An aviation manufacturing equipment big data fault diagnosis method

Abstract

An aviation manufacturing equipment big data fault diagnosis method, steps include: to aerospace product manufacturing data such as vibration, temperature, noise, pose, electromagnetic state signal, for the characteristics of large amount of data and low value density of such aviation equipment detection data, the data is objectively selected by applying entropy weight method; BP neural network is applied to training to realize fault diagnosis; the present application can better play the advantages of BP neural network, and make up for its defects, the data is objectively selected by applying entropy weight method in the processing of diagnosis data, the optimal weight and threshold are searched by ABC algorithm, and the BP neural network is optimized; the position search method of artificial bee colony algorithm and the selection of honey source method of following bee are improved, and the optimization speed and effect are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis of aerospace product manufacturing equipment, and particularly relates to a large data fault diagnosis method for aerospace manufacturing equipment based on an ABC-BP neural network. BACKGROUND

[0002] With the gradual development of science and technology, the demand for aerospace products is increasing, and the product technology is becoming more and more advanced. At the same time, the efficiency and reliability of aerospace production equipment are also increasingly high. In the production system with high automation and intelligence, it is very important to automatically identify and predict the failure of production equipment. In the field of fault diagnosis, there are many mature solutions in China, which are widely used in mechanical equipment diagnosis. However, the diagnosis of aerospace production equipment has problems such as difficulty in knowledge acquisition, complex equipment, and low data quality density, so the fault diagnosis field in China is still not mature. It is of great significance to apply existing fault diagnosis technology to this field to realize higher automation and intelligent integration of aerospace production equipment.

[0003] Fault diagnosis is the process of finding the fault of equipment or system, understanding the running state of the machine, and early discovering the fault and making trend prediction. The fault diagnosis method based on expert system is a relatively mature method at present. The knowledge system is constructed from the experience knowledge of field experts, which is also the reason for its natural deficiency. It is difficult to collect all the problems, and the diagnosis process is easy to fall into a dilemma. The fault diagnosis method based on support vector machine can linearly transform nonlinear problems. Since SVM solves support vectors by means of quadratic programming, it is difficult to realize in the case of large samples such as aerospace equipment, and the efficiency is low. The fuzzy fault diagnosis method mainly uses fuzzy theory to carry out fuzzy modeling and fuzzy clustering processing on data information. It does not need to establish an accurate mathematical model, but its defect is that it depends on human experience to a great extent and does not have learning ability. The fault diagnosis method based on artificial network does not need to summarize and digest the knowledge of field experts, but uses the examples or cases of field experts to solve problems for training. It has high universality, large-scale parallel processing ability, self-learning ability and self-adaptive ability, and plays a specific advantage role in fault diagnosis.

[0004] Backpropagation (BP) neural networks, a widely used and technologically mature type of artificial neural network, have shown good performance in many fields. A BP neural network is a multi-layer feedforward neural network that backpropagates errors during model training. Even without knowing the mapping relationship between input and output, it can obtain a near-desired output through self-training, exhibiting excellent pattern classification and multi-dimensional function mapping capabilities. A BP neural network has a three-layer structure: one input layer, one output layer, and several hidden layers. The relationship between input and output can be altered during computation. In finding the optimal solution, it employs gradient descent, which minimizes the mean squared error between the actual and desired output values. However, this method is prone to stalling at local extrema when dealing with complex multi-dimensional grid spaces, failing to train a truly optimal solution globally. To address this issue, methods for global optimization can be applied to BP neural networks, such as particle swarm optimization, genetic algorithms, and artificial bee colony optimization.

[0005] The Artificial Bee Colony Algorithm (APA) is derived from research on bee colony foraging behavior. It highlights the concept of swarm intelligence and has unique advantages in solving global optimization problems. Before optimization, the algorithm does not need specific problem information; during optimization, it only judges the quality of the problem. It continuously updates the optimal solution and eliminates nectar sources during local optimization, ultimately revealing the global optimal solution. This process converges faster than other similar algorithms. The algorithm employs three types of bees with different task assignments: the leader bee has a one-to-one correspondence with nectar sources, constantly searching for new solutions and remembering and updating the optimal solution found during the search process, transmitting the solution information to other bees; follower bees, upon receiving information from the leader bee, search for nectar sources, playing a cooperative role and effectively improving the convergence speed; scout bees, when trapped in a local optimum, break out of the current loop and search for new nectar sources near the hive, greatly increasing the probability of finding the global optimal solution. The ABC algorithm adopts the above principles and methods, and has the characteristics of simple operation, easy parameter tuning, good optimization effect and good robustness. It has a strong ability to find the global optimal solution and a fast convergence speed, which can make up for the defect of BP neural network that is prone to getting trapped in local extrema.

[0006] Combining the ABC algorithm with a BP neural network for fault diagnosis can compensate for the shortcomings of the BP network. For example, patent number CN201910525924.6 is titled "A Deep Computational Model for Fault Diagnosis of Air Path in Aero-engines." However, they still have technical defects: the selection of training samples without screening will lead to excessively high data dimensionality, and manual screening inevitably involves subjectivity. At the same time, the ABC algorithm itself has certain shortcomings. When searching, it starts from the neighborhood of the nectar source, which limits its ability to find new nectar sources. When selecting nectar sources, it adopts a simple roulette wheel method. If a better solution with a high fitness value is selected, the search stops. Therefore, how to propose corresponding improvement schemes to address these two problems has always been a new approach to fault diagnosis that those skilled in the art are exploring. Summary of the Invention

[0007] To address the problem of fault diagnosis of manufacturing equipment based on big data in aerospace manufacturing processes, this invention aims to provide a fault diagnosis method for aerospace manufacturing equipment based on big data. This method combines the ABC algorithm and the BP neural network for fault diagnosis, better leveraging the advantages of the BP neural network while compensating for its shortcomings. In the processing of diagnostic data, the entropy weight method is applied to objectively select data, and the ABC algorithm is used to search for optimal weights and thresholds. Finally, model training is performed, ensuring training effectiveness while improving training speed.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A big data-based fault diagnosis method for aerospace manufacturing equipment includes the following steps:

[0010] Step 1: Collect and preprocess aerospace product manufacturing data, including vibration, temperature, noise, pose, and electromagnetic state signals, and extract features. Compare the importance of the extracted data to obtain the input data directly used for model training. Use the entropy weight method to assign weights to fixed indicators, and comprehensively evaluate the feature vectors through the weight values. Select the top-ranked feature vectors to participate in the subsequent network training model.

[0011] The method of using entropy weighting to assign weights to fixed indicators and then comprehensively evaluating the feature vectors using these weights specifically includes:

[0012] (1.1) To prevent negative values ​​in the input matrix, the original eigenvectors are normalized to obtain X = (x′). ij ) m×n The formula is:

[0013]

[0014] (1.2) Calculate the information entropy e for each indicator j :

[0015]

[0016] in, K is a constant.

[0017] (1.3) Calculate the weight w of each eigenvector. j :

[0018]

[0019] The feature vectors are comprehensively evaluated by weight values, and the top-ranked feature vectors are selected for subsequent network training.

[0020] Step 2: Use the Artificial Bee Colony ABC algorithm to determine the weight matrix w and bias matrix b parameters needed in the network training model to initialize the neural network. Specific implementation steps include:

[0021] (2.1) Initialize parameters, set parameters, including population size, maximum number of iterations, control parameters, set nectar sources, and the location of each nectar source corresponds to a set of weights and thresholds of the BP network;

[0022] (2.2) Assess the fitness of the initial nectar source quality;

[0023] (2.3) After setting the cycle conditions, the population enters the cycle phase and begins to cycle;

[0024] (2.4) Guide the bees to search the neighborhood of the nectar source to generate new nectar sources, calculate the fitness of the new nectar sources, and select between the old and new nectar sources according to the greed principle;

[0025] (2.5) Calculate the probability of each nectar source being selected. After the follower bee selects a nectar source based on the probability, it searches for new nectar sources, calculates the fitness of the new nectar sources, and selects between the old and new nectar sources according to the greedy principle.

[0026] (2.6) Determine if there are any nectar sources that need to be abandoned, and have the scout bees handle the nectar sources that have failed to evolve to generate new nectar sources;

[0027] (2.7) Continue until the maximum number of iterations is reached, which satisfies the loop termination condition, and the loop ends, thus finding the optimal honey source.

[0028] In step (2.4), when the lead bee and follower bees search for nectar sources, the search starts from the neighborhood of the nectar source. The parameter of the currently found optimal solution is added to the search position update formula, that is:

[0029]

[0030] in, y is a random number between (0, 1) jbest Let j be the j-th variable representing the optimal nectar source for the population.

[0031] In step (2.5), the probability calculation formula used when calculating the probability of a nectar source being selected is:

[0032]

[0033] Where max(f) i t i () represents the maximum value of the nectar source's adaptability.

[0034] Step 3: Use the optimal honey source found by the ABC algorithm as the initial weight matrix w and bias matrix b parameters for the BP neural network, and then train the BP neural network. Specific implementation steps include:

[0035] (3.1) Input the selected feature vector after weight evaluation into the network;

[0036] (3.2) Calculate the output of hidden layer neurons and calculate the output of output layer neurons by weighted summation; the input of each hidden layer neuron is calculated by the following formula:

[0037]

[0038] Among them, w ih θ is the weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer. h It is the threshold of the h-th neuron in the hidden layer;

[0039] The hidden layer performs linear and nonlinear transformations on the input layer data, processes the output of the hidden layer through activation functions, and calculates the output of the output layer neurons through weighted summation. The activation function chosen is sigmoid.

[0040]

[0041] (3.3) Calculate the output layer error and enter the error backpropagation process. Correct the w and b of the BP network by gradient descent.

[0042] (3.4) Determine whether the global error of the model is less than the ideal error. If it is less, proceed to the next step. Otherwise, continue updating the iteration count from step (2) and calculate the global error under the new weights and thresholds until the ideal error requirement is met.

[0043] (3.5) After the BP neural network training is completed, the model testing process begins to evaluate the training results.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. This invention applies the fault diagnosis method based on ABC-BP neural network to the field of big data fault diagnosis of aerospace manufacturing equipment, explores the fault diagnosis in this field, and uses artificial bee colony algorithm to solve the initial weights and thresholds required by BP neural network, effectively avoiding the BP network from getting trapped in local optima.

[0046] 2. For complex aerospace manufacturing equipment data, the importance of the data is selected by using the entropy weight method to objectively select the feature vectors for training. This effectively reduces the dimensionality of the data involved in training and improves efficiency without affecting the training effect.

[0047] 3. The artificial bee colony algorithm does not use traditional formulas in terms of location search and probability calculation. Instead, it adopts new calculation methods, which improves the optimization speed and effectiveness. Attached Figure Description

[0048] Figure 1 This is a flowchart of a big data fault diagnosis method for aerospace manufacturing equipment according to an embodiment of the present invention.

[0049] Figure 2 This is a schematic diagram of the artificial bee colony ABC algorithm training model of the present invention.

[0050] Figure 3 This is a schematic diagram of the BP neural network training model of the present invention.

[0051] Figure 4 This is a flowchart of the ABC-BP neural network training model of the present invention. Detailed Implementation

[0052] The technical solutions in this application will now be clearly and completely described in conjunction with the accompanying drawings.

[0053] like Figure 1 The diagram shows the overall flow of this fault diagnosis method, and the specific steps include:

[0054] Step 1: Collect and preprocess aerospace product manufacturing data, including vibration, temperature, noise, pose, and electromagnetic state signals, and extract features. Compare the importance of the extracted data to obtain the input data directly used for model training. Use the entropy weight method to assign weights to fixed indicators, and comprehensively evaluate the feature vectors through the weight values. Select the top-ranked feature vectors to participate in subsequent network training.

[0055] The method of using entropy weighting to assign weights to fixed indicators and then comprehensively evaluating the feature vectors using these weights specifically includes:

[0056] (1.1) To prevent negative values ​​in the input matrix, the original eigenvectors are normalized to obtain X = (x′). ij ) m×n The formula is:

[0057]

[0058] (1.2) Calculate the information entropy e for each indicator j :

[0059]

[0060] in, K is a constant.

[0061] (1.3) Calculate the weight w of each eigenvector. j :

[0062]

[0063] The feature vectors are comprehensively evaluated by weight values, and the top-ranked feature vectors are selected for subsequent network training.

[0064] like Figure 2 The diagram illustrates the process of finding the optimal solution using the Artificial Bee Colony (APC) algorithm. The APC algorithm is used to determine the weight matrix w and bias matrix b parameters required for model training, which are then used to initialize the neural network. The specific implementation steps include:

[0065] (2.1) Initialize the bee colony. Use a random method to generate T initial solutions in the search space. The formula is as follows, which is used to correspond to the corresponding number of initial nectar sources. Set the population size, maximum number of iterations, control parameters limit, etc. The position of each nectar source corresponds to a set of weights and thresholds of the BP network.

[0066] x ij =L min +rand(0,1)(L max -L min )

[0067] Where, x ij ∈(L min ,L max ), L min and L max Let represent the minimum and maximum values ​​of the search space, respectively, where j∈(1,2,…,D).

[0068] (2.2) Assess the fitness of the initial nectar source quality;

[0069] (2.3) After setting the cycling conditions, the population enters the cycling phase and begins cycling. The fitness function is calculated as follows:

[0070]

[0071] Among them, f i t i Let f be the fitness of the i-th nectar source. i The objective function for optimizing artificial bee colonies.

[0072] (2.4) The lead bee conducts a neighborhood search to generate new nectar sources and calculates the fitness of the new nectar sources. When the lead bee and follower bees search for nectar sources, the search starts from the neighborhood of the nectar source, which limits the search ability to a certain extent. This invention considers adding the parameter of the currently found optimal solution to the search position update formula to improve the optimization speed.

[0073]

[0074] in, y is a random number between (0, 1) jbest Let j be the j-th variable representing the optimal nectar source for the population.

[0075] Using the greedy principle in x′ ij and x ij Choose between the following: If the new honey source x′ ij Its fitness value is higher than that of the original nectar source x ij Then x′ ij Replace x ij Otherwise, keep x ij Keep it unchanged, and record x. ij The update failure count is incremented by 1.

[0076] (2.5) Calculate the probability of each nectar source being selected. After the follower bees select a nectar source based on the probability, they search for new nectar sources and calculate the fitness of the new nectar sources. Based on the greedy principle, they select between old and new nectar sources. To address the problem of some nectar sources having excessively high fitness values, during the optimization process, some optimal solutions have very high fitness values, which may cause the hired bees to prematurely stop searching for new nectar sources. To solve this problem and maintain population diversity, the probability calculation uses the following formula:

[0077]

[0078] Where max(f) i t i () represents the maximum value of the nectar source's adaptability.

[0079] (2.6) Scout bees handle failed nectar sources: If a nectar source does not improve after a specified limit of iterations, the corresponding mercenary bee becomes a scout bee. The scout bee then handles the failed nectar source and continues searching for new nectar sources. Its fitness is calculated, and then the process returns to the mercenary bee stage to begin the next iteration.

[0080] (2.7) Continue until the maximum number of iterations is reached, which satisfies the loop termination condition, and the loop ends, thus finding the optimal honey source.

[0081] Step 3, as follows Figure 3 The diagram shows the flowchart of the BP neural network algorithm. The topology of the neural network is determined, and the optimal honey source found by the ABC algorithm is used as the initial weight matrix w and bias matrix b parameters of the BP neural network. The specific implementation steps during BP neural network training include:

[0082] (3.1) Input the selected feature vector after weight evaluation into the network. The input vector is X = (x1, x2, ... x...). n ), set the error function, and give the calculation accuracy value and the maximum number of learning iterations.

[0083] (3.2) During the forward propagation process, data is transmitted from the input layer to the hidden layer. The input to each neuron in the hidden layer is calculated using the following formula:

[0084]

[0085] Among them, w ih θ is the weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer. h It is the threshold of the h-th neuron in the hidden layer.

[0086] The hidden layer performs linear and nonlinear transformations on the input layer data, processes the output of the hidden layer through activation functions, and calculates the output of the output layer neurons through weighted summation. The activation function chosen is sigmoid.

[0087]

[0088] (3.3) Calculate the output layer error and enter the error backpropagation process. Correct the w and b of the BP network by gradient descent, that is, update the weights by the degree of influence of each weight on the error.

[0089] (3.5) Determine if the model's global error is less than the ideal error. If it is, proceed to the next step; otherwise, update the iteration count and continue training, calculating the global error under the new weights and threshold, until the ideal error requirement is met and training ends. If the number of learning iterations reaches the set maximum during this process, training should also end. The BP neural network training ends, and the model testing process begins to evaluate the training results.

[0090] Figure 4 This is a flowchart of the ABC-BP neural network training model of the present invention. The big data fault diagnosis method provided by the present invention can be applied to the fault prediction and diagnosis of aerospace equipment. In view of the characteristics of large amount of aerospace equipment detection data and low value density, the entropy weight method is applied to objectively select data and incorporate it into the training model, which effectively reduces the dimensionality of the data involved in training. In order to improve the local optimum problem that is prone to occur in BP neural network training, an artificial bee colony algorithm is added to solve the optimal initial weights and thresholds. The position search method and the follower bee selection of nectar source method of the artificial bee colony algorithm are improved to avoid the abnormal individual problem that occurs in the roulette wheel strategy, maintain the diversity of the population, and improve the optimization speed and effect.

Claims

1. A big data-based fault diagnosis method for aerospace manufacturing equipment, characterized in that, The steps are as follows: Step 1: Collect and preprocess aerospace product manufacturing data, including vibration, temperature, noise, pose, and electromagnetic state signals, and extract features. Compare the importance of the extracted data to obtain the input data directly used for model training. Use the entropy weight method to assign weights to fixed indicators, and comprehensively evaluate the feature vectors through the weight values. Select the top-ranked feature vectors to participate in the subsequent network training model. Step 2: Use the Artificial Bee Colony ABC algorithm to determine the weight matrix w and bias matrix b parameters needed in the network training model, in order to initialize the neural network. Step 3: Use the optimal honey source found by the ABC algorithm as the initial weight matrix w and bias matrix b parameters of the BP neural network, and train the BP neural network. Step two includes the following specific implementation steps: (2.1) Initialize parameters, set parameters, including population size, maximum number of iterations, control parameters, set nectar sources, and the location of each nectar source corresponds to a set of weights and thresholds of the BP network; (2.2) Assess the fitness of the initial nectar source quality; (2.3) After setting the cycle conditions, the population enters the cycle phase and begins to cycle; (2.4) Guide the bees to search the neighborhood of the nectar source to generate new nectar sources, calculate the fitness of the new nectar sources, and select between the old and new nectar sources according to the greed principle; (2.5) Calculate the probability of each nectar source being selected. After the follower bee selects a nectar source based on the probability, it searches for new nectar sources, calculates the fitness of the new nectar sources, and selects between the old and new nectar sources according to the greedy principle. (2.6) Determine if there are any nectar sources that need to be abandoned, and have the scout bees handle the nectar sources that have failed to evolve to generate new nectar sources; (2.7) Continue until the maximum number of iterations is reached, which satisfies the loop termination condition, and the loop ends, thus finding the optimal honey source. In step (2.4), when the lead bee and follower bees search for nectar sources, the search starts from the neighborhood of the nectar source, and the currently found optimal solution is added to the search position update formula, that is: in, y is a random number between (0, 1) jbest Let j be the j-th variable representing the optimal nectar source for the population.

2. The method for big data fault diagnosis of aerospace manufacturing equipment according to claim 1, characterized in that, The method of using entropy weighting to assign weights to fixed indicators and then comprehensively evaluating the feature vectors using these weights specifically includes: (1.1) To prevent negative values ​​in the input matrix, the original eigenvectors are normalized to obtain X = (x' ij ) m×n The formula is: (1.2) Calculate the information entropy e for each indicator j : in, K is a constant. (1.3) Calculate the weight w of each eigenvector. j : The feature vectors are comprehensively evaluated by weight values, and the top-ranked feature vectors are selected for subsequent network training.

3. The method for big data fault diagnosis of aerospace manufacturing equipment according to claim 1, characterized in that, In step (2.5), the probability calculation formula used when calculating the probability of a nectar source being selected is: Where max(f) i t i () represents the maximum value of the nectar source's adaptability.

4. The big data fault diagnosis method for aerospace manufacturing equipment according to claim 1, characterized in that, The specific implementation steps of step three include: (3.1) Input the selected feature vector after weight evaluation into the network; (3.2) Calculate the output of hidden layer neurons and calculate the output of output layer neurons by weighted summation; the input of each hidden layer neuron is calculated by the following formula: Among them, w ih θ is the weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer. h It is the threshold of the h-th neuron in the hidden layer; The hidden layer performs linear and nonlinear transformations on the input layer data, processes the output of the hidden layer through activation functions, and calculates the output of the output layer neurons through weighted summation. The activation function chosen is sigmoid. (3.3) Calculate the output layer error and enter the error backpropagation process. Correct the w and b of the BP network by gradient descent. (3.4) Determine whether the global error of the model is less than the ideal error. If it is less, proceed to the next step. Otherwise, continue updating the iteration count from step (2) and calculate the global error under the new weights and thresholds until the ideal error requirement is met. (3.5) After the BP neural network training is completed, the model testing process begins to evaluate the training results.

Images