neural network weight combination data p iter of the parent p iter and the weight combination data q iter of q iter to generate an offspring individual o iterWeight combination data o iter .dec; thus obtaining the offspring individuals o iter = o iter .mask × o iter .dec; thus generating the offspring population O of the iter-th generation iter for each offspring individual in it. Further, step S6 is iterated and environmental selection is as follows:

[0035] S6.1, assign the value obtained by adding the number of offspring |Q| to the previous round of FE' value to FE, completing the update of the evaluation iteration count FE, FE = FE' + |Q|, where FE' represents the previous round of FE value and |Q| represents the number of offspring;

[0036] S6.2, if FE < FEmax, return to step S3 to continue execution; otherwise, select N solutions from P ∪ Q using the environmental selection method of SPEA2 to obtain the final solution.

[0037] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0038] 1. Improve network sparsity: DVCEA can effectively generate a sparse neural network model, thereby reducing model complexity, reducing computational resource consumption, and improving the generalization ability of the model

[0039] 2. Improve fault detection accuracy: DVCEA can optimize the weights of the neural network, thereby improving the accuracy and robustness of fault detection and more accurately identifying the fault state of the electric shovel.

[0040] 3. Improve optimization efficiency: DVCEA adopts dynamic variable clustering and improved genetic operators, which can effectively improve the search efficiency of the algorithm and reduce the calculation time.

[0041] 4. Reduce computational resource consumption: DVCEA can generate a sparse neural network model, thereby reducing the computational resource consumption for model training and inference, making it more suitable for application in resource-constrained industrial scenarios.

[0042] 5. Easy to implement: The technical means adopted by DVCEA is simple and easy to implement, and can be conveniently applied to the existing electric shovel fault diagnosis system. Brief description of the drawings

[0043] Figure 1 is a schematic flow chart of the electric shovel fault diagnosis method based on the sparse evolutionary algorithm and neural network of the present invention. Detailed implementation manners

[0044] The present invention will be further illustrated below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, those skilled in the art's various equivalent modifications of the present invention all fall within the scope defined by the appended claims of this application.

[0045] A fault diagnosis method for an electric shovel based on a sparse evolutionary algorithm and a neural network disclosed by the present invention is an optimization method for neural network fault detection. It effectively guides the algorithm to search for a sparse solution space by introducing dynamic variable clustering and sparse guiding vectors, so as to find a neural network model with better performance and sparsity. DVCEA adopts a double-coding strategy similar to SparseEA to represent solutions. Each solution consists of a binary vector mask and a real-valued vector dec. Each variable in mask represents whether the corresponding decision variable (neural network weight) is zero, while each variable in dec specifies the real value of the corresponding decision variable. At the same time, a dynamic variable clustering method is introduced to group all decision variables according to their importance (reflected by the guiding vector gv). Variables with lower importance are more likely to be grouped together, and thus are more likely to be set to 0 simultaneously in subsequent crossover and mutation operations, accelerating the algorithm's convergence to a sparse solution. The guiding vector gv reflects the proportion of each decision variable in the current population being set to 0, that is, the sparsity degree of this variable. The larger the value of gv, the less important the variable is, and it is more inclined to be set to 0 in the population. DVCEA uses the pairing selection and environmental selection strategies of the SPEA2 algorithm, and combines specific crossover and mutation operators to generate offspring solutions. The crossover operator randomly selects groups of parent solutions and performs crossover operations on the variables within the groups; the mutation operator explores the solution space by randomly changing the elements in mask. Specifically, it is carried out according to the following steps:

[0046] S1. Obtain electric shovel data;

[0047] S1.1. By installing vibration sensors on the key components of the electric shovel (such as bearings and gears), collect one-dimensional vibration signal data of the system. Each data represents the amplitude, and different data represent the states of the system at different times. Analyze the collected data and label the data as different fault types or normal states according to the fault characteristics. Directly use the collected data as the input of the neural network.

[0048] S1.2. Initialize a population P with randomly generated N solutions; define the current evaluation number FE as the size of the initial population |P|; define the maximum evaluation number FE max ; define the decision variable X, and define the number of decision variables D;

[0049] S1.3. Define the specific structure of the fully connected neural network:

[0050] The established neural network is a single-hidden-layer feedforward neural network, consisting of an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is equal to the dimension of the data features, corresponding to the number of features for each sample. The hidden layer contains 20 neurons and uses the hyperbolic tangent function as the activation function to map the input to the range (-1, 1). The number of neurons in the output layer is equal to the number of classes in the classification task, denoted as C, and a linear activation function is used to generate classification scores. The loss function of the network is the mean squared error.

[0051] S1.4. Define each objective function:

[0052] S1.4.1. Define the objective function f1(w) for minimizing the classification error using Equation (1):

[0053]

[0054] In Equation (1), represents the predicted label of the i-th sample; y i is the true label; δ(.,.) is the indicator function, which is equal to 1 if the predicted label matches the true label, and 0 otherwise;

[0055] S1.4.2. Define the objective function f2(w) for minimizing the sparsity of the neural network using Equation (2):

[0056] f2(w) = |w|0 (2)

[0057] In Equation (2), w is the weight parameter of the network; |w|0 represents the number of non-zero weights; introducing the sparsity preference can optimize the network performance and significantly reduce the number of non-zero weights. This can effectively discover high-performance neural networks with fewer parameters, making it particularly suitable for industrial scenarios sensitive to computing resources;

[0058] S1.4.3. Obtain the aggregated true objective function f(x) using Equation (3):

[0059] Minimizef(x) = (f1(x), f2(x)) (3)

[0060] S2. Establish a neural network;

[0061] Obtain the N×D neural network parameter combination data dec = [dec1, dec2,..., dec i ,..., dec D , where dec i represents the weight data of the i-th neural network, and represents the j-th weight data in the i-th neural network; let The weight data of the j-th in the i-th neural network binary variable indicating that the weight data of the j-th in the i-th neural network is not selected indicating that the weight data of the j-th in the i-th neural network is selected Initialize The rest Thus, construct the gene of the weight data of the j-th in the i-th neural network And obtain the individual of the i-th neural network Furthermore, obtain the neural network solution combination matrix X = [x1, x2,..., x i ,..., x D ;

[0062] S3. Generate the guiding vector gv i And generate the parental generation;

[0063] S3.1. Calculate the guiding vector gv using Equation (4) i :

[0064]

[0065] gv i By calculating the proportion of each decision variable in the current population being set to 0, it reflects the sparsity degree of the variable in the population. The larger the value, the less important the variable is and it is more likely to be set to 0 in the population;

[0066] S3.2. Use the mating selection strategy from SPEA2 to select 2N parental solutions from P to form the intermediate population P'

[0067] S4. Dynamically cluster variables and generate grouped decision variables through the guiding vector;

[0068] S4.1. Calculate the population sparsity Sparsity using Equation (5):

[0069]

[0070] S4.2. Calculate the grouped decision variable size GroupSize using Equation (6):

[0071] GroupSize = D × Sparsity (6)

[0072] S4.3. According to gv iSort the decision variables in descending order. The algorithm sorts all decision variables according to their sparsity, and then groups the sorted variables. Variables with lower sparsity are more likely to be grouped together, making it easier to be simultaneously set to 0 in subsequent crossover and mutation operations, accelerating the algorithm's convergence to a sparse solution;

[0073] S4.4. Divide the sorted decision variables into Groups groups, where the number of decision variables in each group is GroupSize. By using the gv i to guide variable clustering, divide the binary variables into multiple equivalent groups, allowing variables within each group to be simultaneously flipped to zero, which is beneficial for approaching the sparse optimal solution. The algorithm can more effectively search the sparse solution space, thereby finding a neural network model with both good performance and sparsity;

[0074] S5. Generate offspring using the crossover and mutation operators;

[0075] S5.1. Initialize an empty offspring solution set O and determine the value of the total number of groups MaxGroup variable according to the number of input groups of the decision variables Groups;

[0076] S5.2. In each iteration, randomly select two neural network combination individuals P iter and q iter from the population P of the iter-th generation iter as parents, and select a random group group from the decision variable groups Groups;

[0077] S5.3. Apply the bitwise exclusive OR operation xor(p.mask, q.mask) to the binary masks of the parents to generate the indices index of different variables. index is used to flip an entire group at once during the mutation and crossover operations, accelerating the generation speed of the optimal solution; S5.4. Delete the selected parent solutions p iter and q iter from P iter , and assign the binary vector p iter .mask of the parent p iter to the binary vector o iter of an offspring individual in the iter-th generation; iter .mask;

[0078] S5.5. Perform crossover and mutation operations on the binary vector o iter .mask of the offspring individual with the same probability according to Step 5.5.1 or Step 5.5.2;

[0079] S5.5.1. Randomly select half of the decision variables from the intersection of index ∩ group with a probability of 50%, set the binary values of the variables to 1, otherwise, set their binary values to 0;

[0080] S5.5.2. With a probability of 50%, for the elements randomly selected from o iter .mask whose binary values are 0, flip their binary values to 1, otherwise, flip the elements with binary value 1 to 0 for their binary values;

[0081] Step 5.6. Perform simulated binary crossover and polynomial mutation operations on the neural network weight combination data p iter of the parent p iter .dec and the weight combination data q iter of q iter .dec to generate the weight combination data o iter of the offspring individual o iter .dec; thus obtaining the offspring individual o iter = o iter .mask × o iter .dec; thus generating each offspring individual in the offspring population O iter in the iter-th generation;

[0082] S6. Iteration and environmental selection

[0083] S6.1. Update the evaluation times, FE = FE + |Q|

[0084] S6.2. If FE < FE max , return to step three to continue execution; otherwise, use the environmental selection method of SPEA2 to select N solutions from P ∪ Q to obtain the final solution.

[0085] To achieve the above object of the invention, the present invention adopts the following technical solutions:

[0086] 1. Dynamic variable clustering: Divide all decision variables (i.e., the weight matrix and bias vector of the neural network) into multiple groups of equal size, thus significantly reducing the dimension of the decision space. This grouping strategy can effectively guide the generation of offspring solutions and improve the optimization efficiency.

[0087] 2. Dual coding mechanism: Adopt a dual coding mechanism to encode the decision variables into binary vectors and real-value vectors. The binary vector represents whether the weight is zero, and the real-value vector represents the specific value of the non-zero weight. This coding method can effectively achieve network sparsification and improve the search efficiency of the algorithm.

[0088] 3. Improved genetic operators: Improved crossover and mutation operators are designed to enable the offspring solutions to better inherit the excellent characteristics of the parent solutions and effectively explore the search space.

[0089] The above are only partial embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for shovel fault diagnosis based on sparse evolutionary algorithm and neural network, characterized in that: It includes the following steps: S1. Obtain the data of the electric shovel, collect the vibration signal data of the vibration sensors on the key components of the electric shovel, and label the state type for the vibration signal data according to the fault characteristics. The state type includes different fault types or normal states, and is used as input data. S2. Initialize the population P with N randomly generated solutions, obtain and establish a single-hidden-layer feedforward neural network with D decision variables, and then construct the solution combination matrix X of the single-hidden-layer neural network. S3, the guidance vector gv of the group P based on the solution i , 2N parent solutions are generated from the solution population P through the mating selection strategy in SPEA2 to form an intermediate population P'; S4, the guidance vector gv of the group P based on the solution i , sort the decision variables in descending order and group them into clusters; S5. Based on the parent solutions selected in step S3, combined with the clustering grouping of the decision variables in step S4, generate offspring solutions through the crossover and mutation operator. During the generation of offspring solutions, the decision variables in the same group cross and mutate simultaneously. S6, step S5 is completed a preset number of times FE max After iterations, the environment selection strategy of SPEA2 is used to select a population of N solutions, and an optimized neural network model with higher classification accuracy and lower sparsity is obtained.

2. The electric shovel fault diagnosis method based on sparse evolutionary algorithm and neural network according to claim 1 is characterized in that: The construction process of the solution combination matrix X of the single-hidden-layer neural network in step S2 is as follows: S2.1, obtain the neural network parameter combination data dec,dec=[dec1,dec2,…,dec i ,…,dec D ], where dec i represents the decision variable of the ith neural network, D represents the number of decision variables, and represents the jth decision variable in the i-th neural network, S2.2, let the jth decision variable in the i-th neural network The binary variable mask is Indicates that initialization the remaining Thus constructing the j-th decision variable gene in the i-th neural network Thus, we get the i-th neural network individual x i , Then we get the neural network solution combination matrix X, X = [x1, x2, ..., x i ,…,x D ].

3. The electric shovel fault diagnosis method based on sparse evolutionary algorithm and neural network according to claim 1 is characterized in that: The formation process of the intermediate population P' in step S3 is as follows: S3.1, calculate the guidance vector gv of the i-th neural network by the following formula i , In the formula, gv i represents the guidance vector of the ith neural network, reflecting the proportion of each decision variable in the current population set to 0, R represents all neural networks to be optimized, |R| represents the number of all neural networks to be optimized, mask i A binary variable mask matrix representing the parameters in the i-th neural network; S3.

2. Generate 2N parent solutions from the population P of solutions through the mating selection strategy in SPEA2 to form the intermediate population P'.

4. The electric shovel fault diagnosis method based on sparse evolutionary algorithm and neural network according to claim 1 is characterized in that: The process of descending sorting and clustering grouping of the decision variables in step S4 is as follows: S4.

1. Calculate the population sparsity Sparsity and the decision variable grouping size GroupSize of the population P of solutions respectively through the following formula: In the formula, |w i |0 represents the weight vector w in the i-th neural network i The number of non-zero elements of ; S4.2, by calculating the guidance vector gv of the neural network decision variable i The size of the decision variables is calculated and sorted. Then, the size of the group GroupSize is determined based on the value of Sparsity. The decision variables are divided into Groups groups according to the group size.

5. The electric shovel fault diagnosis method based on sparse evolutionary algorithm and neural network according to claim 1 is characterized in that: The process of generating offspring solutions through the crossover and mutation operator in step S5 includes: S5.

1. Initialize the empty offspring solution set O and determine the value of the total number of groups MaxGroup variable according to the number of input groups of the decision variables Groups. S5.2, in each iteration, from the iter generation population P iter Randomly select two neural network combination individuals p iter and q iter as the parent and select a random group groups from the decision variable groups Groups; Step 5.

3. Apply the exclusive OR operation to the binary masks of the parents to generate the indexes index of different variables. The index is used to flip one group at a time during the mutation and crossover operations to accelerate the generation speed of the optimal solution. S5.4, the selected parent solution p iter and q iter From P iter Delete it and change the parent P iter The binary vector p iter .mask is assigned to a descendant individual o of the iter generation iter The binary vector o iter .mask; S5.5, with the same probability, perform step S5.5.1 or step S5.5.2 on the binary vector o of the offspring individual iter .mask performs crossover mutation operation; S5.5.

1. With a probability of 50%, randomly select half of the decision variables from the intersection of index ∩ group, and set the binary values of the variables to 1. Otherwise, set their binary values to 0. S5.5.2, with a probability of 50%, will be iter .Randomly select elements with binary value 0 in mask and flip their binary value to 1, otherwise flip the elements with binary value 1 to their binary value 0; S5.6, for the parent p iter The neural network weight combination data p iter .dec and q iter The weighted combination data q iter .dec performs simulated binary crossover and polynomial mutation operations to generate offspring individuals o iter Weighted combination data o iter .dec; thus obtaining the offspring individual o iter =o iter .mask×o iter .dec; thus generating the iter generation offspring population O iter Each offspring individual.

6. The electric shovel fault diagnosis method based on sparse evolutionary algorithm and neural network according to claim 1 is characterized in that: The iteration and environmental selection in step S6 are as follows: S6.

1. Assign the value obtained by adding the number of offspring |Q| to the previous round of FE' value to FE to complete the update of the evaluation iteration times FE, FE = FE' + |Q|, where FE' represents the previous round of FE value and |Q| represents the number of offspring. S6.

2. If FE < FEmax, return to step S3 to continue execution; otherwise, select N solutions from P ∪ Q using the environmental selection method of SPEA2 to obtain the final solution.

Images