A method for explaining fault detection of hydraulic turbines based on non-intrusive measurements

Abstract

The present invention belongs to the technical field of rotating machinery fault detection, and discloses a method for interpretable fault detection of a turbine based on non-invasive measurement, including obtaining a low-dimensional time-dependent feature representation and a reconstruction feature by using a compression network based on a wavelet kernel convolution autoencoder; obtaining a reconstructed input sample according to the reconstruction feature, and calculating the sample reconstruction error; using a separate estimation network to respectively predict a mixed membership matrix and a state transition probability matrix, estimating the parameters of a Gaussian mixture hidden Markov model according to the predicted values, and calculating the energy of the low-dimensional time-dependent feature representation; constructing a joint optimization loss function based on the sample reconstruction error, the energy of the low-dimensional time-dependent feature representation, and a penalty function, and updating the parameters of the compression network and the estimation network based on the joint optimization loss function for training. The present invention enhances the interpretability of the model, and improves the generalization ability and applicability of fault detection under variable working conditions.

Description

Technical Field

[0001] The invention belongs to the technical field of rotating machinery fault detection, and in particular relates to an interpretable fault detection method for a water turbine based on non-invasive measurement. Background Art

[0002] As the core equipment of a hydropower station, a turbine converts the gravitational potential energy of the water body into kinetic energy to drive the generator to rotate and generate electricity, playing the role of peak regulation, phase regulation and emergency standby in the power grid system. Since the turbine has been in service for a long time in a complex and harsh water flow environment, impurities such as silt in the water flow will inevitably cause erosion and wear on the blade surface, making the blade surface rough and reducing the power generation efficiency of the turbine. In addition, long-term wear on the blade surface will lead to a reduction in blade thickness, affect the structural strength of the blade, and cause the blade to deform, which will lead to turbine failure. Once a unit fails, it will either shut down, affecting the economic benefits of the power station, or cause huge property losses and casualties. Therefore, studying an effective and reliable turbine fault detection method is of great engineering significance for timely discovering potential turbine faults, ensuring the dynamic performance of the unit and maintaining the safety of the power grid.

[0003] Existing turbine fault detection methods can be roughly divided into signal processing-based methods, such as reference: Y.Xin, S. Li, Z. Zhang, Z. An and J. Wang, Adaptive Reinforced Empirical MorletWavelet Transform and Its Application in Fault Diagnosis of RotatingMachinery, inIEEE Access, vol. 7, pp. 65150-65162, 2019 and data-driven methods, such as reference: Wang H, Liu X, Ma L, et al. Anomaly detection for hydropower turbine unitbased on variational modal decomposition and deep autoencoder[J]. EnergyReports, 7: 938-946, 2021. Among them, signal processing-based methods can be divided into time domain, frequency domain and time-frequency domain methods. Due to the complexity of the service environment of the turbine and the interference of noise, it is difficult to effectively extract fault features and establish an accurate fault detection model based on signal processing methods. The data-driven method adaptively extracts fault information hidden in the measurement data to achieve effective detection of equipment faults. However, the data-driven method represented by deep learning methods has a "black box" characteristic due to its multi-layer nonlinear mapping, and the model lacks interpretability. On the other hand, intrusive measurements represented by vibration signals have defects such as inconsistent communication protocols, limited installation space, damage to equipment and affecting system stability. In addition, since the speed and load of the turbine are constantly changing during operation, the monitoring signal is no longer periodic, and it also brings non-stationary characteristics such as frequency modulation, amplitude modulation and phase modulation to the monitoring signal, so the fault detection method for constant working conditions is no longer applicable. Summary of the invention

[0004] The purpose of the present invention is to provide an interpretable fault detection method for a hydraulic turbine based on non-intrusive measurement, which enhances the interpretability of the model and improves the generalization ability and applicability of fault detection under variable operating conditions.

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

[0006] A method for detecting an interpretable fault of a hydraulic turbine based on non-intrusive measurement, the method comprising:

[0007] Obtain the original three-phase current monitoring time series of the turbine in the sample data set as an input sample;

[0008] The compression network based on wavelet kernel convolutional autoencoder is used to compress the original three-phase current monitoring time series under normal working conditions to obtain low-dimensional time series dependent feature representation and reconstruction features;

[0009] A reconstructed input sample is obtained according to the reconstruction feature, and a sample reconstruction error is obtained based on the input sample and the reconstructed input sample;

[0010] Construct a Gaussian mixture hidden Markov model, use separate estimation networks to predict the mixture membership matrix and the state transition probability matrix respectively, estimate the parameters of the Gaussian mixture hidden Markov model according to the predicted values ​​of the mixture membership matrix and the state transition probability matrix, and calculate the energy of the low-dimensional time-dependent feature representation based on the parameters of the Gaussian mixture hidden Markov model;

[0011] A joint optimization loss function is constructed based on sample reconstruction error, energy of low-dimensional time-dependent feature representation, and penalty function, and parameters of the compression network and the estimation network are updated based on the joint optimization loss function for training;

[0012] For the newly collected original three-phase current monitoring time series of the turbine, the corresponding energy is obtained through the trained compression network and estimation network, and compared with the threshold to obtain the turbine fault detection result.

[0013] 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.

[0014] Preferably, the compression network based on wavelet kernel convolution autoencoder is used to compress the original three-phase current monitoring time series under normal working conditions to obtain low-dimensional time series dependent feature representation and reconstruction features, including:

[0015] Given a sample dataset The original three-phase current monitoring time series of the hydro turbine , the feature extraction process after wavelet kernel convolution is expressed as:

[0016] ;

[0017] in, Indicates the first input samples, Represents the total number of input samples in the sample data set, Represents the input sample The sample features after wavelet kernel convolution are: represents the wavelet kernel convolution layer, Represents the weight parameters in the wavelet kernel convolution layer;

[0018] Sample features after wavelet kernel convolution After the 1-D convolution operation, it is mapped to a high-dimensional feature space, which is expressed as:

[0019] ;

[0020] in, represents a 1-D convolutional neural network, represents the weight parameters of the 1-D convolutional neural network, Represents 1-D convolutional neural network for sample features The extracted sample’s significant features;

[0021] For the sample's significant features , the low-dimensional temporal dependency feature representation and reconstructed feature representation obtained by the compression network are:

[0022] ;

[0023] ;

[0024] in, The compression network learns the significant features of samples in low-dimensional space The low-dimensional time-dependent feature representation of the low-dimensional time-dependent feature representation is recorded as , and are the weight parameters of the encoder and decoder in the compression network, Indicates the significant features of the sample The reconstruction features of represents the encoder, Represents a decoder.

[0025] Preferably, obtaining a reconstructed input sample according to the reconstruction feature, and obtaining a sample reconstruction error based on the input sample and the reconstructed input sample, comprises:

[0026] Reconstruction features of sample salient features , then the reconstruction of the input sample is expressed as:

[0027] ;

[0028] In the formula, represents a 1-D transposed convolutional neural network, represents the weight parameters of the 1-D transposed convolutional neural network, Represents the reconstructed input sample;

[0029] The sample reconstruction error corresponding to the original three-phase current monitoring time series is calculated as follows:

[0030] ;

[0031] In the formula, Indicates The sample reconstruction error corresponding to the input samples.

[0032] Preferably, the constructing of Gaussian mixture hidden Markov model comprises:

[0033] The operation process of the turbine under variable operating conditions is described as a double random process with a hidden Markov chain with hidden states and a random observation sequence obeying a Gaussian mixture distribution. The length of the observation sequence is , the number of hidden states is , the number of Gaussian distributions is , then the parameter set of the Gaussian mixture hidden Markov model is expressed as:

[0034] ;

[0035] in, represents the initial state probability matrix of the Gaussian mixture hidden Markov model, represents the state transition probability matrix of the Gaussian mixture hidden Markov model, represents the mixture probability matrix of the Gaussian mixture hidden Markov model, represents the mean matrix of the Gaussian mixture hidden Markov model, represents the covariance matrix of the Gaussian mixture hidden Markov model, Represents the set of parameters of the Gaussian mixture hidden Markov model.

[0036] Preferably, the estimation network is a multi-layer neural network, and the method of using a separate estimation network to respectively predict the mixed membership matrix and the state transition probability matrix includes:

[0037] ;

[0038] ;

[0039] in, and They correspond to the weight parameters of two multi-layer neural networks, Represents a given low-dimensional temporal dependency feature representation The predicted value of the state transition probability matrix under Represents a given low-dimensional temporal dependency feature representation The predicted value of the mixed membership matrix under and Represent multi-layer neural network and activation function respectively.

[0040] Preferably, the estimating parameters of the Gaussian mixture hidden Markov model according to the predicted values ​​of the mixed membership matrix and the state transition probability matrix comprises:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] In the formula, The hidden state is The estimated value of the initial state probability when express The hidden state is Time The predicted value of the mixed membership of a Gaussian distribution fit, , ,combination indivual Get the estimated value of the initial state probability matrix of the Gaussian mixture hidden Markov model , Indicates the hidden state from arrive The estimated value of the state transition probability when Represents the total number of input samples in the sample data set, express Hidden state from arrive The predicted value of the state transition probability when ,combination indivual Get the estimated value of the state transition probability matrix of the Gaussian mixture hidden Markov model , The hidden state is Time The estimated value of the mixture probability of Gaussian distribution fitting, combination indivual Get an estimate of the mixing probability matrix for the Gaussian mixture hidden Markov model , The hidden state is Time An estimate of the mean of a Gaussian distribution fit, Represents low-dimensional temporal dependency feature representation, combination indivual Get an estimate of the mean matrix of the Gaussian mixture hidden Markov model , The hidden state is Time An estimate of the covariance of the Gaussian distribution fits, is a diagonal matrix, Indicates transposition, combination indivual Get an estimate of the covariance matrix of the Gaussian mixture hidden Markov model , and finally the estimated parameter set of the Gaussian mixture hidden Markov model is .

[0047] Preferably, the energy of low-dimensional time-dependent feature representation is calculated based on the parameters of the Gaussian mixture hidden Markov model, including:

[0048] No. The low-dimensional temporal dependency feature representation corresponding to the input samples , given the estimated Gaussian mixture hidden Markov model parameter set The conditional probability under for:

[0049] ;

[0050] Then calculate the low-dimensional temporal dependency feature representation The energy is:

[0051] ;

[0052] in, Represents the set of parameters of the estimated Gaussian mixture hidden Markov model Next, to Observation sequence The probability of Represents a given hidden state Hidden state The estimated state transition probability, For a given hidden state Hidden state The estimated observation probability of It is a low-dimensional temporal dependency feature representation The energy Represents a given sample sequence , about the estimated parameter set of the Gaussian mixture hidden Markov model The likelihood function of The parameter set of the Gaussian mixture hidden Markov model representing the given estimate The subsequent observation sequence is probability.

[0053] Preferably, the joint optimization loss function is constructed based on the sample reconstruction error, the energy represented by the low-dimensional time-dependent feature, and the penalty function, including:

[0054] ;

[0055] in, represents the joint optimization loss function, represents the weight parameter to be optimized, Represents the total number of input samples in the sample data set, Indicates The sample reconstruction error corresponding to the input samples is Indicates the first input samples, Indicates The reconstructed input samples corresponding to the input samples are is the energy weighting coefficient, is the weight coefficient of the penalty function, For the The energy of the low-dimensional time-dependent feature representation corresponding to the input samples, is the penalty function, expressed as:

[0056] ;

[0057] In the formula, represents the estimate of the covariance matrix of the Gaussian mixture hidden Markov model, is the number of hidden states of the Gaussian mixture hidden Markov model, is the number of Gaussian distributions in the Gaussian mixture hidden Markov model, represents the dimensionality of the low-dimensional temporal-dependent feature representation generated by the compression network, The hidden state is Time An estimate of the covariance of the Gaussian distribution fits, is a diagonal matrix, Represents the diagonal matrix Line Covariance values ​​for the columns.

[0058] Preferably, the threshold is the 5th percentile of all samples in the sample data set. If the energy corresponding to the newly collected original three-phase current monitoring time series of the turbine is greater than the threshold, it is determined to be in normal operating conditions; otherwise, it is in abnormal operating conditions.

[0059] The present invention provides a hydraulic turbine explainable fault detection method based on non-intrusive measurement, which has the following beneficial effects compared with the prior art:

[0060] 1. 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.

[0061] 2. A wavelet kernel convolutional autoencoder is proposed to obtain a low-dimensional time-dependent feature representation of the input samples, thereby converting the high-dimensional fault detection task into a low-dimensional probability density estimation task and enhancing the interpretability of the model.

[0062] 3. The hidden Markov model is introduced to describe the complex variable operating condition process as a Markov process whose hidden state is an unknown operating condition and the observation sequence is only related to the operating condition, so as to improve the generalization ability and applicability of fault detection under variable operating conditions.

[0063] 4. A joint optimization loss function is designed to promote the parameter learning of wavelet kernel convolutional autoencoder and Gaussian mixture hidden Markov model, and induce representation learning in low-dimensional space and density estimation for unsupervised fault detection in an end-to-end manner, avoiding the need for pre-training. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The present invention is a flow chart of a method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement. DETAILED DESCRIPTION

[0065] 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.

[0066] 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.

[0067] In view of the problems that it is difficult to accurately model the operation process of a hydraulic turbine under existing complex and variable working conditions, the process fault detection performance is easily affected by the change of working conditions, and the model interpretability is poor, this embodiment proposes a hydraulic turbine interpretable fault detection method based on non-intrusive measurement, such as Figure 1 As shown, the following steps are included:

[0068] Step S1, obtaining the original three-phase current monitoring time series of the turbine in the sample data set as an input sample. Under normal working conditions of the turbine, the original three-phase current monitoring time series of the device is obtained by using the Hall sensor installed in the stator or rotor of the turbine.

[0069] Step S2, using a compression network based on a wavelet kernel convolutional autoencoder, compress the original three-phase current monitoring time series under normal operating conditions, remove redundant information, obtain low-dimensional time series dependent feature representation and reconstruction features, thereby converting the turbine fault detection problem in a high-dimensional space into a probability density estimation problem in a potential low-dimensional space.

[0070] This embodiment uses a compression network based on a wavelet kernel convolutional autoencoder to extract timing-dependent features from the original three-phase current monitoring time series of the turbine, and then uses a one-dimensional convolutional layer to project the features into a low-dimensional space. The compression network based on the wavelet kernel convolutional autoencoder includes a CWConv (continuouswavelet convolutional) layer composed of Laplace wavelets, two 1-D convolutional layers, and six linear layers. The number of output channels of the CWConv layer and the two 1-D convolutional layers is (12, 24, 24), and the ReLU activation function is used as the nonlinear activation function, and the maximum pooling technique is used for downsampling. In addition, Dropout is used to prevent overfitting of the training data. In this embodiment, in the encoder network, a continuous wavelet convolution CWConv layer is used to replace the first convolutional layer, so as to extract basic semantic information with clear physical meaning from high-dimensional data, thereby enhancing the interpretability of the mapping.

[0071] Specifically, we first extract the sample salient features reflecting the complex operating conditions of the equipment through wavelet kernel convolution. Specifically, given a sample data set The original three-phase current monitoring time series of the hydro turbine , the feature extraction process after wavelet kernel convolution is expressed as:

[0072] ;

[0073] in, Indicates the first input samples, Represents the total number of input samples in the sample data set, Represents the input sample The sample features after wavelet kernel convolution are: represents the wavelet kernel convolution layer, Represents the weight parameters in the wavelet kernel convolution layer, which is determined by the scale parameter and time-shift parameters In wavelet kernel convolution, the input sample Using a parameterized wavelet dictionary Perform convolution.

[0074] Sample features after wavelet kernel convolution After the 1-D convolution operation, it is mapped to a high-dimensional feature space, which is expressed as:

[0075] ;

[0076] in, represents a 1-D convolutional neural network, represents the weight parameters of the 1-D convolutional neural network, Represents 1-D convolutional neural network for sample features The extracted sample salient features.

[0077] The second is the generation of low-dimensional representations of sample salient features and feature reconstruction. This embodiment uses an autoencoder stacked with multiple fully connected layers to generate a low-dimensional representation of the salient features of the input sample. This embodiment specifically uses an autoencoder composed of 6 linear layers. The encoder part only learns the salient features in the input information to compress the information from the high-dimensional space, while the decoder part reconstructs the input information from the learned feature space.

[0078] For the sample's significant features , the low-dimensional temporal dependency feature representation and reconstructed feature representation obtained by the compression network are:

[0079] ;

[0080] ;

[0081] in, The compression network learns the significant features of samples in low-dimensional space The low-dimensional time-dependent feature representation of the low-dimensional time-dependent feature representation is recorded as , and are the weight parameters of the encoder and decoder in the compression network, Indicates the significant features of the sample The reconstruction features of represents the encoder, Represents a decoder.

[0082] Step S3: obtaining a reconstructed input sample according to the reconstruction feature, and obtaining a sample reconstruction error based on the input sample and the reconstructed input sample.

[0083] This embodiment uses a 1-D transposed convolutional network to reconstruct the input sample data, that is, upsampling. , then the reconstruction of the input sample is expressed as:

[0084] ;

[0085] In the formula, represents a 1-D transposed convolutional neural network, represents the weight parameters of the 1-D transposed convolutional neural network, Represents the reconstructed input sample.

[0086] The sample reconstruction error corresponding to the original three-phase current monitoring time series is calculated as follows:

[0087] ;

[0088] In the formula, Indicates The sample reconstruction error corresponding to the input samples is measured using the Euclidean distance in this embodiment.

[0089] Step S4, constructing a Gaussian mixture hidden Markov model (GMHMM), using a separate estimation network to predict the mixture membership matrix and the state transition probability matrix respectively, estimating the parameters of the Gaussian mixture hidden Markov model according to the predicted values ​​of the mixture membership matrix and the state transition probability matrix, and calculating the energy of the low-dimensional time-dependent feature representation based on the parameters of the Gaussian mixture hidden Markov model.

[0090] Establishment of long-term continuous observation sequence model. Given the low-dimensional time-dependent feature representation of input samples under complex working conditions, it is considered to be modeled as a dual random process of a hidden Markov chain (HMM) with hidden states and a random observation sequence obeying a Gaussian mixture distribution (GMM).

[0091] Assume that the length of the observation sequence is , the number of hidden states is , the number of Gaussian distributions is , then the parameter set of the Gaussian mixture hidden Markov model is expressed as:

[0092] ;

[0093] in, represents the initial state probability matrix of the Gaussian mixture hidden Markov model, represents the state transition probability matrix of the Gaussian mixture hidden Markov model, represents the mixture probability matrix of the Gaussian mixture hidden Markov model, represents the mean matrix of the Gaussian mixture hidden Markov model, represents the covariance matrix of the Gaussian mixture hidden Markov model, Represents the set of parameters of the Gaussian mixture hidden Markov model.

[0094] Model parameter learning for long-term continuous observation sequences: As an alternative method to alternately estimate model parameters by expectation maximization, we first design two multi-layer stacked estimation networks and use the estimation network to replace the E step of the Expectation Maximization (EM) algorithm to predict some model parameters.

[0095] Specifically, given the model parameters and low-dimensional temporal dependency feature representation , the estimation network uses a multi-layer neural network (MLN) to predict the transition probability matrix of the hidden state and the mixed membership matrix of each sample under the given hidden state:

[0096] ;

[0097] ;

[0098] in, and They correspond to the weight parameters of two multi-layer neural networks, Represents a given low-dimensional temporal dependency feature representation The predicted value of the state transition probability matrix under Represents a given low-dimensional temporal dependency feature representation The predicted value of the mixed membership matrix under and They represent parameterized multi-layer neural networks and activation functions respectively.

[0099] Furthermore, the parameters of the Gaussian mixture hidden Markov model are estimated according to the predicted values ​​of the mixture membership matrix and the state transition probability matrix, including:

[0100] ;

[0101] ;

[0102] ;

[0103] ;

[0104] ;

[0105] In the formula, The hidden state is The estimated value of the initial state probability when express The hidden state is Time The predicted value of the mixed membership of a Gaussian distribution fit, , ,combination indivual Get the estimated value of the initial state probability matrix of the Gaussian mixture hidden Markov model , Indicates the hidden state from arrive The estimated value of the state transition probability when Represents the total number of input samples in the sample data set, express Hidden state from arrive The predicted value of the state transition probability when ,combination indivual Get the estimated value of the state transition probability matrix of the Gaussian mixture hidden Markov model , The hidden state is Time The estimated value of the mixture probability of Gaussian distribution fitting, combination indivual Get an estimate of the mixing probability matrix for the Gaussian mixture hidden Markov model , The hidden state is Time An estimate of the mean of a Gaussian distribution fit, Represents low-dimensional temporal dependency feature representation, combination indivual Get an estimate of the mean matrix of the Gaussian mixture hidden Markov model , The hidden state is Time An estimate of the covariance of the Gaussian distribution fits, is a diagonal matrix, Indicates transposition, combination indivual Get an estimate of the covariance matrix of the Gaussian mixture hidden Markov model .

[0106] The parameter set of the estimated Gaussian mixture hidden Markov model is .

[0107] On this basis, in order to further solve the low-dimensional time-dependent feature representation Given an estimated Gaussian mixture hidden Markov model parameter set The conditional probability of , define and calculate the forward probability and backward probability as follows:

[0108] ;

[0109] ;

[0110] In the formula, is the forward probability, representing the parameter set of the Gaussian mixture hidden Markov model given the estimate , to Observation sequence And the hidden state is The probability of is the backward probability, representing the parameter set of the Gaussian mixture hidden Markov model given the estimate , No. The hidden state of an observation sequence is The following observation sequence is probability.

[0111] Combining the first-order Markov and observation independence assumptions, the forward probability satisfies the following formula:

[0112] ;

[0113] ;

[0114] in, is the forward probability, representing the parameter set of the Gaussian mixture hidden Markov model given the estimate , to the first observation sequence And the hidden state is The probability of The parameter set of the Gaussian mixture hidden Markov model representing the given estimate , No. The hidden state of an observation sequence is The following observation sequence is The probability of The parameter set of the Gaussian mixture hidden Markov model representing the given estimate The hidden state is probability. For a given state Time observation sequence The estimated observation probability of The hidden state is The estimated value of the initial state probability when The parameter set of the Gaussian mixture hidden Markov model representing the given estimate When Observation sequence , and hidden state , The probability of The parameter set of the Gaussian mixture hidden Markov model representing the given estimate and to Observation sequence And hidden state Next Observation sequence and its hidden state is The probability of The parameter set of the Gaussian mixture hidden Markov model representing the given estimate and hidden state Next observation sequence and its hidden state The probability of For a given state hour The estimated state transition probability, For a given hidden state and The estimated value of the state transition probability when .

[0115] Similarly, combining the first-order Markov and observation independence assumptions, the backward probability satisfies the following formula:

[0116] ;

[0117] ;

[0118] in, is the backward probability, representing the parameter set of the Gaussian mixture hidden Markov model given the estimate , the hidden state is The following observation sequence is probability; For a given state Time observation sequence The estimated value of the state transition probability; For a given state Time observation sequence The observed probability estimate of ; For a given hidden state hour The observed probability estimate of ; thus, the initial value and recursive form of the backward probability can be obtained.

[0119] Further, The low-dimensional temporal dependency feature representation corresponding to the input samples , given the estimated Gaussian mixture hidden Markov model parameter set The conditional probability under for:

[0120] ;

[0121] Then calculate the low-dimensional temporal dependency feature representation The energy is:

[0122] ;

[0123] in, Represents the set of parameters of the estimated Gaussian mixture hidden Markov model Next, to Observation sequence The probability of Represents a given hidden state Hidden state The estimated state transition probability, For a given hidden state Hidden state The estimated observation probability of It is a low-dimensional temporal dependency feature representation The energy Represents a given sample sequence , about the estimated parameter set of the Gaussian mixture hidden Markov model The likelihood function of The parameter set of the Gaussian mixture hidden Markov model representing the given estimate The subsequent observation sequence is When the turbine is operating under normal operating conditions, the observation probability is usually stable and large. If the system is attacked or fails, the distribution of observations will be different, resulting in a lower probability of observation. By setting the percentile of the energy of all training samples as the fault detection threshold, samples above this threshold are considered to be operating under abnormal conditions.

[0124] Step S5: construct a joint optimization loss function based on the sample reconstruction error, the energy represented by the low-dimensional time-dependent feature, and the penalty function, and update the parameters of the compression network and the estimation network for training based on the joint optimization loss function.

[0125] In this embodiment, when optimizing the network model induced by the knowledge constraint of the sequence model, a joint optimization objective function is designed to induce the optimal combination of the compression network and the estimation network under the supervision of the training density estimation task. The sample data set of samples, the constructed joint optimization objective function is as follows:

[0126] ;

[0127] in, represents the joint optimization loss function, represents the weight parameter to be optimized, , Represents the total number of input samples in the sample data set, Indicates The sample reconstruction error corresponding to the input samples is Indicates the first input samples, Indicates The reconstructed input samples corresponding to the input samples are is the energy weighting coefficient, is the weight coefficient of the penalty function, For the The energy of the low-dimensional time-dependent feature representation corresponding to the input samples, is the penalty function.

[0128] It is not difficult to find that the objective function is composed of the reconstruction error , the energy of low-dimensional time-dependent feature representation and penalty function It consists of three parts. Among them, It is the distance measure between the input sample and the reconstructed sample output by the compression network, which is used to characterize the sample reconstruction error introduced by the deep convolutional autoencoder compression network; secondly, is the energy of the low-dimensional time-dependent feature representation defined, which is used to model the possibility of observing a sequence under given parameters; in addition, considering that the Gaussian mixture model is prone to triggering a trivial solution when the diagonal elements of the covariance matrix are 0, a penalty function is introduced To avoid the singularity problem, the penalty function is expressed as:

[0129] ;

[0130] In the formula, represents the estimate of the covariance matrix of the Gaussian mixture hidden Markov model, is the number of hidden states of the Gaussian mixture hidden Markov model, is the number of Gaussian distributions in the Gaussian mixture hidden Markov model, represents the dimensionality of the low-dimensional temporal-dependent feature representation generated by the compression network, The hidden state is Time An estimate of the covariance of the Gaussian distribution fits, is a diagonal matrix, Represents the diagonal matrix Line Covariance values ​​for the columns.

[0131] The Adam optimizer is used to automatically update the model parameters of the compression network and the estimation network until convergence, thus completing the entire model training.

[0132] Step S6: for the newly collected original three-phase current monitoring time series of the turbine, the corresponding energy is obtained through the trained compression network and estimation network, and compared with the threshold value to obtain the turbine fault detection result.

[0133] In the fault detection stage, the newly collected three-phase current monitoring time series is input into the compression network to obtain the low-dimensional time series dependency feature representation of the new sample, and the low-dimensional time series dependency feature representation is input into the trained Gaussian mixture hidden Markov model to calculate the energy of the turbine three-phase current monitoring time series, and then compared with the 5th percentile (threshold) of all training samples to obtain the equipment fault detection result, that is, if the energy corresponding to the original three-phase current monitoring time series of the newly collected turbine is greater than the threshold, it is judged to be in normal working condition; otherwise, it is in abnormal working condition.

[0134] It was further verified using a rotating mechanical platform in the laboratory.

[0135] In order to simulate the variable working condition rotating machinery movement process in the physical world, processing tasks with different speeds were carried out on the spindle rotating machinery platform. During the experiment, three speeds (1800 rpm, 2400 rpm, 3000 rpm) were simulated. Under the condition of sampling frequency of 20khz, each working condition was collected for 8 seconds, with a total of 160,000 data. In addition, the experimental device simulated three types of bearing faults, namely dynamic balance (BD), blade crack (Crack), and blade bending (Bent).

[0136] In terms of data preprocessing, the original three-phase current time series collected under three normal working conditions are first spliced ​​together in sequence. Then, the spliced ​​data are processed using the sliding window method, and the window size and step size are set to 1600 to ensure that the data segments do not overlap. In this way, 300 sample data sets under normal working conditions can be obtained. The fault data are processed in the same way. Subsequently, half of the 300 normal sample data sets are selected as training sets, and the remaining normal sample and fault sample data sets are used as test sets.

[0137] The area under the ROC curve (AUC-ROC), the area under the PR curve (AUC-PR) and the F1 score were further used for evaluation. The experimental results are shown in Tables 1 to 3.

[0138] Table 1 Experimental results under dynamic balance failure

[0139] method AUC-ROC AUC-PR F1 DeepSVDD 0.6669 0.6662 0.6250 GOAD 0.5664 0.5425 0.5583 NeuTral 0.4963 0.5060 0.4167 WCAE-GMHMM 0.7083 0.6759 0.5882

[0140] Table 2 Experimental results under blade crack fault

[0141] method AUC-ROC AUC-PR F1 DeepSVDD 0.6146 0.4794 0.5417 GOAD 0.5000 0.4444 0.6154 NeuTral 0.6523 0.6268 0.5583 WCAE-GMHMM 0.7708 0.7454 0.7027

[0142] Table 3 Experimental results under blade bending fault

[0143] method AUC-ROC AUC-PR F1 DeepSVDD 0.6206 0.6123 0.5667 GOAD 0.6422 0.5123 0.5560 NeuTral 0.6179 0.5931 0.5750 WCAE-GMHMM 1.0000 1.0000 1.0000

[0144] According to the experimental results, the WCAE-GMHMM method of this application obtained a high evaluation score in the experiment. In contrast, the existing deep support vector data description method (DeepSVDD) assumes that normal data points are concentrated in a compact area, and tries to find a minimum spherical boundary to surround it, but performs poorly in this complex scenario. The existing graphic observer-based alarm diagnosis (GOAD) can intuitively display the system's operating conditions and possible faults in a graphical manner by monitoring the system's state variables, input and output information, but in a dynamically changing system, the adaptability to new working conditions is weak, resulting in poor performance. The existing neural network-based method (NeuTral) can capture the relationship between feature subsets, but cannot accurately capture the relationship between features in complex time series.

[0145] 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.

[0146] 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 method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement, characterized in that: The water turbine explainable fault detection method based on non-intrusive measurement includes: Obtain the original three-phase current monitoring time series of the turbine in the sample data set as an input sample; The compression network based on wavelet kernel convolutional autoencoder is used to compress the original three-phase current monitoring time series under normal working conditions to obtain low-dimensional time series dependent feature representation and reconstruction features; A reconstructed input sample is obtained according to the reconstruction feature, and a sample reconstruction error is obtained based on the input sample and the reconstructed input sample; Construct a Gaussian mixture hidden Markov model, use separate estimation networks to predict the mixture membership matrix and the state transition probability matrix respectively, estimate the parameters of the Gaussian mixture hidden Markov model according to the predicted values ​​of the mixture membership matrix and the state transition probability matrix, and calculate the energy of the low-dimensional time-dependent feature representation based on the parameters of the Gaussian mixture hidden Markov model; A joint optimization loss function is constructed based on sample reconstruction error, energy of low-dimensional time-dependent feature representation, and penalty function, and parameters of the compression network and the estimation network are updated based on the joint optimization loss function for training; For the newly collected original three-phase current monitoring time series of the turbine, the corresponding energy is obtained through the trained compression network and estimation network, and compared with the threshold to obtain the turbine fault detection result.

2. The method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement according to claim 1, characterized in that: The compression network based on the wavelet kernel convolution autoencoder is used to compress the original three-phase current monitoring time series under normal working conditions to obtain low-dimensional time series dependent feature representation and reconstruction features, including: Given a sample dataset The original three-phase current monitoring time series of the hydro turbine , the feature extraction process after wavelet kernel convolution is expressed as: ； in, Indicates the first input samples, Represents the total number of input samples in the sample data set, Represents the input sample The sample features after wavelet kernel convolution are: represents the wavelet kernel convolution layer, Represents the weight parameters in the wavelet kernel convolution layer; Sample features after wavelet kernel convolution After the 1-D convolution operation, it is mapped to a high-dimensional feature space, which is expressed as: ； in, represents a 1-D convolutional neural network, represents the weight parameters of the 1-D convolutional neural network, Represents 1-D convolutional neural network for sample features The extracted sample’s significant features; For the sample's significant features , the low-dimensional temporal dependency feature representation and reconstructed feature representation obtained by the compression network are: ； ； in, The compression network learns the significant features of samples in low-dimensional space The low-dimensional time-dependent feature representation of the low-dimensional time-dependent feature representation is recorded as , and are the weight parameters of the encoder and decoder in the compression network, Indicates the significant features of the sample The reconstruction features of represents the encoder, Represents a decoder.

3. The method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement according to claim 2, characterized in that: The step of obtaining a reconstructed input sample according to the reconstruction feature, and obtaining a sample reconstruction error based on the input sample and the reconstructed input sample, includes: Reconstruction features of sample salient features , then the reconstruction of the input sample is expressed as: ； In the formula, represents a 1-D transposed convolutional neural network, represents the weight parameters of the 1-D transposed convolutional neural network, Represents the reconstructed input sample; The sample reconstruction error corresponding to the original three-phase current monitoring time series is calculated as follows: ； In the formula, Indicates The sample reconstruction error corresponding to the input samples.

4. The method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement according to claim 1, characterized in that: The Gaussian mixture hidden Markov model is constructed, comprising: The operation process of the turbine under variable operating conditions is described as a double random process with a hidden Markov chain with hidden states and a random observation sequence obeying a Gaussian mixture distribution. The length of the observation sequence is , the number of hidden states is , the number of Gaussian distributions is , then the parameter set of the Gaussian mixture hidden Markov model is expressed as: ； in, represents the initial state probability matrix of the Gaussian mixture hidden Markov model, represents the state transition probability matrix of the Gaussian mixture hidden Markov model, represents the mixture probability matrix of the Gaussian mixture hidden Markov model, represents the mean matrix of the Gaussian mixture hidden Markov model, represents the covariance matrix of the Gaussian mixture hidden Markov model, Represents the set of parameters of the Gaussian mixture hidden Markov model.

5. The method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement according to claim 4, characterized in that: The estimation network is a multi-layer neural network, and the method of using a separate estimation network to respectively predict the mixed membership matrix and the state transition probability matrix includes: ； ； in, and They correspond to the weight parameters of two multi-layer neural networks, Represents a given low-dimensional temporal dependency feature representation The predicted value of the state transition probability matrix under Represents a given low-dimensional temporal dependency feature representation The predicted value of the mixed membership matrix under and Represent multi-layer neural network and activation function respectively.

6. The method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement according to claim 5, characterized in that: The method of estimating the parameters of the Gaussian mixture hidden Markov model according to the predicted values ​​of the mixed membership matrix and the state transition probability matrix includes: ； ； ； ； ； In the formula, The hidden state is The estimated value of the initial state probability when express The hidden state is Time The predicted value of the mixed membership of a Gaussian distribution fit, , ,combination indivual Get the estimated value of the initial state probability matrix of the Gaussian mixture hidden Markov model , Indicates the hidden state from arrive The estimated value of the state transition probability when Represents the total number of input samples in the sample data set, express Hidden state from arrive The predicted value of the state transition probability when ,combination indivual Get the estimated value of the state transition probability matrix of the Gaussian mixture hidden Markov model , The hidden state is Time The estimated value of the mixture probability of Gaussian distribution fitting, combination indivual Get an estimate of the mixing probability matrix for the Gaussian mixture hidden Markov model , The hidden state is Time An estimate of the mean of a Gaussian distribution fit, Represents low-dimensional temporal dependency feature representation, combination indivual Get an estimate of the mean matrix of the Gaussian mixture hidden Markov model , The hidden state is Time An estimate of the covariance of the Gaussian distribution fits, is a diagonal matrix, Indicates transposition, combination indivual Get an estimate of the covariance matrix of the Gaussian mixture hidden Markov model , and finally the estimated parameter set of the Gaussian mixture hidden Markov model is .

7. The method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement according to claim 1, characterized in that: The energy of low-dimensional time-series dependent feature representation is calculated based on the parameters of the Gaussian mixture hidden Markov model, including: No. The low-dimensional temporal dependency feature representation corresponding to the input samples , given the estimated Gaussian mixture hidden Markov model parameter set The conditional probability under for: ； Then calculate the low-dimensional temporal dependency feature representation The energy is: ； in, Represents the set of parameters of the estimated Gaussian mixture hidden Markov model Next, to Observation sequence The probability of Represents a given hidden state Hidden state The estimated state transition probability, For a given hidden state Hidden state The estimated observation probability of It is a low-dimensional temporal dependency feature representation The energy Represents a given sample sequence , about the estimated parameter set of the Gaussian mixture hidden Markov model The likelihood function of The parameter set of the Gaussian mixture hidden Markov model representing the given estimate The subsequent observation sequence is probability.

8. The method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement according to claim 1, characterized in that: The joint optimization loss function is constructed based on the sample reconstruction error, the energy represented by the low-dimensional time-dependent feature, and the penalty function, including: ； in, represents the joint optimization loss function, represents the weight parameter to be optimized, Represents the total number of input samples in the sample data set, Indicates The sample reconstruction error corresponding to the input samples is Indicates the first input samples, Indicates The reconstructed input sample corresponding to the input sample, is the energy weighting coefficient, is the weight coefficient of the penalty function, For the The energy of the low-dimensional time-dependent feature representation corresponding to the input samples, is the penalty function, expressed as: ； In the formula, represents the estimate of the covariance matrix of the Gaussian mixture hidden Markov model, is the number of hidden states of the Gaussian mixture hidden Markov model, is the number of Gaussian distributions in the Gaussian mixture hidden Markov model, represents the dimensionality of the low-dimensional temporal-dependent feature representation generated by the compression network, The hidden state is Time An estimate of the covariance of the Gaussian distribution fits, is a diagonal matrix, Represents the diagonal matrix Line Covariance values ​​for the columns.

9. The method for explaining fault detection of a hydraulic turbine based on non-intrusive measurement according to claim 1, characterized in that: The threshold is the 5th percentile of all samples in the sample data set. If the energy corresponding to the newly collected original three-phase current monitoring time series of the turbine is greater than the threshold, it is determined to be in normal working condition; otherwise, it is in abnormal working condition.

Images