Full life cycle adaptive fault diagnosis methods, systems, equipment and media for major aerospace product manufacturing equipment

Abstract

The invention relates to a full-life cycle adaptive fault diagnosis method, system, equipment and medium for aerospace major product manufacturing equipment. A large amount of labeled fault data of aerospace major product manufacturing equipment is used as a training set for training to preliminarily determine the model parameters of the fault diagnosis system based on deep learning. Subsequently, a time series data anomaly detection system is established. The abnormal data filtered out by the time series data anomaly detection system is used to determine what kind of fault the abnormal data is and update the deep neural network parameters for the data. The system, equipment and medium can perform full-life cycle adaptive fault diagnosis on aerospace major product manufacturing equipment based on the fault diagnosis method, which greatly prevents the failure of the fault diagnosis model caused by long-term use of production equipment, and improves the accuracy and versatility of the fault diagnosis system.

Description

Technical Field

[0001] The present invention relates to the field of product fault detection technology, and in particular to a full life cycle adaptive fault diagnosis method, system, equipment and medium for manufacturing equipment of major aerospace products. Background Art

[0002] During the production of major aerospace products, such as aircraft, the reliability of manufacturing equipment must be guaranteed. Any minor malfunction could potentially lead to safety and quality issues. To ensure the safety and high quality of major aerospace products and reduce maintenance costs, it is crucial to accurately identify potential equipment failures.

[0003] Existing fault diagnosis methods generally do not rely on expert knowledge. Instead, they build neural network models and train them with a large amount of labeled sensor data to establish a correlation between sensor data and equipment fault status.

[0004] However, as manufacturing equipment operates, its internal joints, fit, lubrication, and wear conditions constantly change, altering the correlation between sensor data and fault conditions, and the aforementioned models gradually become ineffective. A universal and robust method for diagnosing major aerospace product faults is lacking.

[0005] Patent application CN107505126B discloses a method for evaluating the flight test performance of an airborne product. The method first obtains the number of flight test samples N, the number of diagnostic failures c, and the test performance index requirement RL; a confidence level C is given; then, the minimum flight test sample size n is calculated according to a formula. If the number of flight test samples N is greater than or equal to the minimum flight test sample size n, the flight test sample number N is substituted for n in the formula, and RL is calculated in reverse order as the airborne product test performance index evaluation value. Otherwise, the airborne product test performance index evaluation value r is evaluated using a point estimation method. This method provides a reasonable and feasible sample size determination method and confidence interval index evaluation method for field test performance evaluation, enabling a more realistic and objective assessment of the flight test performance level of airborne products. However, its accuracy and versatility still need to be further improved. Summary of the Invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a full life cycle adaptive fault diagnosis method, system, equipment and medium for aerospace major product manufacturing equipment; after preliminarily training a fault diagnosis model on labeled data, the abnormal data diagnosis model is used to detect abnormal data of sensors in the equipment during the operation of the major aerospace production and manufacturing equipment, and the parameters of the fault diagnosis model are continuously updated using the abnormal data as a new training set, so that the fault diagnosis model proposed by the present invention can have good fault diagnosis effect throughout the life cycle of the equipment.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment specifically includes the following steps:

[0009] Step 1. Preliminary establishment of a deep learning-based fault diagnosis system: A large amount of labeled aerospace major product manufacturing equipment fault data was used as a training set to preliminarily determine the model parameters of the deep learning-based fault diagnosis system and establish a convolutional neural network (CNN) fault diagnosis model that incorporates EMD-SDP features.

[0010] Step 2. Based on the convolutional neural network (CNN) fault diagnosis model with EMD-SDP feature fusion established in Step 1, an improved autoencoder network is built to establish a time series data anomaly detection system to detect abnormal data sequences during the operation of major aerospace product manufacturing equipment;

[0011] Step 3. Use the time series data anomaly detection system established in step 2 to determine whether there are any anomalies in the production process data of the major aerospace product manufacturing equipment to be tested. For the data determined to be abnormal, perform fault diagnosis using the deep learning-based fault diagnosis system established in step 1.

[0012] The specific method of step 1 is:

[0013] 1.1) Stabilize the original vibration signal through EMD decomposition, thereby adaptively extracting local feature signals at different time scales:

[0014] 1.1.1) Determine all local extreme points of the signal x(t) and connect them with a cubic spline curve to form the envelope of the signal data;

[0015] 1.1.2) Calculate the mean of the envelope and find the difference between it and the original signal, that is:

[0016] y1(t)=x(t)-μ1 (1)

[0017] Where: x(t) is the original signal, μ1 is the mean value of the envelope, and y1(t) is the difference signal;

[0018] 1.1.3) Determine whether the result of step 1.1.2) satisfies the modal component function condition. If it does not, use y1(t) as the data to be processed and iterate steps 1.1.1) and 1.1.2) until y1(t) satisfies the modal component function condition. Then, let y1(t) = c1(t) to obtain the first-order modal component c1(t) of the original signal x(t). The first-order modal component c1(t) includes the highest-frequency fluctuation component in the original signal x(t). Whether the SD value of the results of two adjacent iterative processes is less than a threshold is used as the condition for determining whether the first-order modal component c1(t) is a modal component function. SD is defined as:

[0019]

[0020] 1.1.4) Separate the modal components from the original signal x(t), that is:

[0021] r1(t)=x(t)-c1(t) (3)

[0022] Where: x(t) is the original signal, c1(t) is the first-order modal component, and r1(t) is the difference signal;

[0023] 1.1.5) Using the difference signal obtained in step 1.1.4) as the original signal, repeat the above steps to obtain the other n modal components in sequence, and normalize the modal components, that is:

[0024]

[0025] Where: c i (t) is the modal component of each order, r i (t) is the difference signal of each order;

[0026] When the modal component c n (t) reaches the set threshold or remainder r n When (t) becomes a monotonic function, the loop process is terminated and the equations (3) and (4) are added together to obtain:

[0027]

[0028] In formula (5): r n (t) represents the residual function of the average trend of the signal, c i (t) is the signal component from high to low in the signal frequency band;

[0029] 1.2) Using the SDP analysis method to fuse the feature information of the modal components to obtain the SDP feature image under the vibration signal;

[0030] Map the normalized time domain waveform of the modal component in step 1.1.5) to the polar coordinates to generate a symmetrical point pattern, that is, the time domain vibration signal X = {x1, x2, ..., x i ,...,x n}Convert to a point in polar coordinate space

[0031] In the polar coordinate plane, a point in the time domain waveform is mapped to a radial component, and the adjacent point is mapped to an angular component; r(i) is the radial component, θ(i) is the counterclockwise rotation angle of the mirror symmetry plane, is the clockwise rotation angle of the mirror image plane. The specific calculation formulas are as follows:

[0032]

[0033]

[0034]

[0035] In formulas (6) to (8), lt is the time lag coefficient; x i is the i-th sample point of the data signal; x max is the maximum amplitude of the signal x(t); min is the minimum amplitude of the signal x(t); θ is the deflection angle of the mirror symmetry plane ξ is the amplification factor (ξ≤θ), and the time domain waveform is converted into a polar coordinate image using the SDP method to present the fault vibration signal characteristics;

[0036] Each mirror plane of the SDP image represents a set of data features. The modal component functions obtained by EMD decomposition in step 1.1.4) are expressed in the same polar coordinates using the SDP image. By fusing the information features of each modal component, a more complete representation of the original vibration signal is achieved.

[0037] 1.3) Using the SDP image from step 1.2) as input to a convolutional neural network model (CNN) to implement data-driven feature learning, thus enabling adaptive feature extraction of vibration signals and classification and identification of fault types;

[0038] CNN performs repeated convolution and pooling operations on the input image by setting up multi-level feature filters to achieve the purpose of automatic data feature extraction and dimensionality reduction, and obtain data features that are rotation and translation invariant;

[0039] 1.3.1) The convolution layer extracts the feature matrix of the local area through convolution kernel weight sharing:

[0040]

[0041] Where: M j is the input data set; l is the lth layer of the network; is the activation function value of the i-th feature vector in the l-1th layer; is the convolution kernel vector of the jth eigenvector in the lth layer and the i-th eigenvector in the l-1th layer; * is the convolution operator; is the deviation of the jth feature vector in the lth layer; f1(·) is the activation function of the convolutional layer, using the S-type function or ReLU function;

[0042] 1.3.2) Add a pooling layer to reduce the feature matrix obtained by equation (9) and reduce the number of network parameters. By performing the local extreme value extraction operation on the feature matrix, the number of network parameters is reduced while obtaining a feature matrix with displacement invariance. The pooling operation process is as follows:

[0043]

[0044] Where: down(·) is the downsampling function; is the connection weight; —— represents the bias of this layer; f2(·) is the activation function of the pooling layer;

[0045] 1.3.3) Add the same number of neurons as the fault categories to the CNN model as the fully connected output layer:

[0046] O=f3(b o +ω o f v ) (11)

[0047] Where: b o is the deviation vector; ω o is the weight matrix; f v is the feature vector; f3(·) is the activation function of the fully connected layer;

[0048] 1.3.4) Train the CNN model obtained in step 1.3.3)

[0049] The backpropagation algorithm is used to adjust the network parameters through the output labels of the training samples. The hyperparameters used include the convolution kernel size k, the weight coefficients β and ω, and the bias b of each layer. The model parameters are adjusted by calculating the difference between the actual output and the ideal output, so that the actual output is closer to the ideal output. Finally, a model suitable for the fault classification task is obtained. The loss function is calculated as follows:

[0050]

[0051] Where: O pj Output of actual fault type; Y pj Output for ideal fault type: E p is the loss function difference.

[0052] The specific method of step 2 is:

[0053] 2.1) Construction of improved autoencoder network

[0054] The feature matrix in step 1.3.1) is recorded as X and input into the encoder. After the convolution operation in the encoder, the output is sent to an attention mechanism model. The temporal attention mechanism is used to adaptively select the step size related to the current step size.

[0055] Perform deconvolution operation in the decoder and output the reconstructed feature matrix X1 of X to obtain the improved autoencoder model;

[0056] The specific calculation is as follows:

[0057] The input data is first encoded and then decoded. The model is trained by comparing the difference between the decoded data and the original input data to obtain stable parameters. Formula (13) is the encoding process, Formula (14) is the decoding process, and Formula (15) is the objective function (mean-square error, MSE), where σ represents the activation function, W represents the weight of the autoencoder, and b represents the bias of the autoencoder.

[0058] z=σ(W xh x+b xh ) (13)

[0059]

[0060]

[0061] 2.2) Building a time series anomaly detection model

[0062] 2.2.1) Construct a time series data feature matrix to represent the multi-level system states at different time steps;

[0063] The relationship between multiple features from time t-ω to time t is recorded as the sequence from t-ω to t as L ωt , let the total number of features be n, obviously L ωt is n vectors of length ω; using sequence L ωt The cosine similarity of the two variables within the constructs an n×n feature matrix M t , M tAs shown in formula (16);

[0064]

[0065] In formula (16), Defined as a variable and variables In order to characterize the system state at different scales, multiple feature matrices with different lengths of ω are constructed at each time step, that is, the size of the unit feature matrix is ​​n×n;

[0066] 2.2.2) Based on the feature matrix constructed in step 2.2.1), a convolutional encoder is used to encode time series correlations, and an attention-based long short-term memory (ALSTM) network is used to capture more complex temporal patterns, thereby improving the accuracy and scalability of the time series anomaly detection model.

[0067] For the constructed time series data feature matrix, a convolutional encoder is used to encode the spatial pattern of the system feature matrix: connect M at different scales t As a tensor X t,0 ∈R n×n×s , and then output it to several convolutional layers; assuming Represents the feature matrix of the l-1 layer, then the output of the l-th layer is shown in formula (17):

[0068] X t,l =f4(W l *X t,l-1 +b l ) (17)

[0069] Where: * represents the convolution operation, f4(·) represents the activation function, Indicates size k l ×k l ×s l-1 d l convolution kernel, b l ∈R sl is a bias term, Represents the output feature matrix of the lth layer; feature extraction is performed through CNN to obtain a temporal feature matrix with spatial correlation, which is then sequentially input into the attention-based LSTM model to adaptively select relevant hidden states at different time steps; specifically, given the feature matrix X of the lth convolutional layer t,l and the previous hidden state The current state H t,l , through H t,l =(X t,l ,H t-1,l ) to update, the LSTM cell is calculated by formula;

[0070]

[0071] Where: * represents the convolution operation, ° represents the Hadamard product, σ represents the activation function, and Indicates size of convolution kernels, and is the bias parameter of the lth layer of the LSTM model;

[0072] The output size of each layer in the LSTM model corresponds to the output size of the convolution; all input X t,l , cell output c t ,l , hidden state H t,l and three gates (i t,l 、f t,l and o t,l ) are all three-dimensional tensors; tune the step size h (sequence length of the feature matrix), use the temporal attention mechanism to adaptively select the step size related to the current step size, and aggregate the representations of these information feature matrices to form an accurate output feature matrix, which is calculated as shown in formula (19):

[0073]

[0074] Where: Vec(·) represents the vector, γ is the adjustment factor; that is, the last hidden state H t,l As the context vector of the group, the importance weight α of the previous step is measured by the softmax function i ;

[0075] 2.2.3) The precise output feature matrix obtained in step 2.2.2) is used for anomaly detection and diagnosis using residual loss: When given an abnormal feature matrix, the model cannot reconstruct the feature matrix very well, i.e., the residual error generated is higher than that of the precise output feature matrix of the normal sequence.

[0076] The specific method of step 3 is:

[0077] Step 3.1. For the production process data of the major aerospace product manufacturing equipment to be tested, first use the data anomaly detection system in Step 2 to detect whether there are anomalies in the data, that is, whether the equipment may have malfunctioned during the period;

[0078] Step 3.2. After determining that the data to be tested is abnormal in step 3.1, input the abnormal data into the deep learning-based fault diagnosis system established in step 1 to detect the type of fault corresponding to the data;

[0079] Step 3.3. Use the data completed in the fault detection of step 3.2 as a training set for the deep learning-based fault diagnosis system to continuously update the parameters of the deep learning-based fault diagnosis system.

[0080] The system based on the above-mentioned full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment includes:

[0081] The fault diagnosis model pre-training module preliminarily establishes a deep learning-based fault diagnosis system through step 1. It uses a large amount of labeled aerospace major product manufacturing equipment fault data as a training set to preliminarily determine the model parameters of the deep learning-based fault diagnosis system.

[0082] The abnormal data identification module establishes a time series data anomaly detection system through step 2 to detect abnormal data sequences in the operation process of major aerospace product manufacturing equipment;

[0083] The fault diagnosis model adaptive adjustment module inputs the abnormal data identified by the abnormal data identification module into the deep learning-based fault diagnosis system initially established in step 1 to identify the fault type, and uses this data as a training set for the deep learning-based fault diagnosis system for learning and training, so as to continuously update the parameters of the deep learning-based fault diagnosis system.

[0084] Equipment based on the above-mentioned full life cycle adaptive fault diagnosis method for major aerospace product manufacturing equipment includes:

[0085] memory for storing computer programs;

[0086] A processor is used to implement the full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment described in steps 1 to 3 when executing the computer program.

[0087] A computer-readable storage medium stores a computer program, which, when executed by a processor, can implement full-life cycle adaptive fault diagnosis of aerospace major product manufacturing equipment based on the method described in steps 1 to 3.

[0088] Compared with the prior art, the present invention has the following benefits:

[0089] The present invention preliminarily establishes a fault diagnosis system based on deep learning through step 1, and trains a large amount of labeled fault data of major aerospace product manufacturing equipment as a training set to preliminarily determine the model parameters of the fault diagnosis system based on deep learning; through step 2, a time series data anomaly detection system is established to detect abnormal data sequences in the operation process of major aerospace product manufacturing equipment, and the detected abnormal data sequences are input into the fault diagnosis system based on deep learning established in step 1 to detect what kind of fault the data corresponds to; finally, the abnormal data and its data type are used as a training set of the fault diagnosis system based on deep learning for learning and training, and the parameters of the fault diagnosis system based on deep learning are continuously updated, so that the fault diagnosis model parameters are continuously adjusted as the equipment operates and will not become invalid with the operation of the equipment, thereby realizing full life cycle adaptive fault diagnosis of major aerospace product manufacturing equipment.

[0090] The present invention uses labeled data to preliminarily train a fault diagnosis model, uses the abnormal data diagnosis model to detect abnormal data from sensors in the equipment during the operation of major aerospace production and manufacturing equipment, and uses the abnormal data as a new training set to continuously update the parameters of the fault diagnosis model, so that the fault diagnosis model proposed by the present invention can have good fault diagnosis effects throughout the entire life cycle of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is a system flow chart of the present invention.

[0092] Figure 2 It is a structural diagram of the CNN model of the present invention.

[0093] Figure 3 It is the overall architecture of the autoencoder network of the present invention.

[0094] Figure 4 Schematic diagram of the autoencoder structure of the present invention.

[0095] Figure 5 It is the time series anomaly detection model of the present invention.

[0096] Figure 6 It is the time series data feature matrix of the present invention. DETAILED DESCRIPTION

[0097] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0098] A full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment, see Figure 1 , specifically including the following steps:

[0099] Step 1. Initially establish a deep learning-based fault diagnosis system. A large amount of labeled aerospace major product manufacturing equipment fault data is used as a training set to preliminarily determine the model parameters of the deep learning-based fault diagnosis system. A convolutional neural network (CNN) fault diagnosis model is established that incorporates EMD-SDP feature fusion.

[0100] 1.1) Stabilize the original vibration signal of the extended endpoint data through EMD decomposition, thereby adaptively extracting local feature signals of different time scales:

[0101] 1.1.1) Determine all local extreme points of the signal x(t) and connect them with a cubic spline curve to form the envelope of the signal data;

[0102] 1.1.2) Calculate the mean of the envelope and find the difference between it and the original signal, that is,

[0103] y1(t)=x(t)-μ1 (1)

[0104] Where x(t) is the original signal, μ1 is the mean of the envelope, and y1(t) is the difference signal;

[0105] 1.1.3) Determine whether the result of step 1.2) satisfies the modal component condition. If it is not a modal component function, use y1(t) as the data to be processed and continue with steps 1.1) and 1.2) above until y1(t) satisfies the modal component function condition. Then, let y1(t) = c1(t) to obtain the first modal component c1(t) of signal x(t). This component includes the highest-frequency fluctuation component in signal x(t). However, in practical applications, a stopping condition is required to terminate the iteration process. The SD value of the results of two adjacent iterations is less than a threshold value as the condition for determining whether it is a modal component function. SD is defined as:

[0106]

[0107] 1.1.4) Separate the modal components from the original signal, that is:

[0108] r1(t)=x(t)-c1(t) (3)

[0109] Where: x(t) is the original signal, c1(t) is the first-order modal component, and r1(t) is the difference signal;

[0110] 1.1.5) Using the difference signal as the original signal, repeat the above steps to obtain the other n modal components, namely:

[0111]

[0112] Where: c i (t) is the modal component of each order, r i (t) is the difference signal of each order;

[0113] When the modal component c n (t) reaches a certain threshold or remainder r n When (t) becomes a monotonic function, the loop process is terminated and the equations (3) and (4) are added together to obtain:

[0114]

[0115] In formula (5): r n (t) represents the residual function of the average trend of the signal, c i (t) is the signal component from high to low in the signal frequency band; generally, the first few components obtained by EMD decomposition usually contain the most important feature information of the original signal;

[0116] At the same time, considering that in EMD decomposition, there is an endpoint effect when the Paul curve is obtained by performing spline interpolation fitting on the extreme points of the original signal, which may cause signal divergence and thus affect the decomposition accuracy; the method of extending the signal sampling time and extending the endpoint data is adopted to suppress the endpoint effect during EMD decomposition. Specifically, when extracting the fault vibration signal data, a data signal with a signal cycle length is intercepted at both ends. The EMD method can effectively decompose the signal into the sum of several modal components, and the signal decomposition result after the endpoint data is intercepted is close to the ideal decomposition result, so that the signal characteristics can be displayed at different resolutions.

[0117] 1.2) Using the SDP analysis method to fuse the feature information of the modal components to obtain the SDP feature image under the vibration signal;

[0118] The normalized time domain waveform is mapped to polar coordinates to generate a symmetrical point pattern, that is, the time domain vibration signal X={x1,x2,...,x i ,...,x n}Convert to a point in polar coordinate space

[0119] In the polar coordinate plane, a point in the time domain waveform is mapped to a radial component, and the adjacent point is mapped to an angular component; r(i) is the radial component, θ(i) is the counterclockwise rotation angle of the mirror symmetry plane, It is the clockwise rotation angle of the mirror image plane. The specific calculation formula is as follows:

[0120]

[0121]

[0122]

[0123] Where: lt is the time lag coefficient; x i is the i-th sample point of the data signal; x max is the maximum amplitude of the signal x(t); min is the minimum amplitude of the signal x(t); θ is the deflection angle of the mirror symmetry plane ξ is the amplification factor (ξ≤θ), and the time domain waveform is converted into a polar coordinate image using the SDP method to present the fault vibration signal characteristics;

[0124] Each mirror plane of the SDP image can represent a set of data features. The modal component functions obtained by EMD decomposition in step 1 are expressed in the same polar coordinates using the SDP image. By fusing the information features of each modal component, a more complete expression of the original vibration signal is achieved.

[0125] 1.3) The SDP image from step 2) is used as the input of the convolutional neural network model CNN to achieve data-driven feature learning, adaptive feature extraction of vibration signals and classification and identification of fault types. Figure 2 ;

[0126] CNN performs repeated convolution and pooling operations on the input image by setting up multi-level feature filters to achieve the purpose of automatic data feature extraction and dimensionality reduction, and obtain data features with rotation and translation invariance.

[0127] 1.3.1) The convolution layer extracts data features of local areas through convolution kernel weight sharing:

[0128] ( Figure 4 ) The convolution operation can not only reduce the complexity of the network model, but also significantly improve the learning efficiency of the network. The specific operation process is shown in formula (9).

[0129]

[0130] Where: M j is the input data set; l is the lth layer of the network; is the activation function value of the i-th feature vector in the l-1th layer; is the convolution kernel vector of the jth eigenvector in the lth layer and the i-th eigenvector in the l-1th layer; * is the convolution operator; is the deviation of the jth feature vector in the lth layer; f1(·) is the activation function of the convolutional layer, using the S-type function or ReLU function;

[0131] 1.3.2) By adding a pooling layer to reduce the size of the feature space and the number of network parameters, maximum pooling can reduce the parameters while obtaining features with displacement invariance by extracting local extreme values ​​from the feature matrix. The pooling operation process is as follows:

[0132]

[0133] Where: down(·) is the downsampling function; is the connection weight; —— represents the bias of this layer; f2(·) is the activation function of the pooling layer;

[0134] 1.3.3) Add the same number of neurons as the fault categories to the CNN model as the fully connected output layer:

[0135] O=f3(b o +ω o f v ) (11)

[0136] Where: b o is the deviation vector; ω o is the weight matrix; f v is the feature vector; f3(·) is the activation function of the fully connected layer;

[0137] 1.3.4) Model Training

[0138] During model training, CNN uses the backpropagation algorithm to adjust network parameters based on the output labels of training samples. The hyperparameters used include the convolution kernel size k, weight coefficients β and ω, and the bias b of each layer. The model parameters are adjusted by calculating the difference between the actual output and the ideal output, making the actual output closer to the ideal output. Ultimately, a model suitable for fault classification tasks is obtained. The loss function is calculated as follows:

[0139]

[0140] Where: O pj Output of actual fault type; Y pj Output for ideal fault type: E p is the loss function difference.

[0141] Step 2. Based on the convolutional neural network (CNN) fault diagnosis model with EMD-SDP feature fusion established in Step 1, an improved autoencoder network is built, and then a time series data anomaly detection system is established to detect abnormal data sequences during the operation of major aerospace product manufacturing equipment. The specific method is as follows:

[0142] For a given historical time series data X with a sequence length of T and a dimension of n,n ) T ∈R n×T ,Assuming that there is no abnormal data in X or the proportion of abnormal data is small, the goal of the time series data anomaly detection model includes two parts: anomaly detection and anomaly diagnosis. Anomaly detection is to detect abnormal sequences at a certain time step after T, and anomaly diagnosis is to identify the features that are most likely to cause anomalies;

[0143] 2.1) Construction of improved autoencoder network, see Figure 3 、 Figure 4 ;

[0144] First, the feature matrix X is input into the encoder, and a convolution operation is performed in the encoder;

[0145] The feature matrix output by the encoder is then input into an attention mechanism model, and the temporal attention mechanism is used to adaptively select the step size related to the current step size;

[0146] Finally, a deconvolution operation is performed in the decoder to output the reconstructed feature matrix X1 of X to obtain the improved autoencoder model;

[0147] The specific calculation is as follows:

[0148] The input data is first encoded and then decoded. The model is trained by comparing the difference between the decoded data and the original input data to obtain stable parameters. Formula (13) is encoding, formula (14) is decoding, and formula (15) is the objective function (mean-square error, MSE), where: σ represents the activation function, W represents the weight of the autoencoder, and b represents the bias of the autoencoder.

[0149] z=σ(W xh x+b xh ) (13)

[0150]

[0151]

[0152] 2.2) Build a time series anomaly detection model, see Figure 5 ;

[0153] 2.2.1) Construct a time series data feature matrix to represent the multi-level system states at different time steps; see Figure 6 ;

[0154] The relationship between multiple features from time t-ω to time t is recorded as the sequence from t-ω to t as L ωt , let the total number of features be n, obviously L ωtis n vectors of length ω; using sequence L ωt The cosine similarity of the two variables within the constructs an n×n feature matrix M t , M t As shown in formula (16);

[0155]

[0156] In formula (16), Defined as a variable and variables In order to characterize the system state at different scales, multiple feature matrices with different lengths of ω are constructed at each time step, that is, the size of the unit feature matrix is ​​n×n;

[0157] 2.2.2) Based on the feature matrix constructed in step 2.2.1), a convolutional encoder is used to encode time series correlations, and an attention-based long short-term memory (ALSTM) network is used to capture more complex temporal patterns, thereby improving the accuracy and scalability of the time series anomaly detection model.

[0158] For the constructed time series data feature matrix, a convolutional encoder is used to encode the spatial pattern of the system feature matrix: connect M at different scales t As a tensor X t,0 ∈R n×n×s , and then output it to several convolutional layers; assuming Represents the feature matrix of the l-1 layer, then the output of the l-th layer is shown in formula (17):

[0159] X t,l =f4(W l *X t,l-1 +b l ) (17)

[0160] Table 1 Encoder partial architecture

[0161]

[0162] Where: * represents the convolution operation, f4(·) represents the activation function, Indicates size k l ×k l ×s l-1 d l convolution kernel, b l ∈R sl is a bias term, Represents the output feature matrix of the lth layer; feature extraction is performed through CNN to obtain a temporal feature matrix with spatial correlation, which is then sequentially input into the attention-based LSTM model to adaptively select relevant hidden states at different time steps; specifically, given the feature matrix X of the lth convolutional layer t,l and the previous hidden state The current state H t,l , through H t,l =(X t,l ,H t-1,l ) to update, the LSTM cell is calculated by formula;

[0163]

[0164] Among them: * represents the convolution operation, ° represents the Hadamard product, σ represents the activation function, and Indicates size s l convolution kernels, and is the bias parameter of the lth layer of the LSTM model;

[0165] The output size of each layer in the LSTM model corresponds to the output size of the convolution; it should be noted that all input X t ,l , cell output c t,l , hidden state H t,l and three gates (i t,l 、f t,l and o t,l ) are all three-dimensional tensors; the step size h (sequence length of the feature matrix) is tuned, and the temporal attention mechanism is used to adaptively select the step size related to the current step size. The representations of these information feature matrices are aggregated to form an accurate output feature matrix. The calculation formula is shown in Equation (19):

[0166]

[0167] Where: Vec(·) represents the vector, γ is the adjustment factor; that is, the last hidden state H t,l As the context vector of the group, the importance weight α of the previous step is measured by the softmax function i ; This model can reduce the probability of deterioration of detection model performance.

[0168] 2.2.3) Based on the feature matrix that encodes the intrinsic correlation and temporal information of the features, the convolution decoder is used to reconstruct the input feature matrix and the residual loss is used for anomaly detection and diagnosis. Here, four convolutional layers are used for feature extraction, as shown in Table 1.

[0169] Step 3. Use the time series data anomaly detection system established in Step 2 to determine whether there are any anomalies in the production process data of the major aerospace product manufacturing equipment to be tested. For the data determined to be abnormal, use the deep learning-based fault diagnosis system established in Step 1 to perform fault diagnosis. The specific method is as follows:

[0170] Step 3.1. For the production process data of the major aerospace product manufacturing equipment to be tested, first use the data anomaly detection system in Step 2 to detect whether there are anomalies in the data, that is, whether the equipment may have malfunctioned during the period;

[0171] Step 3.2. After determining that the data to be tested is abnormal in step 3.1, input the abnormal data into the deep learning-based fault diagnosis system established in step 1 to detect the type of fault corresponding to the data;

[0172] Step 3.3. Use the data completed in the fault detection of step 3.2 as a training set for the deep learning-based fault diagnosis system to continuously update the parameters of the deep learning-based fault diagnosis system.

[0173] First, in step 1, a deep learning-based fault diagnosis system was initially established. A large amount of labeled fault data from major aerospace product manufacturing equipment was used as a training set to preliminarily determine the model parameters of the deep learning-based fault diagnosis system. At this point, the model had the ability to identify faults. However, as the equipment ran for a period of time, its operating conditions changed, and the data characteristics measured by the sensors also changed accordingly. At this time, the model had no adaptive parameter adjustment function and might fail as the equipment ran.

[0174] Secondly, in step 2, a time series data anomaly detection system is established to detect abnormal data sequences in the operation process of major aerospace product manufacturing equipment. After the abnormal data sequence is determined, the abnormal data sequence is input into the deep learning-based fault diagnosis system established in step 1 to detect what kind of fault the data segment corresponds to.

[0175] Finally, the abnormal data and its data type are used as the training set of the deep learning-based fault diagnosis system for learning and training, so as to continuously update the parameters of the deep learning-based fault diagnosis system.

[0176] The fault diagnosis model parameters proposed in this invention continuously adjust as the equipment operates, preventing them from becoming ineffective during operation. This enables adaptive fault diagnosis throughout the lifecycle of critical aerospace manufacturing equipment. This prevents the failure of the fault diagnosis model due to long-term use of production equipment, improving the accuracy and versatility of the fault diagnosis system.

[0177] After initially training a fault diagnosis model using labeled data, the present invention uses the abnormal data diagnosis model to detect abnormal data from sensors in the equipment during the operation of major aerospace production and manufacturing equipment, and uses the abnormal data as a new training set to continuously update the parameters of the fault diagnosis model, so that the fault diagnosis model proposed by the present invention can have good fault diagnosis effects throughout the entire life cycle of the equipment.

Claims

1. A full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment, characterized by: The specific steps include: Step 1. Preliminary establishment of a deep learning-based fault diagnosis system: A large amount of labeled aerospace major product manufacturing equipment fault data was used as a training set to preliminarily determine the model parameters of the deep learning-based fault diagnosis system and establish a convolutional neural network (CNN) fault diagnosis model that incorporates EMD-SDP features. 1.1) Stabilize the original vibration signal through EMD decomposition, thereby adaptively extracting local feature signals at different time scales: 1.2) Using the SDP analysis method to fuse the feature information of the modal components to obtain the SDP feature image under the vibration signal; 1.3) Using the SDP image from step 1.2) as input to a convolutional neural network model (CNN) to implement data-driven feature learning, thus enabling adaptive feature extraction of vibration signals and classification and identification of fault types; CNN performs repeated convolution and pooling operations on the input image by setting up multi-level feature filters to achieve the purpose of automatic data feature extraction and dimensionality reduction, and obtain data features that are rotation and translation invariant; 1.3.1) The convolution layer extracts the feature matrix of the local area through convolution kernel weight sharing: Where: M j is the input data set; l is the lth layer of the network; is the activation function value of the i-th feature vector in the l-1th layer; is the convolution kernel vector of the jth eigenvector in the lth layer and the i-th eigenvector in the l-1th layer; * is the convolution operator; is the deviation of the jth feature vector in the lth layer; f1(·) is the activation function of the convolutional layer, using the S-type function or ReLU function; 1.3.2) Add a pooling layer to reduce the feature matrix obtained by equation (9) and reduce the number of network parameters. By performing the local extreme value extraction operation on the feature matrix, the number of network parameters is reduced while obtaining a feature matrix with displacement invariance. The pooling operation process is as follows: Where: down(·) is the downsampling function; is the connection weight; —— represents the bias of this layer; f2(·) is the activation function of the pooling layer; 1.3.3) Add the same number of neurons as the fault categories to the CNN model as the fully connected output layer: O=f3(b o +ω o f v ) (11) Where: b o is the deviation vector; ω o is the weight matrix; f v is the feature vector; f3(·) is the activation function of the fully connected layer; 1.3.4) Training the CNN model obtained in step 1.3.3); Step 2. Based on the EMD-SDP feature fusion convolutional neural network (CNN) fault diagnosis model established in Step 1, an improved autoencoder network is constructed. The shift-invariant feature matrix obtained in Step 1.3) is recorded as X and input into the encoder. After performing a convolution operation in the encoder, the output is fed into an attention mechanism model. The temporal attention mechanism is used to adaptively select a step size related to the current step size. A deconvolution operation is performed in the decoder, and the reconstructed feature matrix X1 of X is output to obtain the improved autoencoder model. Then, a time series data anomaly detection system is established to detect abnormal data sequences during the operation of major aerospace product manufacturing equipment. Step 3. Use the time series data anomaly detection system established in step 2 to determine whether there are any anomalies in the production process data of the major aerospace product manufacturing equipment to be tested. For the data determined to be abnormal, perform fault diagnosis using the deep learning-based fault diagnosis system established in step 1.

2. The full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment according to claim 1 is characterized in that: The specific method of step 1 is: 1.1) Stabilize the original vibration signal through EMD decomposition, thereby adaptively extracting local feature signals at different time scales: 1.1.1) Determine all local extreme points of the signal x(t) and connect them with a cubic spline curve to form the envelope of the signal data; 1.1.2) Calculate the mean of the envelope and find the difference between it and the original signal, that is: y1(t)=x(t)-μ1 (1) Where: x(t) is the original signal, μ1 is the mean value of the envelope, and y1(t) is the difference signal; 1.1.3) Determine whether the result of step 1.1.2) satisfies the modal component function condition. If it does not, use y1(t) as the data to be processed and iterate steps 1.1.1) and 1.1.2) until y1(t) satisfies the modal component function condition. Then, let y1(t) = c1(t) to obtain the first-order modal component c1(t) of the original signal x(t). The first-order modal component c1(t) includes the highest-frequency fluctuation component in the original signal x(t). Whether the SD value of the results of two adjacent iterative processes is less than a threshold is used as the condition for determining whether the first-order modal component c1(t) is a modal component function. SD is defined as: 1.1.4) Separate the modal components from the original signal x(t), that is: r1(t)=x(t)-c1(t) (3) Where: x(t) is the original signal, c1(t) is the first-order modal component, and r1(t) is the difference signal; 1.1.5) Using the difference signal obtained in step 1.1.4) as the original signal, repeat the above steps to obtain the other n modal components in sequence, and normalize the modal components, that is: Where: c i (t) is the modal component of each order, r i (t) is the difference signal of each order; When the modal component c n (t) reaches the set threshold or remainder r n When (t) becomes a monotonic function, the loop process is terminated and the equations (3) and (4) are added together to obtain: In formula (5): r n (t) represents the residual function of the average trend of the signal, c i (t) is the signal component from high to low in the signal frequency band; 1.2) Using the SDP analysis method to fuse the feature information of the modal components to obtain the SDP feature image under the vibration signal; Map the time domain waveform of the modal component after normalization in step 1.1.5) to the polar coordinates to generate a symmetrical point pattern, that is, the time domain vibration signal X={x1,x2,...,x i ,...,x n }Convert to a point in polar coordinate space In the polar coordinate plane, a point in the time domain waveform is mapped to a radial component, and the adjacent point is mapped to an angular component; r(i) is the radial component, θ(i) is the counterclockwise rotation angle of the mirror symmetry plane, is the clockwise rotation angle of the mirror image plane. The specific calculation formulas are as follows: In formulas (6) to (8), lt is the time lag coefficient; x i is the i-th sample point of the data signal; x max is the maximum amplitude of the signal x(t); min is the minimum amplitude of the signal x(t); θ is the deflection angle of the mirror symmetry plane ξ is the amplification factor (ξ≤θ), and the time domain waveform is converted into a polar coordinate image using the SDP method to present the fault vibration signal characteristics; Each mirror plane of the SDP image represents a set of data features; The modal component functions obtained by EMD decomposition in step 1.1.4) are expressed in the same polar coordinates using SDP images. By fusing the information features of each modal component, a more complete representation of the original vibration signal is achieved. 1.3) Using the SDP image from step 1.2) as input to a convolutional neural network model (CNN) to implement data-driven feature learning, thus enabling adaptive feature extraction of vibration signals and classification and identification of fault types; CNN performs repeated convolution and pooling operations on the input image by setting up multi-level feature filters to achieve the purpose of automatic data feature extraction and dimensionality reduction, and obtain data features that are rotation and translation invariant; 1.3.1) The convolution layer extracts the feature matrix of the local area through convolution kernel weight sharing: Where: M j is the input data set; l is the lth layer of the network; is the activation function value of the i-th feature vector in the l-1th layer; is the convolution kernel vector of the jth eigenvector in the lth layer and the i-th eigenvector in the l-1th layer; * is the convolution operator; is the deviation of the jth feature vector in the lth layer; f1(·) is the activation function of the convolutional layer, using the S-type function or ReLU function; 1.3.2) Add a pooling layer to reduce the feature matrix obtained by equation (9) and reduce the number of network parameters. By performing the local extreme value extraction operation on the feature matrix, the number of network parameters is reduced while obtaining a feature matrix with displacement invariance. The pooling operation process is as follows: Where: down(·) is the downsampling function; is the connection weight; —— represents the bias of this layer; f2(·) is the activation function of the pooling layer; 1.3.3) Add the same number of neurons as the fault categories to the CNN model as the fully connected output layer: O=f3(b o +ω o f v ) (11) Where: b o is the deviation vector; ω o is the weight matrix; f v is the feature vector; f3(·) is the activation function of the fully connected layer; 1.3.4) Train the CNN model obtained in step 1.3.3) The backpropagation algorithm is used to adjust the network parameters through the output labels of the training samples. The hyperparameters used include the convolution kernel size k, the weight coefficients β and ω, and the bias b of each layer. The model parameters are adjusted by calculating the difference between the actual output and the ideal output, so that the actual output is closer to the ideal output. Finally, a model suitable for the fault classification task is obtained. The loss function is calculated as follows: Where: O pj Output of actual fault type; Y pj Output for ideal fault type: E p is the loss function difference.

3. The full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment according to claim 1 is characterized in that: The specific method of step 2 is: 2.1) Construction of improved autoencoder network The shift-invariant feature matrix obtained in step 1.3) is recorded as X and input into the encoder. After convolution operation in the encoder, the output is sent to an attention mechanism model. The temporal attention mechanism is used to adaptively select the step size related to the current step size. Perform deconvolution operation in the decoder and output the reconstructed feature matrix X1 of X to obtain the improved autoencoder model; The specific calculation is as follows: The input data is first encoded and then decoded. The model is trained by comparing the difference between the decoded data and the original input data to obtain stable parameters. Formula (13) is the encoding process, Formula (14) is the decoding process, and Formula (15) is the objective function (mean-square error, MSE), where σ represents the activation function, W represents the weight of the autoencoder, and b represents the bias of the autoencoder. z=σ(W xh x+b xh ) (13) 2.2) Building a time series anomaly detection model 2.2.1) Construct a time series data feature matrix to represent the multi-level system states at different time steps; The relationship between multiple features from time t-ω to time t is recorded as the sequence from t-ω to t as L ωt , let the total number of features be n, obviously L ωt is n vectors of length ω; using sequence L ωt The cosine similarity of the two variables within the constructs an n×n feature matrix M t , M t As shown in formula (16); In formula (16), Defined as a variable and variables In order to characterize the system state at different scales, multiple feature matrices with different lengths of ω are constructed at each time step, that is, the size of the unit feature matrix is ​​n×n; 2.2.2) Based on the feature matrix constructed in step 2.2.1), a convolutional encoder is used to encode time series correlations, and an attention-based long short-term memory (ALSTM) network is used to capture more complex temporal patterns, thereby improving the accuracy and scalability of the time series anomaly detection model. For the constructed time series data feature matrix, a convolutional encoder is used to encode the spatial pattern of the system feature matrix: connect M at different scales t As a tensor X t,0 ∈R n×n×s , and then output it to several convolutional layers; assuming Represents the feature matrix of the l-1 layer, then the output of the l-th layer is shown in formula (17): X t,l =f4(W l *X t,l-1 +b l ) (17) Among them, * represents the convolution operation, f4(·) represents the activation function, Indicates size k l ×k l ×s l-1 d l convolution kernel, b l ∈R sl is a bias term, Represents the output feature matrix of the lth layer; feature extraction is performed through CNN to obtain a temporal feature matrix with spatial correlation, which is then sequentially input into the attention-based LSTM model to adaptively select relevant hidden states at different time steps; specifically, given the feature matrix X of the lth convolutional layer t,l and the previous hidden state The current state H t,l , through H t,l =(X t,l ,H t -1,l ) to update, the LSTM cell is calculated by formula; Among them: * represents the convolution operation, ° represents the Hadamard product, σ represents the activation function, and Indicates size of convolution kernels, and is the bias parameter of the lth layer of the LSTM model; The output size of each layer in the LSTM model corresponds to the output size of the convolution; all input X t,l , cell output c t,l , hidden state H t,l and three gates (i t,l 、f t,l and o t,l ) are all three-dimensional tensors; tune the step size h (sequence length of the feature matrix), use the temporal attention mechanism to adaptively select the step size related to the current step size, and aggregate the representations of these information feature matrices to form an accurate output feature matrix, which is calculated as shown in formula (19): Where: Vec(·) represents the vector, γ is the adjustment factor; that is, the last hidden state H t,l As the context vector of the group, the importance weight α of the previous step is measured by the softmax function i ; 2.2.3) The exact output feature matrix obtained in step 2.2.2) is used for anomaly detection and diagnosis using residual loss: When given an abnormal feature matrix, the model cannot reconstruct the feature matrix, that is, the residual error generated is higher than the exact output feature matrix of the normal sequence.

4. The full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment according to claim 1 is characterized in that: The specific method of step 3 is: Step 3.

1. For the production process data of the major aerospace product manufacturing equipment to be tested, first use the data anomaly detection system in Step 2 to detect whether there are anomalies in the data, that is, whether the equipment may have malfunctioned during the period; Step 3.

2. After determining that the data to be tested is abnormal in step 3.1, input the abnormal data into the deep learning-based fault diagnosis system established in step 1 to detect the type of fault corresponding to the data; Step 3.

3. Use the data completed in the fault detection of step 3.2 as a training set for the deep learning-based fault diagnosis system to continuously update the parameters of the deep learning-based fault diagnosis system.

5. A system based on the full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment according to any one of claims 1 to 4, characterized in that: include: The fault diagnosis model pre-training module preliminarily establishes a deep learning-based fault diagnosis system through step 1. It uses a large amount of labeled aerospace major product manufacturing equipment fault data as a training set to preliminarily determine the model parameters of the deep learning-based fault diagnosis system. The abnormal data identification module establishes a time series data anomaly detection system through step 2 to detect abnormal data sequences in the operation process of major aerospace product manufacturing equipment; The fault diagnosis model adaptive adjustment module inputs the abnormal data identified by the abnormal data identification module into the deep learning-based fault diagnosis system initially established in step 1 to identify the fault type, and uses this data as a training set for the deep learning-based fault diagnosis system for learning and training, so as to continuously update the parameters of the deep learning-based fault diagnosis system.

6. A device based on the full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment according to any one of claims 1 to 4, characterized in that: include: Memory for storing computer programs; A processor for implementing the full life cycle adaptive fault diagnosis method for aerospace major product manufacturing equipment as described in any one of claims 1 to 4 when executing the computer program.

7. A computer-readable storage medium storing a computer program, which, when executed by a processor, can implement full life cycle adaptive fault diagnosis of major aerospace product manufacturing equipment based on the method described in any one of claims 1 to 4.

Images