Main shaft machining composite fault detection method based on robust multi-working-condition process model

Abstract

The invention belongs to the technical field of rotating machinery fault detection, and discloses a spindle machining composite fault detection method based on a robust multi-working-condition process model, which comprises the following steps: collecting and preprocessing current signals of spindle machining equipment in a normal operation state; the preprocessed current signal passes through a wavelet kernel attention convolution layer and a convolution auto-encoder in sequence; constructing a Gaussian mixture hidden Markov model, obtaining the number of optimal Gaussian components through a Dirichlet process, and calculating the energy corresponding to the low-dimensional representation of the sample; and constructing a joint optimization framework loss function comprising a sample reconstruction error, energy corresponding to sample low-dimensional representation, a penalty term and an evidence lower bound of variational Bayesian inference, and updating network parameters to convergence. According to the invention, the interpretability of the model is improved, and the effectiveness of fault detection is enhanced.

Description

Technical Field

[0001] The invention belongs to the technical field of rotating machinery fault detection, and in particular relates to a spindle machining composite fault detection method based on a robust multi-operating condition process model. Background Art

[0002] Spindle processing equipment occupies a key position in the field of mechanical manufacturing, and its performance directly affects the processing accuracy and quality of products. However, in actual operation, spindle processing equipment is often in complex working conditions, such as sudden load changes, speed fluctuations, etc., coupled with the interaction of factors such as mechanical wear and electrical failures, which can easily lead to compound faults. Traditional fault detection methods are mostly targeted at a single working condition or a single fault type, and it is difficult to effectively detect compound faults under variable working conditions, resulting in frequent false alarms and missed alarms, which seriously affects production efficiency and equipment reliability. Therefore, the study of spindle processing compound fault detection based on a robust multi-condition process model can accurately identify compound faults under variable working conditions, improve the safety and stability of equipment operation, and is of great significance to promoting the development of the mechanical processing industry.

[0003] Existing composite fault detection methods can be roughly divided into three categories: supervised, semi-supervised, and unsupervised. Among them, unsupervised methods are used when there is no labeled data, and anomalies are identified based on the characteristics of the data itself. For example, K-Means divides clusters based on the distance between data points and cluster centers. The defect is that it is sensitive to the initial center, it is difficult to determine the number of clusters, and it is easy to cause unreasonable clustering due to the lack of prior knowledge. Similar methods also include density methods, such as DBSCAN, which judges based on the density around data points. In high dimensions, density estimation is inaccurate, and the parameters have a large impact. The density measurement is difficult due to the complexity of high-dimensional data. Semi-supervised methods are used in scenarios with a small amount of labeled data. Common ones include graph methods, such as the label propagation algorithm, which uses the similarity of data points to construct a graph to judge anomalies. The calculation is complex and sensitive to graph construction because there are many calculation relationships and the construction strategy affects the similarity measurement. Supervised methods use a large amount of labeled data to train models. Common neural network methods such as multi-layer perceptrons and convolutional neural networks learn data and fault label mapping. Training consumes a lot of resources, the model is complex and prone to overfitting, and a large number of weights need to be adjusted and it is easy to over-learn training details. In addition, the composite fault detection method represented by deep learning has multi-layer nonlinear mapping, which causes the network to have "black box" characteristics and the model lacks interpretability. Summary of the invention

[0004] The purpose of the present invention is to provide a spindle machining composite fault detection method based on a robust multi-condition process model, so as to improve the interpretability of the model and enhance the effectiveness of fault detection.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A spindle machining composite fault detection method based on a robust multi-condition process model, the spindle machining composite fault detection method based on a robust multi-condition process model comprising:

[0007] Collect and pre-process the current signal of the spindle processing equipment under normal operating conditions;

[0008] The preprocessed current signal passes through the wavelet kernel attention convolution layer and the convolutional autoencoder in sequence. The wavelet kernel attention convolution layer first uses a variety of wavelet basis functions to extract multi-scale interpretable features of the preprocessed current signal, and then uses the softmax function to process the interpretable features to obtain the sample's significant features. The convolutional autoencoder composed of 1-D convolutional layers obtains a low-dimensional representation of the sample containing multiple wavelet decomposition results based on the sample's significant features.

[0009] Construct a Gaussian mixture hidden Markov model, obtain the optimal number of Gaussian components through the Dirichlet process, use the estimation network to estimate the mixture membership and state transition probability of each hidden state in the Gaussian mixture hidden Markov model corresponding to the observation sequence, calculate the parameters of the Gaussian mixture hidden Markov model based on the estimated values ​​of the mixture membership and state transition probability, and then calculate the energy corresponding to the low-dimensional representation of the sample based on the Gaussian mixture hidden Markov model;

[0010] Construct a joint optimization framework loss function including sample reconstruction error, energy corresponding to the sample low-dimensional representation, penalty term and the lower bound of the evidence of variational Bayesian inference, and update the parameters of the wavelet kernel attention convolution layer, convolutional autoencoder and estimation network until convergence by minimizing the joint optimization framework loss function;

[0011] In the fault detection stage, the newly collected current signal is passed through the wavelet kernel attention convolution layer and the convolution autoencoder to obtain a low-dimensional representation of the sample to be detected, and the low-dimensional representation of the sample to be detected is input into the Gaussian mixture hidden Markov model to obtain energy. The energy corresponding to the low-dimensional representation of the sample to be detected is compared with the threshold to obtain the composite fault detection result of the spindle machining equipment. The threshold is the 99th percentile of all training samples.

[0012] Several optional methods are also provided below, but they are not intended to be additional limitations on the above-mentioned overall solution, but are merely further supplements or preferences. Under the premise that there are no technical or logical contradictions, each optional method can be combined with the above-mentioned overall solution separately, and multiple optional methods can also be combined.

[0013] Preferably, the wavelet kernel attention convolution layer first uses a variety of wavelet basis functions to extract multi-scale interpretable features of the preprocessed current signal, and then uses a softmax function to process the interpretable features to obtain sample salient features, including:

[0014] Take the preprocessed current signal X i ={x i (1),x i (2),…x i (L)} and various wavelet bases φ q (·), X i represents the i-th preprocessed current signal, i.e., the i-th input sample, x i (l) represents the input sample X i For the lth data in , the similarity coefficient between the input sample and the wavelet kernel function is expressed as:

[0015]

[0016] in, It means that the qth type for the lth data depends on two adjustable parameters σ w and u w The wavelet kernel function, σ w is the scaling factor, u w is the shift factor, q=1,…,Q, Q represents the total number of selected wavelet bases, w represents the number of channels, ρ i,w,q (l) represents the input sample X i With wavelet kernel function The similarity coefficient between i (m) represents the input sample X i The mth data in, m is the index and m=1,…,L, Represents the wavelet kernel function The complex conjugate of It means that the qth type for the lmth data depends on two adjustable parameters σ w and u w Wavelet kernel function, N represents the number of input samples, L represents the length of each input sample;

[0017] Select wavelet kernel function Extracting interpretable features from an input sample of length L is Then the interpretable features extracted from the input samples by various wavelet kernel functions are expressed as:

[0018]

[0019] Among them, S i represents the interpretable features extracted from the i-th input sample, s i,w,q represents the interpretable features of the ith input sample extracted by the qth wavelet kernel function in the wth channel, Indicates i,w,q The transpose of , W represents the total number of convolution channels, w = 1, ..., W, q = 1, ..., Q;

[0020] Then the softmax function is used to process the interpretable features to obtain the significant features of the sample The formula is:

[0021]

[0022] Among them, q * is the index of the wavelet kernel function that extracts the significant features of the sample in the channel among the Q wavelet kernel functions, s i,w,q* Indicates the qth * The interpretable features of the ith input sample extracted by the wavelet kernel function in the wth channel, Indicates i,w,q* The transpose of .

[0023] Preferably, the convolutional autoencoder composed of 1-D convolutional layers obtains a low-dimensional representation of the sample containing multiple wavelet decomposition results according to the significant features of the sample, including:

[0024]

[0025] Among them, Z i represents the low-dimensional representation of the sample of the i-th input sample, represents the sample salient features of the i-th input sample, Θ e denotes the weight of a 1-D convolutional layer, and Conv1d(;) denotes a 1-D convolutional layer.

[0026] Preferably, the step of obtaining the optimal number of Gaussian components by using the Dirichlet process comprises:

[0027] Let G be a distribution that satisfies the Dirichlet process, then:

[0028] G~DP(α,G0)

[0029] Where α is the concentration parameter and G0 is a Beta distribution used to generate the mean μ of each Gaussian component. j and covariance Σ j ;

[0030] The weight of the jth Gaussian component is π j The weight v drawn from the Beta distribution for the jth time j , and the unused weights 1-v of the first j-1 Gaussian components i OK, the formula is as follows:

[0031]

[0032] Among them, v i represents the used weight of the i-th Gaussian component among the first j-1 Gaussian components;

[0033] Continue to sample from the Beta distribution and assign the remaining weights to the next Gaussian component until all weights are assigned and satisfy

[0034] When adaptively determining the number of Gaussian components J, we filter out Gaussian components with weights less than the threshold ε by setting a threshold ε, and count all weights π j The number of Gaussian components greater than the threshold ε, the specific process is as follows:

[0035]

[0036] in, is the indicator function, when π j It is 1 when ≥ε, otherwise it is 0.

[0037] Preferably, the method of estimating the mixed membership and state transition probability of each hidden state corresponding to the observation sequence in the Gaussian mixture hidden Markov model by using the estimation network includes:

[0038] Take two multi-layer neural networks as estimation networks, and use the multi-layer neural network to estimate the state transition probability matrix of the hidden state and the mixed membership matrix of each sample under the given hidden state respectively:

[0039]

[0040] Among them, θ A and θ B They correspond to the weight parameters of two multi-layer neural networks, represents the estimate of the state transition probability matrix, represents the estimation of the mixture membership matrix, MLN(·) and softmax(·) represent the multi-layer neural network and activation function respectively, M represents the number of hidden states, and J represents the number of Gaussian components.

[0041] Preferably, the construction includes a joint optimization framework loss function of sample reconstruction error, energy corresponding to the sample low-dimensional representation, a penalty term, and a lower bound of the evidence of variational Bayesian inference, including:

[0042] Constructing the joint optimization framework loss function as follows:

[0043]

[0044] Where N is the number of input samples, λ1, λ2 and λ3 are weighting coefficients, Dis(X i ,X i′) is the reconstruction error between the i-th input sample and the i-th reconstructed input sample, E(χ) is the energy corresponding to the low-dimensional representation of the sample, represents the penalty term, The lower bound of the evidence for variational Bayesian inference with the Dirichlet process as the prior is calculated as follows:

[0045]

[0046] Among them, ELBO represents the lower bound of the evidence for variational Bayesian inference with Dirichlet process as prior, P(Z,π,Q) represents the joint probability distribution containing the sample low-dimensional representation Z, the mixing coefficient π and the hidden state Q, q(π,Q) represents the variational distribution used to approximate the posterior distribution, is the expected value of the joint probability distribution, is the entropy term of the variational distribution.

[0047] The present invention provides a spindle machining composite fault detection method based on a robust multi-condition process model, which has the following beneficial effects compared with the prior art:

[0048] 1. A wavelet kernel attention convolutional autoencoder is proposed. By selecting a wavelet kernel function composed of multiple different wavelet bases, multi-scale signal processing is performed on time series data, and the time-frequency domain features obtained by wavelet decomposition are fully utilized to improve the interpretability of the model.

[0049] 2. A Gaussian mixture hidden Markov model is introduced to characterize the current signal modeling process suitable for variable working conditions, and the Dirichlet process is used to adaptively determine the number of Gaussian components to adapt to the data distribution characteristics under different working conditions and improve the applicability of this method.

[0050] 3. The evidence lower bound of variational Bayesian inference is designed as the loss function term. By estimating the true distribution of variable operating condition data, the stability and robustness of the model are improved, and the generalization ability of the model in semi-supervised learning is optimized, thereby enhancing the effectiveness of fault detection.

[0051] 4. A joint optimization framework of a compression network based on wavelet kernel attention convolutional autoencoder and an estimation network based on Gaussian mixture hidden Markov model is designed to induce representation learning of variable operating condition data in the compression network and density estimation of the estimation network in an end-to-end manner.

[0052] 5. A fault detection method based on non-invasive measurement is proposed, which makes full use of the built-in sensor information of the equipment, avoids the damage to the equipment caused by invasive measurement, and improves the stability of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1The present invention is a flow chart of a spindle machining composite fault detection method based on a robust multi-condition process model. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0056] like Figure 1 As shown, this embodiment proposes a spindle machining composite fault detection method based on a robust multi-condition process model, comprising the following steps:

[0057] Step S1, collecting and preprocessing the current signal of the spindle processing equipment under normal operation. The current sensor installed on the spindle processing equipment is used to collect the current signal of normal operation under various working conditions, and the current signal is mapped to the same scale through maximum-minimum normalization.

[0058] The current signal of the spindle machining equipment under normal operation under unknown working conditions is collected, where the number of samples is N, the monitoring sample timestamp is l=1,…,L, and the i-th monitoring sample time series is recorded as {y i (l)}, i = 1, ..., N. Since the raw current signal collected by the current sensor has a large numerical difference, the raw current data is preprocessed by maximum-minimum normalization to improve the data quality of the raw current data. The specific formula is as follows:

[0059]

[0060] Among them, y i,min and i,max Respectively represent the sequence {y i The minimum and maximum values ​​in (l)} are recorded as X after the maximum-minimum normalization preprocessing operation. i ={x i (1),x i (2),…x i (l)},X i as the i-th input sample.

[0061] Step S2, the preprocessed current signal passes through the wavelet kernel attention convolution layer and the convolutional autoencoder in sequence, wherein the wavelet kernel attention convolution layer first uses a variety of wavelet basis functions to extract multi-scale interpretable features of the preprocessed current signal, and then uses the softmax function to process the interpretable features to obtain sample significant features. The convolutional autoencoder composed of 1-D convolutional layers obtains a low-dimensional representation of the sample containing multiple wavelet decomposition results according to the sample significant features.

[0062] The wavelet kernel attention convolution layer of this embodiment uses three wavelet basis functions, Laplace, Morlet and Mexh, to extract multi-scale physical features of the current signal, generate 12-dimensional physically meaningful interpretable features, and then apply the softmax function to normalize the attention weights to obtain 4-dimensional physically meaningful sample significant features, and then use a convolutional autoencoder composed of 1-D convolutional layers to obtain a 10-dimensional sample low-dimensional representation containing multiple wavelet decomposition results.

[0063] Specifically, given the preprocessed data X i ={x i (1),x i (2),…x i (l)} and various specific wavelet bases φ q (l), the similarity coefficient between the input sample and the wavelet kernel function is expressed as:

[0064]

[0065] in, It means that the qth type for the lth data depends on two adjustable parameters σ w and u w The wavelet kernel function, σ w is the scaling factor, u w is the shift factor, the selected wavelet basis is denoted as q=1,…,Q, Q represents the total type of selected wavelet basis, w represents the number of channels, and m represents each sample x in the input sequence processed by the convolution operation. i (m) is the index or position of the input signal, and the summation over m covers all possible positions of the input signal, m=1,…,L. Represents the wavelet kernel function The complex conjugate of It means that the qth type for the lmth data depends on two adjustable parameters σ w and u w The wavelet kernel function, ρ i,w,q (l) represents the input sample X i With wavelet kernel function The similarity coefficient between them, N represents the number of samples, and L represents the length of each sample.

[0066] Select wavelet kernel function Extracting interpretable features from an input sample of length L is The physically meaningful interpretable features extracted from the input samples by multiple sets of wavelet kernel functions are expressed as:

[0067]

[0068] Among them, S i represents the interpretable features extracted from the i-th input sample, s i,w,q represents the interpretable features of the ith input sample extracted by the qth wavelet kernel function in the wth channel, Indicates i,w,q , W represents the total number of convolution channels, q=1,…,Q, w=1,…,W.

[0069] Then, the self-attention mechanism after applying the softmax function to normalize the attention weights is used to obtain the sample’s significant features. The formula is:

[0070]

[0071] Among them, q * is the index of the wavelet kernel function extracted in this channel that can significantly express the sensitive features (i.e., the significant features of the sample), Indicates the qth * The interpretable features of the ith input sample extracted by the wavelet kernel function in the wth channel, express The transposed matrix of .

[0072] Furthermore, through the 1-D convolution layer, the interpretable salient features are transformed into Projected into a low-dimensional space, it can be expressed as:

[0073]

[0074] Among them, Conv1d(·) represents a 1-D convolution operation, Z i represents the low-dimensional representation of the sample generated by the i-th input sample, Θ e represents the weights of the 1-D convolutional layer.

[0075] After that, a 1-D transposed convolution layer is used to reconstruct the low-dimensional representation of the sample back to the high-dimensional space, which can be expressed as:

[0076] X i ′=ConvTranspose1d(Z i ; Θ d )

[0077] Among them, ConvTranspose1d(·) represents the operation of 1-D transposed convolution, X i ′ represents the i-th reconstructed input sample, Θ d Represents the weights of a 1-D transposed convolutional layer.

[0078] Therefore, the input sample X i The reconstruction error of is defined as follows:

[0079]

[0080] Among them, Dis(X i ,X i ′) is the reconstruction error between the ith input sample and the ith reconstructed sample.

[0081] Step S3, construct a Gaussian mixture hidden Markov model, obtain the optimal number of Gaussian components through the Dirichlet process, design an estimation network to estimate the mixed membership and state transition probability of each hidden state in the Gaussian mixture hidden Markov model corresponding to the observation sequence, calculate the parameters of the Gaussian mixture hidden Markov model according to the predicted values ​​of the mixed membership and state transition probability, and then calculate the energy of the low-dimensional representation of the sample.

[0082] The Gaussian mixture hidden Markov model (GMM-HMM) is used to model the time series data to capture the changes in the working condition characteristics in the time series. Assuming that the observation sequence length, that is, the sample length is L, the number of hidden states is M, and the number of Gaussian distributions is J, the parameter set δ of the model is expressed as:

[0083] δ={τ,A,π,μ,Σ}

[0084] Among them, τ is the initial state probability matrix of the model, A represents the state transition probability matrix, π, μ, Σ are the mixture probability matrix, mean matrix and covariance matrix of the Gaussian mixture model respectively.

[0085] Assume that the sample has a low-dimensional representation Z i The corresponding observation sequence is composed of multiple hidden states s m Generated, each hidden state corresponds to a different operating state of the device, and the observation sequence obeys the Gaussian mixture distribution under each hidden state:

[0086]

[0087] Among them, p(Z i |s m ) represents the low-dimensional representation Z of the sample i The corresponding observation sequence in the hidden state s m is a Gaussian mixture distribution, J is the number of components in the Gaussian mixture distribution, and πj is the weight of the jth Gaussian component, satisfying Indicates that the j-th mean is μ j , the covariance is Σ j Gaussian distribution.

[0088] In order to adaptively obtain the number of Gaussian components J, this embodiment introduces the Dirichlet process, a non-parametric Bayesian method. Assuming that G is a distribution that satisfies the Dirichlet process, then:

[0089] G~DP(α,G0)

[0090] Where α is the concentration parameter and G0 is a Beta distribution used to generate the mean μ of each Gaussian component. j and covariance Σ j .

[0091] The implementation of the Dirichlet process can be represented by the Stick-Breaking Construction process, where each weight π j It can be defined in the following ways:

[0092]

[0093] Among them, v i represents the used weight of the i-th Gaussian component among the first j-1 Gaussian components. This formula indicates that the weight of the j-th component is drawn from the Beta distribution for the jth time. j , and the unused weights of the first j-1 components (i.e. 1-v i ) is determined. Through this process, we continuously sample from the Beta distribution and assign the remaining weights to the next Gaussian component until all weights are assigned and satisfy

[0094] When adaptively determining the number of Gaussian components J, we filter out components with weights less than the threshold ε by setting a threshold ε, and count all weights π j The number of components in the equation that is greater than the threshold ε, the specific process is as follows:

[0095]

[0096] in is the indicator function, when π j >ε, it is 1, otherwise it is 0.

[0097] As an alternative method to estimate model parameters by alternately maximizing expectation, this embodiment first designs two multi-layer stacked estimation networks to predict some model parameters. Specifically, given the model parameters and the sample low-dimensional representation Z = {Z1, Z2, ... ZN}, a multi-layer neural network (MLN) is used to estimate the transition probability matrix of the hidden state and the mixed membership of each sample under a given hidden state:

[0098]

[0099] Among them, θ A and θ B They correspond to the weight parameters of two multi-layer neural networks, represents the estimate of the state transition probability matrix, represents the estimation of the sample mixture membership matrix, MLN(·) and softmax(·) represent the parameterized multilayer neural network and activation function, respectively.

[0100] Secondly, according to the income and Further estimate the parameters of the model in, They are the estimates of the initial state probability matrix, state transition probability matrix, mixing probability matrix, mean matrix and covariance matrix respectively.

[0101] In the process of establishing the Gaussian mixture hidden Markov model, the first-order Markov assumption and the observation independence assumption are obeyed, and the joint probability distribution of the model is obtained as follows:

[0102]

[0103] Among them, P(Q N ,O1,…O N ) represents a given observation sequence O1,…O N and the hidden state sequence Q N The joint probability distribution of the Gaussian mixture hidden Markov model, P(Q i |Q i-1 ) is from state Q i-1 Transfer to state Q i The transition probability, P(O i |Q i ) is a given state Q i Next observation sequence O i The generation probability of , P(Q1) is the initial state distribution.

[0104] After obtaining the joint probability distribution of the model, we solve P(Q 1:N |O 1:N ), that is, given the observation sequence O 1:N (O1,…O N ) under the condition that Q 1:N (Q1,…QN ). Variational inference replaces the true posterior distribution P(Q 1:N |O 1:N ), and optimize this approximate posterior distribution by maximizing the evidence lower bound (ELBO), the specific process is as follows:

[0105]

[0106] Where P(Z,π,Q) represents the joint probability distribution of the low-dimensional representation Z, the mixing coefficient π, and the hidden state Q. q(π,Q) represents the variational distribution used to approximate the posterior distribution. is the expected value of the joint probability, which represents the joint log-likelihood of the observed data and the hidden state, It is the entropy term of the variational distribution, ensuring that the approximate distribution is not too concentrated. By maximizing ELBO, the variational distribution can be made as close to the posterior distribution as possible, thereby achieving effective inference of the model parameters.

[0107] On this basis, we get the input sample X i ={x i (1),x i (2),…x i (l)} In the given model The conditional probability under for:

[0108]

[0109] Among them, α n,m is the forward probability, indicating that given the model parameters The observation sequence O1, O2, ..., O m And the hidden state is Q m = the probability of n, β n,m is the backward probability, indicating that given the model parameters The hidden state at time m is Q m =nThe following observation sequence is O m+1 ,O m+2 ,…,O M The probability of is the estimate of the forward probability, is the estimated value of the backward probability. n,m and β n,m The calculation formula is as follows:

[0110]

[0111] in, For a given state Q m =j time Q m+1= m-1 estimated value of state transition probability, For a given hidden state Q m =j and Q m+1 = n when the estimated state transition probability is For a given state Q m =n when the observation sequence O m = the observed probability estimate of Z, For a given state Q m =j observation sequence O m = estimated value of state transition probability of n; For a given state Q m =j observation sequence O m = the observed probability estimate of Z; For a given hidden state Q m =j time Q m+1 =The observed probability estimate of m+1.

[0112] According to the estimated parameters and The energy of the observation sequence is defined as:

[0113]

[0114] in, It means that given a sample sequence O, about the parameter The likelihood function of The parameter set of the Gaussian mixture hidden Markov model representing the given estimate The probability that the subsequent observation sequence is O.

[0115] Step S4, construct a joint optimization framework loss function including sample reconstruction error, energy corresponding to the sample low-dimensional representation, penalty term and the lower bound of the evidence of variational Bayesian inference, and update the parameters of the wavelet kernel attention convolution layer, convolutional autoencoder and network estimation until convergence by minimizing the joint optimization framework loss function.

[0116] This embodiment designs the evidence lower bound of variational Bayesian inference as the loss function term, and constructs a joint optimization framework of a compression network based on wavelet kernel attention convolutional autoencoder and an estimation network based on a Gaussian mixture hidden Markov model, inducing representation learning of variable operating condition data in the compression network and density estimation of the estimation network in an end-to-end manner.

[0117] Constructed joint optimization framework loss function as follows:

[0118]

[0119] Where N is the number of input samples, λ1, λ2 and λ3 are weighting coefficients, Dis(Xi ,X i ′) is the reconstruction error between the i-th input sample and the i-th reconstructed input sample, E(χ) is the energy corresponding to the low-dimensional representation of the sample, Represents the penalty item; the specific expression is as follows:

[0120]

[0121] Among them, Δ j,m,v,v Represents the values ​​on the diagonal of the penalty matrix, and V represents the dimension of the low-dimensional representation of the sample.

[0122] Then, the Adam optimizer is used to minimize the optimization objective function Implement parameter update.

[0123] In the fault detection stage, the newly collected current signal is passed through the wavelet kernel attention convolution layer and the convolutional autoencoder to obtain a low-dimensional representation of the sample to be detected, and the energy obtained by inputting the low-dimensional representation of the sample to be detected into the Gaussian mixture hidden Markov model (the parameters of the model are estimated based on the trained estimation network) is compared with the threshold (the 99th percentile of all training samples is taken as the threshold) to obtain the equipment fault detection result.

[0124] In a specific embodiment, a multi-operating condition data set is collected on an existing electromechanical fault simulation platform for verification, wherein the experimental results of fault detection under unknown operating conditions are shown in Table 1.

[0125] In the multi-condition data set, three speed conditions and two load conditions are set. The speed conditions and load conditions are combined to obtain a total of six operating conditions. 150 normal samples and 150 compound fault samples under the six operating conditions are randomly selected as the test set, and the remaining normal samples are used for training, and the initial parameters of the experiment are set.

[0126] The area under the ROC curve (AUC-ROC), the area under the PR curve (AUC-PR) and the F1 score are used to evaluate the proposed method and related comparative methods (including Deep SVDD (deep single-class support vector data description), NeuTral AD (neural conversion analysis anomaly detection), SLAD (semi-supervised learning anomaly detection)). As can be seen from Table 1, the proposed method obtains a higher score, with AUC-ROC reaching 97.00, AUC-PR reaching 93.02, and F1 score reaching 96.39. Deep SVDD performs poorly in complex scenes that do not conform to geometric boundaries. NeuTral AD's reliance on learnable data transformation makes it powerful in a series of anomaly detection tasks, but it is not effective in dealing with complex anomaly patterns. SLAD introduces scale as a new data feature, but when faced with complex anomaly patterns, scale differences cannot well represent anomalies. In summary, compared with the existing anomaly detection methods, the proposed method has improved the comprehensive performance by about 20%, which verifies the effectiveness and superiority of the proposed method.

[0127] Table 1 Experimental results (combined fault of gear wear + blade misalignment)

[0128] method AUC-ROC AUC-PR F1 score Deep SVDD 87.32％ 83.62％ 72.62％ NeuTral AD 94.14％ 90.59％ 82.14％ SLAD 87.29％ 87.29％ 63.10％ Method of the present invention 97.00％ 93.02％ 96.39％

[0129] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A spindle machining composite fault detection method based on a robust multi-condition process model, characterized in that: The spindle machining composite fault detection method based on a robust multi-condition process model comprises: Collect and pre-process the current signal of the spindle processing equipment under normal operating conditions; The preprocessed current signal passes through the wavelet kernel attention convolution layer and the convolutional autoencoder in sequence. The wavelet kernel attention convolution layer first uses a variety of wavelet basis functions to extract multi-scale interpretable features of the preprocessed current signal, and then uses the softmax function to process the interpretable features to obtain the sample's significant features. The convolutional autoencoder composed of 1-D convolutional layers obtains a low-dimensional representation of the sample containing multiple wavelet decomposition results based on the sample's significant features. Construct a Gaussian mixture hidden Markov model, obtain the optimal number of Gaussian components through the Dirichlet process, use the estimation network to estimate the mixture membership and state transition probability of each hidden state in the Gaussian mixture hidden Markov model corresponding to the observation sequence, calculate the parameters of the Gaussian mixture hidden Markov model based on the estimated values ​​of the mixture membership and state transition probability, and then calculate the energy corresponding to the low-dimensional representation of the sample based on the Gaussian mixture hidden Markov model; Construct a joint optimization framework loss function including sample reconstruction error, energy corresponding to the sample low-dimensional representation, penalty term and the lower bound of the evidence of variational Bayesian inference, and update the parameters of the wavelet kernel attention convolution layer, convolutional autoencoder and estimation network until convergence by minimizing the joint optimization framework loss function; In the fault detection stage, the newly collected current signal is passed through the wavelet kernel attention convolution layer and the convolution autoencoder to obtain a low-dimensional representation of the sample to be detected, and the low-dimensional representation of the sample to be detected is input into the Gaussian mixture hidden Markov model to obtain energy. The energy corresponding to the low-dimensional representation of the sample to be detected is compared with the threshold to obtain the composite fault detection result of the spindle machining equipment. The threshold is the 99th percentile of all training samples.

2. The spindle machining composite fault detection method based on a robust multi-condition process model according to claim 1 is characterized in that: The wavelet kernel attention convolution layer first uses a variety of wavelet basis functions to extract multi-scale interpretable features of the preprocessed current signal, and then uses the softmax function to process the interpretable features to obtain sample salient features, including: Take the preprocessed current signal X i ={x i (1),x i (2),…x i (L)} and various wavelet bases φ q (·), X i represents the i-th preprocessed current signal, i.e., the i-th input sample, x i (l) represents the input sample X i For the lth data in , the similarity coefficient between the input sample and the wavelet kernel function is expressed as: in, It means that the qth type for the lth data depends on two adjustable parameters σ w and u w The wavelet kernel function, σ w is the scaling factor, u w is the shift factor, q=1,…,Q, Q represents the total number of selected wavelet bases, w represents the number of channels, ρ i,w,q (l) represents the input sample X i With wavelet kernel function The similarity coefficient between i (m) represents the input sample X i The mth data in, m is the index and m=1,…,L, Represents the wavelet kernel function The complex conjugate of It means that the qth type for the lmth data depends on two adjustable parameters σ w and u w Wavelet kernel function, N represents the number of input samples, L represents the length of each input sample; Select wavelet kernel function Extracting interpretable features from an input sample of length L is Then the interpretable features extracted from the input samples by various wavelet kernel functions are expressed as: Among them, S i represents the interpretable features extracted from the i-th input sample, s i,w,q represents the interpretable features of the ith input sample extracted by the qth wavelet kernel function in the wth channel, Indicates i,w,q The transpose of , W represents the total number of convolution channels, w = 1, ..., W, q = 1, ..., Q; Then the softmax function is used to process the interpretable features to obtain the significant features of the sample The formula is: Among them, q * is the index of the wavelet kernel function that extracts the significant features of the sample in the channel among the Q wavelet kernel functions, s i,w,q* Indicates the qth * The interpretable features of the ith input sample extracted by the wavelet kernel function in the wth channel, express The transpose of .

3. The spindle machining composite fault detection method based on a robust multi-condition process model according to claim 1 is characterized in that: The convolutional autoencoder composed of 1-D convolutional layers obtains a low-dimensional representation of the sample containing multiple wavelet decomposition results according to the sample's significant features, including: Among them, Z i represents the low-dimensional representation of the sample of the i-th input sample, represents the sample salient features of the i-th input sample, Θ e denotes the weight of a 1-D convolutional layer, and Conv1d(;) denotes a 1-D convolutional layer.

4. The spindle machining composite fault detection method based on a robust multi-condition process model according to claim 1 is characterized in that: The method of obtaining the optimal number of Gaussian components by using the Dirichlet process includes: Let G be a distribution that satisfies the Dirichlet process, then: G~DP(α,G0) Where α is the concentration parameter and G0 is a Beta distribution used to generate the mean μ of each Gaussian component. j and covariance Σ j ; The weight of the jth Gaussian component is π j The weight v drawn from the Beta distribution for the jth time j , and the unused weights 1-v of the first j-1 Gaussian components i OK, the formula is as follows: Among them, v i represents the used weight of the i-th Gaussian component among the first j-1 Gaussian components; Continue to sample from the Beta distribution and assign the remaining weights to the next Gaussian component until all weights are assigned and satisfy When adaptively determining the number of Gaussian components J, we filter out Gaussian components with weights less than the threshold ε by setting a threshold ε, and count all weights π j The number of Gaussian components greater than the threshold ε, the specific process is as follows: in, is the indicator function, when π j It is 1 when ≥ε, otherwise it is 0.

5. The spindle machining composite fault detection method based on a robust multi-condition process model according to claim 1 is characterized in that: The method of estimating the mixed membership and state transition probability of each hidden state corresponding to the observation sequence in the Gaussian mixture hidden Markov model by using the estimation network includes: Take two multi-layer neural networks as estimation networks, and use the multi-layer neural network to estimate the state transition probability matrix of the hidden state and the mixed membership matrix of each sample under the given hidden state respectively: Among them, θ A and θ B They correspond to the weight parameters of two multi-layer neural networks, represents the estimate of the state transition probability matrix, represents the estimation of the mixture membership matrix, MLN(·) and softmax(·) represent the multi-layer neural network and activation function respectively, M represents the number of hidden states, and J represents the number of Gaussian components.

6. The spindle machining composite fault detection method based on a robust multi-condition process model according to claim 1 is characterized in that: The construction includes a joint optimization framework loss function of sample reconstruction error, energy corresponding to the sample low-dimensional representation, penalty term and the lower bound of the evidence of variational Bayesian inference, including: Constructing the joint optimization framework loss function as follows: Where N is the number of input samples, λ1, λ2 and λ3 are weighting coefficients, Dis(X i ,X i ′) is the reconstruction error between the i-th input sample and the i-th reconstructed input sample, E(χ) is the energy corresponding to the low-dimensional representation of the sample, represents the penalty term, The lower bound of the evidence for variational Bayesian inference with the Dirichlet process as the prior is calculated as follows: Among them, ELBO represents the lower bound of the evidence for variational Bayesian inference with Dirichlet process as prior, P(Z,π,Q) represents the joint probability distribution containing the sample low-dimensional representation Z, the mixing coefficient π and the hidden state Q, q(π,Q) represents the variational distribution used to approximate the posterior distribution, is the expected value of the joint probability distribution, is the entropy term of the variational distribution.

Images