Fault diagnosis method, system and medium for pumped storage unit based on vibration signal

Abstract

The present invention provides a pumped-storage unit fault diagnosis method, system and medium based on vibration signals. The method collects normal vibration signals and known fault vibration signals of the pumped-storage unit, and processes them to obtain fused feature vector samples of the normal vibration signals and fused feature vector samples of the known fault vibration signals; based on a variational autoencoder, a state single classifier and a fault classification unit are constructed according to the fused feature vector samples; based on the whale optimization algorithm, a support vector machine is optimized, and a known fault type diagnosis model for diagnosing the types of known faults of the pumped-storage unit is constructed according to the fused feature vector samples of the known fault vibration signals; finally, the vibration signals of the pumped-storage unit collected in real time are used to diagnose the faults of the pumped-storage unit through a state single classifier, a fault classification unit and a known fault type diagnosis model, thereby greatly improving the timeliness and effectiveness of the diagnosis.

Description

Technical Field

[0001] The invention belongs to the technical field of electric power engineering and relates to a fault diagnosis method, system and medium. Background Art

[0002] The utilization and development of wind power and photovoltaic power generation can effectively reduce carbon emissions. However, renewable energy sources such as wind power and photovoltaic power generation are intermittent and have random fluctuations. The access of a large number of renewable energy sources to the power grid has put forward higher requirements for maintaining the safe, stable and efficient operation of the power grid.

[0003] Pumped-storage power stations are currently the most mature and large-scale energy storage and regulation power sources. They have the important functions of peak load regulation, frequency regulation, and ensuring the safety and stability of the power grid. As the core equipment for peak load regulation, the normal and stable operation of pumped-storage units is the basis for the power station to play its role. Due to the complex and changeable working conditions of pumped-storage units and the influence of water-electromechanical coupling factors, pumped-storage units inevitably have certain vibrations. When the vibration exceeds the rated range, it will affect the normal operation of the pumped-storage units and even cause serious failures. Therefore, it is of great significance to study and analyze vibration monitoring data and conduct timely and effective fault diagnosis of pumped-storage units. Due to the complexity of the operating environment of mechanical equipment and the continuous variability of working conditions, unknown faults will inevitably occur with the continuous operation of pumped-storage units. However, the current fault diagnosis methods for pumped-storage units based on vibration signals often rely on known fault cases, and require supervised learning of the mapping relationship between samples and faults. It is impossible to complete the diagnosis of unknown faults in a purely data-driven way, resulting in poor timeliness and effectiveness of diagnosis. Summary of the invention

[0004] In order to solve the problem of poor timeliness and effectiveness of the fault diagnosis method of the pumped storage unit described in the background technology, the present invention provides a pumped storage unit fault diagnosis method, system and medium based on vibration signals.

[0005] The method of the present invention comprises:

[0006] Collect normal vibration signals of pumped storage units and known fault vibration signals of pumped storage units;

[0007] Based on the amplitude domain analysis method, Fourier transform theory and one-dimensional convolutional neural network, a vibration signal multi-dimensional fusion feature vector extraction model is constructed;

[0008] The normal vibration signal of the pumped-storage unit and the known fault vibration signal of the pumped-storage unit are processed by a vibration signal multi-dimensional fusion feature vector extraction model, and the fusion feature vector samples of the normal vibration signal of the pumped-storage unit and the fusion feature vector samples of the known fault vibration signal of the pumped-storage unit are extracted;

[0009] According to the fused feature vector samples of the normal vibration signal of the pumped storage unit, a state single classifier for identifying the normal state and abnormal state of the pumped storage unit is constructed based on the variational autoencoder.

[0010] According to the fused feature vector samples of known fault vibration signals of pumped storage units, a fault classification unit for identifying known and unknown faults of pumped storage units is constructed based on variational autoencoder.

[0011] According to the fusion feature vector samples of known fault vibration signals of pumped storage units, the support vector machine is optimized based on the whale optimization algorithm to build a known fault type diagnosis model for diagnosing known fault types of pumped storage units.

[0012] When the pumped-storage unit is in operation, the vibration signal of the pumped-storage unit is collected in real time, and the fused feature vector samples of the vibration signal of the pumped-storage unit are extracted through the vibration signal multi-dimensional fusion feature vector extraction model; the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the state single classifier to identify the normal state and abnormal state of the pumped-storage unit; when the pumped-storage unit is identified as being in an abnormal state, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the fault classification unit to identify whether the pumped-storage unit is in a known fault or an unknown fault; when the pumped-storage unit is identified as being in a known fault, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the known fault type diagnosis model to diagnose the known fault type of the pumped-storage unit; when the pumped-storage unit is identified as being in an unknown fault, an unknown fault label is output, and the unknown fault type is diagnosed by combining methods including expert experience and spectrum analysis.

[0013] Furthermore, the method for constructing the vibration signal multi-dimensional fusion feature vector extraction model includes:

[0014] The typical time domain characteristics of the vibration signal of the pumped storage unit are extracted based on the amplitude domain analysis method, and the characteristic parameters of the typical time domain characteristics include: simple statistical characteristic parameters and high-order statistical characteristic parameters; the simple statistical characteristic parameters include mean value, root mean square value, peak-to-peak value; the high-order statistical characteristic parameters include slope, kurtosis, and peak state;

[0015] Based on Fourier transform theory, typical frequency domain features of the vibration signal of the pumped storage unit are extracted, wherein the typical frequency domain features include: frequency amplitude mean, frequency centroid value, frequency standard deviation, frequency variation coefficient, frequency slope and frequency amplitude root mean square;

[0016] The implicit features of the vibration signal of the pumped storage unit are mined through a one-dimensional convolutional neural network. The structure of the one-dimensional convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer. The internal spatial features of the data are extracted by alternating the convolutional layer and the pooling layer in a way of local connection and shared weights, so as to obtain the potential deep features of the vibration signal of the pumped storage unit. The specific operations include:

[0017] The convolution layer extracts feature information by sliding convolution operation of the convolution kernel on the input data. For a given input image I and convolution kernel K, the convolution operation can be expressed as:

[0018] P i,j =∑ m ∑ n I i+m,j+n K m,n (1)

[0019] Where P i,j is the pixel value of the output feature map at position (i, j), (m, n) represents the row and column index of the convolution kernel. In actual calculation, the vibration signal feature vector is one-dimensional data, and a one-dimensional convolutional neural network is used to perform convolution operation on one dimension of the data matrix;

[0020] Apply the activation function to the result of the convolution operation and set the ReLU function as the activation function of the one-dimensional convolutional neural network. Its expression is as follows:

[0021] f ReLU (x) = max(0, x)(2);

[0022] The pooling layer downsamples the features output by the convolutional layer through a pooling operation, wherein the pooling operation includes maximum pooling and average pooling;

[0023] In the maximum pooling, the pooling operation selects the maximum value in the input area as the output, and the expression is as follows:

[0024]

[0025] In the formula, y l Represents the output value after pooling; Represents the value of the feature map at position (i, j) of the input pooling layer;

[0026] In the average pooling, the average value in the input area is selected as the output. Assuming the size of the pooling window is m×n, the expression is as follows:

[0027]

[0028] The fully connected layer is used to integrate the local information with category distinction in the convolution layer or the pooling layer. Based on the softmax function, the output of the fully connected layer will be converted into the probability distribution of the fault category. The output layer will take the label output corresponding to the maximum probability value as the recognition result of the sample type;

[0029] When the one-dimensional convolutional neural network performs feature extraction, the model is first pre-trained to obtain the optimal pre-trained model. At this time, sample data is input and the output vector of the fully connected layer is taken out as the multi-dimensional fusion feature vector of the vibration signal of the pumped storage unit.

[0030] Furthermore, the method for constructing the state single classifier includes: constructing a variational autoencoder model for identifying the normal state and abnormal state of the pumped-storage unit as the state single classifier based on the fused feature vector samples of the normal vibration signal of the pumped-storage unit;

[0031] In the variational autoencoder model, assuming a given D-dimensional sample time series dataset For each sample x, the encoding process of the variational autoencoder maps complex data to the latent variable space, as shown below:

[0032]

[0033] In the formula, is the distribution function of the encoder, z is the potential feature, and x is the input data;

[0034] The decoding process is to restore the latent variable space data to the original data space, by learning the potential features μ and σ of the data and generating new data samples. The decoding expression is as follows:

[0035] x'~Dec(z)=p θ (x|z)(8),

[0036] Where θ is the distribution function of the decoder; x' is the reconstructed data;

[0037] Variational autoencoder introduces a probability distribution Instead of the true posterior distribution p θ (z|x), then from Sampling from the posterior distribution p θ (z|x) sampling, in order to make and p θ (z|x) are as equal as possible, and KL divergence is used to evaluate the similarity between different distributions. If and only if KL divergence is 0, it means and p θ (z|x) are exactly equal, and by optimizing the parameters Minimize the KL divergence with θ, as follows:

[0038]

[0039] make but:

[0040]

[0041] In the formula, It is called the variational lower bound, which minimizes By maximizing accomplish;

[0042] Assume that from Take S samples, variational lower bound Further expressed as:

[0043]

[0044] Where z′ is the lth sample; p θ (z) is the standard normal distribution;

[0045] Reparameterization is used to simplify the operation. An auxiliary variable ε is obtained from the standard normal N(0, 1), and then it is calculated using μ and σ. The expression is as follows:

[0046]

[0047] When the batch size is large enough, we take S = 1, that is, only one sample data is sampled each time, and formula (11) is expressed as:

[0048]

[0049] In the formula, the first term is the approximate posterior distribution With the prior distribution p θ KL divergence of (z); the second term is the reconstruction loss term of the variational autoencoder; the variational autoencoder training process aims to maximize This is equivalent to minimizing the KL divergence and reconstruction loss at the same time, so Can also be seen as Regularization term that approximates the standard normal;

[0050] Through the above operations, the variational autoencoder maps the input normal vibration signal fusion feature data of the pumped-storage unit into an ideal Gaussian distribution through the encoder, and then inputs the samples sampled by the Gaussian distribution into the decoder. The decoder generates reconstructed data and constructs a fault symptom indicator by comparing the difference between the input data and the reconstructed data. When the fault symptom indicator exceeds the threshold, it is judged that the pumped-storage unit is in an abnormal state, otherwise it is in a normal state.

[0051] Furthermore, the method for constructing the fault classification unit includes: constructing a variational autoencoder model for identifying normal and abnormal states of the pumped-storage unit as a fault classification unit based on the fused feature vector samples of the normal vibration signal of the pumped-storage unit;

[0052] In the variational autoencoder model, assuming a given D-dimensional sample time series dataset For each sample x, the encoding process of the variational autoencoder maps complex data to the latent variable space, as shown below:

[0053]

[0054] In the formula, is the distribution function of the encoder, z is the potential feature, and x is the input data;

[0055] The decoding process is to restore the latent variable space data to the original data space, by learning the potential features μ and σ of the data and generating new data samples. The decoding expression is as follows:

[0056] x'~Dec(z)=p θ (x|z)(8),

[0057] Where θ is the distribution function of the decoder; x' is the reconstructed data;

[0058] Variational autoencoder introduces a probability distribution Instead of the true posterior distribution p θ (z|x), then from Sampling from the posterior distribution p θ (z|x) sampling, in order to make and p θ (z|x) are as equal as possible, and KL divergence is used to evaluate the similarity between different distributions. If and only if KL divergence is 0, it means and p θ (z|x) are exactly equal, and by optimizing the parameters Minimize the KL divergence with θ, as follows:

[0059]

[0060] make but:

[0061]

[0062] In the formula, It is called the variational lower bound;

[0063] Assume that from Take S samples, variational lower bound Further expressed as:

[0064]

[0065] Where z′ is the lth sample; p θ (z) is Gaussian distribution;

[0066] Reparameterization is used to simplify the operation. An auxiliary variable ε is obtained from the standard normal N(0, 1), and then it is calculated using μ and σ. The expression is as follows:

[0067]

[0068] When the batch size is large enough, we take S = 1, that is, only one sample data is sampled each time, and formula (11) is expressed as:

[0069]

[0070] In the formula, the first term is the approximate posterior distribution With the prior distribution p θ KL divergence of (z); the second term is the reconstruction loss term of the variational autoencoder; the variational autoencoder training process aims to maximize This is equivalent to minimizing the KL divergence and reconstruction loss at the same time, so Can also be seen as Regularization term that approximates the standard normal;

[0071] Through the above operations, the variational autoencoder maps the input known fault vibration signal fusion feature data into an ideal Gaussian distribution through the encoder, and then inputs the samples sampled by the Gaussian distribution into the decoder, which generates reconstructed data. The fault classification index is constructed by comparing the difference between the input data and the reconstructed data. When the fault classification index exceeds the threshold, it is judged that the pumped storage unit is in an unknown fault, otherwise it is in this type of known fault;

[0072] For each known fault type of the pumped-storage unit, the variational autoencoder model is independently trained with its multi-dimensional vibration signal fusion features as input, and then a fault classification unit for known and unknown faults of the pumped-storage unit composed of multiple variational autoencoders is constructed. Each variational autoencoder focuses on judging whether the unit has a corresponding known type of fault. Through the collaborative work of each variational autoencoder model, the fault classification unit can realize real-time diagnosis of known and unknown faults of the pumped-storage unit.

[0073] Furthermore, the method for constructing the known fault type diagnosis model includes:

[0074] Based on the fused feature vector samples of known fault vibration signals of pumped storage units, the whale optimization algorithm is used to optimize the parameters of the support vector machine. Combining the global optimization ability of the whale optimization algorithm and the classification ability of the support vector machine, a known fault type diagnosis model for diagnosing the known fault types of pumped storage units is constructed.

[0075] The support vector machine maps the fault data to a high-dimensional space through a nonlinear kernel function, constructs a Lagrangian function to obtain the optimal classification hyperplane; assuming that a set of training samples {(x 1 ,y 1 ),(x 2 ,y 2 ),...,(x N ,y N )}, where x i belongs to the input vector, y i is the corresponding output vector. For nonlinear problems, the support vector machine introduces relaxation variables With the penalty coefficient C, the Lagrangian function is defined as follows:

[0076]

[0077] In the formula, a i ≥0 is the Lagrange multiplier, w represents the weight, and b is the bias vector;

[0078] Introducing the radial basis kernel function K(x i ,x j ) = exp(||x i -x j || 2 / (2γ 2 )), optimize the objective function and constraints, the expression is as follows:

[0079]

[0080]

[0081] This problem is a quadratic function optimization problem with a unique solution. The optimal solution a is defined as i * , that is:

[0082]

[0083] a i * ≠0 corresponding sample data x i * It is the support vector. Finally, the optimal classification hyperplane of the support vector machine is:

[0084]

[0085] Among them, b * is the classification threshold, by solving a i [y i (w·x i )+b]-1=0 to obtain;

[0086] The whale optimization algorithm is used to optimize the penalty coefficient C and the kernel function parameter γ. The specific operations are as follows:

[0087] The whale optimization algorithm simulates the hunting behavior of humpback whales, which includes surrounding prey, attacking with a bubble net, and searching for prey;

[0088] The formula for the encirclement phase is as follows:

[0089]

[0090] Where: t represents the number of iterations; t max Indicates the maximum number of iterations; is the coefficient vector; A random number between the range [0, 1]; represents the search particle, Represents the target prey, which is the position vector of the current optimal solution, and is continuously updated in iterations to obtain a better solution; is the convergence factor that decreases linearly from 2 to 0 during the iteration process;

[0091] The whale optimization algorithm mathematically models the bubble network attack behavior through the shrinking and surrounding mechanism and the spiral update mechanism, which are used alternately according to the probability p, where p is a probability randomly generated from [0, 1]. When p < 0.5, the shrinking and surrounding mechanism is used; and when p ≥ 0.5, the spiral mechanism is used to update the position of the search particle. The expression is as follows:

[0092]

[0093] Where d is the shape parameter of the logarithmic spiral, l is a random number between [-1, 1], Represents the distance between the search particle and the current optimal solution;

[0094] It decreases linearly, resulting in The value of will be Changes between Then the search particles gradually surround the current optimal solution, which belongs to the local optimization stage in the whale optimization algorithm; The distance data will be randomly updated, causing the whale to deviate from the original optimal individual and update the population according to the random position to perform a global search. In the process of searching for prey, its mathematical model is defined as follows:

[0095]

[0096] In the formula, represents a search particle randomly selected from the population;

[0097] Based on the vibration signal feature set of known faults of pumped storage units, combined with the strong global optimization ability of the whale optimization algorithm and the classification advantage of the support vector machine, the known fault types can be identified.

[0098] The present invention also proposes a pumped-storage unit fault diagnosis system based on vibration signals, including a vibration signal collection module, a multi-dimensional fusion feature vector extraction model construction module, a vibration signal fusion feature vector sample extraction module, a state single classifier construction module, a fault classification unit construction module, a known fault type diagnosis model construction module, and a pumped-storage unit fault diagnosis module.

[0099] The vibration signal collection module is used to collect normal vibration signals of the pumped-storage unit and known fault vibration signals of the pumped-storage unit.

[0100] The multi-dimensional fusion feature vector extraction model construction module constructs a vibration signal multi-dimensional fusion feature vector extraction model based on an amplitude domain analysis method, Fourier transform theory and a one-dimensional convolutional neural network.

[0101] The vibration signal fusion feature vector sample extraction module processes the normal vibration signal of the pumped-storage unit and the known fault vibration signal of the pumped-storage unit through a vibration signal multi-dimensional fusion feature vector extraction model, and extracts the fusion feature vector samples of the normal vibration signal of the pumped-storage unit and the fusion feature vector samples of the known fault vibration signal of the pumped-storage unit.

[0102] The state single classifier construction module constructs a state single classifier for identifying the normal state and abnormal state of the pumped-storage unit based on the fused feature vector samples of the normal vibration signal of the pumped-storage unit and based on the variational autoencoder.

[0103] The fault classifier unit construction module constructs a fault classification unit for identifying known faults and unknown faults of the pumped-storage unit based on a variational autoencoder according to the fused feature vector samples of the known fault vibration signal of the pumped-storage unit.

[0104] The known fault type diagnosis model construction module constructs a known fault type diagnosis model for diagnosing the types of known faults of the pumped-storage unit based on the fused feature vector samples of the known fault vibration signal of the pumped-storage unit and the support vector machine optimized based on the whale optimization algorithm.

[0105] The fault diagnosis module of the pumped-storage unit is used to collect the vibration signal of the pumped-storage unit in real time when the pumped-storage unit is running, and extract the fused feature vector samples of the vibration signal of the pumped-storage unit through the vibration signal multi-dimensional fusion feature vector extraction model; input the fused feature vector samples of the vibration signal of the pumped-storage unit into the state single classifier to identify the normal state and abnormal state of the pumped-storage unit; when the pumped-storage unit is identified as being in an abnormal state, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the fault classification unit to identify whether the pumped-storage unit is in a known fault or an unknown fault; when the pumped-storage unit is identified as being in a known fault, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the known fault type diagnosis model to diagnose the known fault type of the pumped-storage unit; when the pumped-storage unit is identified as being in an unknown fault, an unknown fault label is output, and the unknown fault type is diagnosed by combining methods including expert experience and spectrum analysis.

[0106] The present invention also proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the vibration signal-based pumped storage unit fault diagnosis method as described above is implemented.

[0107] The present invention collects normal vibration signals and known fault vibration signals of the pumped-storage unit, and processes them to obtain fused feature vector samples of the normal vibration signal and fused feature vector samples of the known fault vibration signal; based on the variational autoencoder, according to the fused feature vector samples of the normal vibration signal, a state single classifier for identifying the normal state and abnormal state of the pumped-storage unit is constructed; based on the variational autoencoder, according to the fused feature vector samples of the known fault vibration signal, a fault classification unit for identifying the known fault and unknown fault of the pumped-storage unit is constructed; based on the whale optimization algorithm to optimize the support vector machine, according to the fusion feature vector samples of the known fault vibration signal Feature vector samples are used to construct a known fault type diagnosis model for diagnosing the known fault types of the pumped-storage unit; finally, for the vibration signal of the pumped-storage unit collected in real time, the normal state and abnormal state of the pumped-storage unit are identified by a state single classifier, and the pumped-storage unit is identified as a known fault or an unknown fault by a fault classification unit. The known fault type diagnosis model is used to diagnose the known fault type of the pumped-storage unit. When it is identified that the pumped-storage unit is in an unknown fault, the unknown fault label is output, and the unknown fault type is diagnosed by combining methods including expert experience and spectrum analysis.

[0108] Compared with the prior art, the present invention has the following beneficial effects: First, in view of the problem that the vibration signal of the pumped-storage unit generally has nonlinear and non-stationary characteristics, and a single feature is difficult to accurately reflect the state of the unit, a multi-dimensional fusion feature vector extraction model of the vibration signal is constructed based on the amplitude domain analysis method, Fourier transform theory and one-dimensional convolutional neural network, which realizes a more comprehensive and accurate characterization of the operating state of the pumped-storage unit, and lays a foundation for improving the robustness and accuracy of fault diagnosis; Second, in view of the problem that traditional fault diagnosis methods are difficult to identify unknown faults, a state single classifier and a fault classification unit are constructed by combining the known state vibration signal fusion feature vector samples and variational autoencoders of the pumped-storage unit, so as to realize the monitoring of the abnormal state of the unit and the diagnosis of the fault type, and reduce the misidentification of unknown faults; Third, based on the whale optimization algorithm to optimize the support vector machine, according to the fusion feature vector samples of the known fault vibration signal, a known fault type diagnosis model for diagnosing the known fault types of the pumped-storage unit is constructed, so as to realize the accurate classification of known fault types.

[0109] In general, the present invention uses a purely data-driven method such as vibration signals to simultaneously identify known faults and unknown faults of the pumped-storage unit during its operation, and accurately diagnose known fault types and mark unknown faults to achieve abnormal state diagnosis of the pumped-storage unit. This breaks through the application limitations of traditional intelligent fault diagnosis models, greatly improves the timeliness and effectiveness of diagnosis, and is of great significance for ensuring the safe and stable operation of pumped-storage unit equipment, reducing production safety risks, and improving production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] Figure 1 The figure is a flow chart of the method of the present invention.

[0111] Figure 2 This is a fault diagnosis flow chart of the pumped storage unit of the present invention.

[0112] Figure 3 It is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0113] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0114] Fault diagnosis method of pumped storage unit based on vibration signal, the flow chart is as follows Figure 1 As shown, the details are as follows.

[0115] Firstly, the normal vibration signal of the pumped storage unit and the known fault vibration signal of the pumped storage unit are collected.

[0116] Then, based on the amplitude domain analysis method, Fourier transform theory and one-dimensional convolutional neural network, a multi-dimensional fusion feature vector extraction model of vibration signal was constructed.

[0117] Specifically, the construction method of the vibration signal multi-dimensional fusion feature vector extraction model is:

[0118] The typical time domain characteristics of the vibration signal of the pumped storage unit are extracted based on the amplitude domain analysis method, and the characteristic parameters of the typical time domain characteristics include: simple statistical characteristic parameters and high-order statistical characteristic parameters; the simple statistical characteristic parameters include mean value, root mean square value, peak-to-peak value; the high-order statistical characteristic parameters include slope, kurtosis, and peak state;

[0119] Based on Fourier transform theory, typical frequency domain features of the vibration signal of the pumped storage unit are extracted, wherein the typical frequency domain features include: frequency amplitude mean, frequency centroid value, frequency standard deviation, frequency variation coefficient, frequency slope and frequency amplitude root mean square;

[0120] The implicit features of the vibration signal of the pumped storage unit are mined through a one-dimensional convolutional neural network. The structure of the one-dimensional convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer. The internal spatial features of the data are extracted by alternating the convolutional layer and the pooling layer in a way of local connection and shared weights, so as to obtain the potential deep features of the vibration signal of the pumped storage unit. The specific operations include:

[0121] The convolution layer extracts feature information by sliding convolution operations on the input data. For a given input image I and convolution kernel K, the convolution operation can be expressed as:

[0122] P i,j =∑ m ∑ n I i+m,j+n K m,n (1)

[0123] Where P i,j is the pixel value of the output feature map at position (i, j), and (m, n) represents the row and column index of the convolution kernel;

[0124] Apply the activation function to the result of the convolution operation and set the ReLU function as the activation function of the one-dimensional convolutional neural network. Its expression is as follows:

[0125] f ReLU (x) = max(0, x)(2);

[0126] The pooling layer downsamples the features output by the convolutional layer through a pooling operation, wherein the pooling operation includes maximum pooling and average pooling;

[0127] In the maximum pooling, the pooling operation selects the maximum value in the input area as the output, and the expression is as follows:

[0128]

[0129] In the formula, y l Represents the output value after pooling; Represents the value of the feature map at position (i, j) of the input pooling layer;

[0130] In average pooling, the average value in the input area is selected as the output. Assuming the size of the pooling window is m×n, the expression is as follows:

[0131]

[0132] The fully connected layer is used to integrate the local information with category distinction in the convolution layer or pooling layer. Based on the softmax function, the output of the fully connected layer will be converted into the probability distribution of the fault category. The output layer will take the label output corresponding to the maximum probability value as the identification result of the sample type;

[0133] When the one-dimensional convolutional neural network is used for feature extraction, the model is first pre-trained to obtain the optimal pre-trained model. At this time, the sample data is input and the output vector of the fully connected layer is taken out as the multi-dimensional fusion feature vector of the vibration signal of the pumped storage unit.

[0134] Then, the normal vibration signal of the pumped-storage unit and the vibration signal of the known fault of the pumped-storage unit are processed through the multi-dimensional fusion feature vector extraction model of the vibration signal, and the fused feature vector samples of the normal vibration signal of the pumped-storage unit and the fused feature vector samples of the known fault vibration signal of the pumped-storage unit are extracted.

[0135] Then, according to the fused feature vector samples of the normal vibration signal of the pumped-storage unit, a state single classifier for identifying the normal and abnormal states of the pumped-storage unit is constructed based on the variational autoencoder.

[0136] Specifically, the construction method of the state single classifier is as follows: based on the fused feature vector samples of the normal vibration signal of the pumped storage unit, a variational autoencoder model for identifying the normal state and abnormal state of the pumped storage unit is constructed as the state single classifier;

[0137] In the variational autoencoder model, assuming a given D-dimensional sample time series dataset For each sample x, the encoding process of the variational autoencoder maps complex data to the latent variable space, as shown below:

[0138]

[0139] In the formula, is the distribution function of the encoder, z is the potential feature, and x is the input data;

[0140] The decoding process is to restore the latent variable space data to the original data space, by learning the potential features μ and σ of the data and generating new data samples. The decoding expression is as follows:

[0141] x'~Dec(z)=p θ (x|z)(8),

[0142] Where θ is the distribution function of the decoder; x' is the reconstructed data;

[0143] Variational autoencoder introduces a probability distribution Instead of the true posterior distribution p θ (z|x), then from Sampling from the posterior distribution p θ (z|x) sampling, in order to make and p θ (z|x) are as equal as possible, and KL divergence is used to evaluate the similarity between different distributions. If and only if KL divergence is 0, it means and p θ (z|x) are exactly equal, and by optimizing the parameters Minimize the KL divergence with θ, as follows:

[0144]

[0145] make but:

[0146]

[0147] In the formula, It is called the variational lower bound, which minimizes By maximizing accomplish;

[0148] Assume that from Take S samples, variational lower bound Further expressed as:

[0149]

[0150] Where z′ is the lth sample; p θ (z) is the standard normal distribution;

[0151] Reparameterization is used to simplify the operation. An auxiliary variable ε is obtained from the standard normal N(0, 1), and then it is calculated using μ and σ. The expression is as follows:

[0152]

[0153] When the batch size is large enough, we take S = 1, that is, only one sample data is sampled each time, and formula (11) is expressed as:

[0154]

[0155] In the formula, the first term is the approximate posterior distribution With the prior distribution p θ KL divergence of (z); the second term is the reconstruction loss term of the variational autoencoder; the variational autoencoder training process aims to maximize This is equivalent to minimizing the KL divergence and reconstruction loss at the same time, so Can also be seen as Regularization term that approximates the standard normal;

[0156] Through the above operations, the variational autoencoder maps the fused feature data of the normal vibration signal of the input pumped-storage unit into an ideal Gaussian distribution through the encoder, and then inputs the samples sampled by the Gaussian distribution into the decoder. The decoder generates reconstructed data, and constructs a fault symptom indicator by comparing the difference between the input data and the reconstructed data. When the fault symptom indicator exceeds the threshold, it is judged that the pumped-storage unit is in an abnormal state, otherwise it is in a normal state.

[0157] The fault symptom indicator may be the mean square error between the input data and the reconstructed data, and the threshold may be set according to actual conditions by combining manual experience and the 3σ criterion.

[0158] Then, according to the fused feature vector samples of the known fault vibration signals of the pumped-storage unit, a fault classification unit for identifying known and unknown faults of the pumped-storage unit is constructed based on the variational autoencoder.

[0159] Specifically, the fault classification unit is constructed as follows: based on the vibration signal fusion feature set of a known fault of the pumped storage unit, a variational autoencoder model for judging whether the known fault occurs is constructed as the fault classification unit;

[0160] In the variational autoencoder model, assuming a given D-dimensional sample time series dataset For each sample x, the encoding process of the variational autoencoder maps complex data to the latent variable space, as shown below:

[0161]

[0162] In the formula, is the distribution function of the encoder, z is the potential feature, and x is the input data;

[0163] The decoding process is to restore the latent variable space data to the original data space, by learning the potential features μ and σ of the data and generating new data samples. The decoding expression is as follows:

[0164] x'~Dec(z)=p θ (x|z)(8),

[0165] Where θ is the distribution function of the decoder; x' is the reconstructed data;

[0166] Variational autoencoder introduces a probability distribution Instead of the true posterior distribution p θ (z|x), then from Sampling from the posterior distribution p θ (z|x) sampling, in order to make and p θ (z|x) are as equal as possible, and KL divergence is used to evaluate the similarity between different distributions. If and only if KL divergence is 0, it means and p θ (z|x) are exactly equal, and by optimizing the parameters Minimize the KL divergence with θ, as follows:

[0167]

[0168] make but:

[0169]

[0170] In the formula, It is called the variational lower bound;

[0171] Assume that from Take S samples, variational lower bound Further expressed as:

[0172]

[0173] Where z is the lth sample; p θ (z) is Gaussian distribution;

[0174] Reparameterization is used to simplify the operation. An auxiliary variable ε is obtained from the standard normal N(0, 1), and then it is calculated using μ and σ. The expression is as follows:

[0175]

[0176] When the batch size is large enough, we take S = 1, that is, only one sample data is sampled each time, and formula (11) is expressed as:

[0177]

[0178] In the formula, the first term is the approximate posterior distribution With the prior distribution p θ KL divergence of (z); the second term is the reconstruction loss term of the variational autoencoder; the variational autoencoder training process aims to maximize This is equivalent to minimizing the KL divergence and reconstruction loss at the same time, so Can also be seen as Regularization term that approximates the standard normal;

[0179] Through the above operations, the variational autoencoder maps the fused feature vector samples of the known fault vibration signal of the input data pumped-storage unit into an ideal Gaussian distribution through the encoder, and then inputs the samples sampled by the Gaussian distribution into the decoder. The decoder generates reconstructed data, and constructs a fault classification index by comparing the difference between the input data and the reconstructed data. When the fault classification index exceeds the threshold, it is judged that the pumped-storage unit is in an unknown fault, otherwise it is in this type of known fault.

[0180] Among them, the fault classification index can be the mean square error between the input data and the reconstructed data, and the threshold can be set according to the actual situation by combining manual experience and the 3σ criterion and other methods.

[0181] For each known fault type of the pumped-storage unit, the variational autoencoder model is independently trained with its multi-dimensional vibration signal fusion features as input, and then a fault classification unit for known and unknown faults of the pumped-storage unit composed of multiple variational autoencoders is constructed. Each variational autoencoder focuses on judging whether the unit has a corresponding known type of fault. Through the collaborative work of each variational autoencoder model, the fault classification unit can realize real-time diagnosis of known and unknown faults of the pumped-storage unit.

[0182] Then, according to the fused feature vector samples of the known fault vibration signals of the pumped-storage unit, the support vector machine is optimized based on the whale optimization algorithm, and a known fault type diagnosis model is constructed to diagnose the known fault types of the pumped-storage unit.

[0183] Specifically, the construction method of the known fault type diagnosis model is:

[0184] Based on the fused feature vector samples of known fault vibration signals of pumped storage units, the whale optimization algorithm is used to optimize the parameters of the support vector machine. Combining the global optimization ability of the whale optimization algorithm and the classification ability of the support vector machine, a known fault type diagnosis model for diagnosing the known fault types of pumped storage units is constructed.

[0185] The support vector machine maps the fault data to a high-dimensional space through a nonlinear kernel function and constructs a Lagrangian function to obtain the optimal classification hyperplane. Assuming that a set of training samples {(x 1 ,y 1 ),(x 2 ,y 2 ),...,(x N ,y N )}, where x i belongs to the input vector, y i is the corresponding output vector. For nonlinear problems, the support vector machine introduces relaxation variables With the penalty coefficient C, the Lagrangian function is defined as follows:

[0186]

[0187] In the formula, a i ≥0 is the Lagrange multiplier, w represents the weight, and b is the bias vector;

[0188] Introducing the radial basis kernel function K(x i ,x j ) = exp(||x i -x j || 2 / (2γ 2 )), optimize the objective function and constraints, the expression is as follows:

[0189]

[0190]

[0191] This problem is a quadratic function optimization problem with a unique solution. The optimal solution a is defined as i * , that is:

[0192]

[0193] a i * ≠0 corresponding sample data x i * It is the support vector. Finally, the optimal classification hyperplane of the support vector machine is:

[0194]

[0195] Among them, b * is the classification threshold, by solving a i [y i (w·x i )+b]-1=0 to obtain;

[0196] When the radial basis function is used as the kernel function, the selection of the penalty coefficient C and the kernel function parameter γ directly affects the anti-interference ability and recognition performance of the support vector machine;

[0197] The whale optimization algorithm is used to optimize the penalty coefficient C and the kernel function parameter γ. The specific operations are as follows:

[0198] The whale optimization algorithm simulates the hunting behavior of humpback whales, which includes surrounding prey, attacking with a bubble net, and searching for prey;

[0199] The formula for the encirclement phase is as follows:

[0200]

[0201] Where: t represents the number of iterations; t max Indicates the maximum number of iterations; is the coefficient vector; A random number between the range [0, 1]; represents the search particle, Represents the target prey, which is the position vector of the current optimal solution, and is continuously updated in iterations to obtain a better solution; is the convergence factor that decreases linearly from 2 to 0 during the iteration process;

[0202] The whale optimization algorithm mathematically models the bubble network attack behavior through the shrinking and surrounding mechanism and the spiral update mechanism, which are used alternately according to the probability p, where p is a probability randomly generated from [0, 1]. When p < 0.5, the shrinking and surrounding mechanism is used; and when p ≥ 0.5, the spiral mechanism is used to update the position of the search particle. The expression is as follows:

[0203]

[0204] Where d is the shape parameter of the logarithmic spiral, l is a random number between [-1, 1], Represents the distance between the search particle and the current optimal solution;

[0205] a decreases linearly, resulting in The value of will be Changes between Then the search particles gradually surround the current optimal solution, which belongs to the local optimization stage in the whale optimization algorithm; The distance data will be randomly updated, causing the whale to deviate from the original optimal individual and update the population according to the random position to perform a global search. In the process of searching for prey, its mathematical model is defined as follows:

[0206]

[0207] In the formula, represents a search particle randomly selected from the population;

[0208] Based on the vibration signal feature set of known faults of pumped storage units, combined with the strong global optimization ability of the whale optimization algorithm and the classification advantage of the support vector machine, the known fault types can be identified.

[0209] Finally, when the pumped-storage unit is in operation, the vibration signal of the pumped-storage unit is collected in real time, and the fused feature vector samples of the vibration signal of the pumped-storage unit are extracted through the vibration signal multi-dimensional fusion feature vector extraction model; the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the state single classifier to identify the normal state and abnormal state of the pumped-storage unit; when the pumped-storage unit is identified as being in an abnormal state, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the fault classification unit to identify whether the pumped-storage unit is in a known fault or an unknown fault; when the pumped-storage unit is identified as being in a known fault, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the known fault type diagnosis model to diagnose the known fault type of the pumped-storage unit; when the pumped-storage unit is identified as being in an unknown fault, the unknown fault label is output, and the unknown fault type is diagnosed by combining methods including expert experience and spectrum analysis.

[0210] Through the above content, the fault diagnosis process of the pumped storage unit is described in detail. The fault diagnosis flow chart is as follows: Figure 2 shown.

[0211] The present invention also proposes a pumped storage unit fault diagnosis system based on vibration signals, and its architecture diagram is as follows: Figure 3 As shown, it consists of a vibration signal collection module, a multi-dimensional fusion feature vector extraction model construction module, a vibration signal fusion feature vector sample extraction module, a state single classifier construction module, a fault classification unit construction module, a known fault type diagnosis model construction module, and a pumped storage unit fault diagnosis module.

[0212] Vibration signal collection module, used to collect normal vibration signals of pumped storage units and known fault vibration signals of pumped storage units

[0213] The multi-dimensional fusion feature vector extraction model construction module constructs a multi-dimensional fusion feature vector extraction model for vibration signals based on the amplitude domain analysis method, Fourier transform theory and one-dimensional convolutional neural network.

[0214] The vibration signal fusion feature vector sample extraction module processes the normal vibration signal of the pumped-storage unit and the known fault vibration signal of the pumped-storage unit through the vibration signal multi-dimensional fusion feature vector extraction model, and extracts the fused feature vector samples of the normal vibration signal of the pumped-storage unit and the fused feature vector samples of the known fault vibration signal of the pumped-storage unit.

[0215] The state single classifier construction module constructs a state single classifier for identifying the normal and abnormal states of the pumped-storage unit based on the fused feature vector samples of the normal vibration signal of the pumped-storage unit and based on the variational autoencoder.

[0216] The fault classification unit construction module constructs a fault classification unit for identifying known faults and unknown faults of the pumped storage unit based on the fused feature vector samples of the known fault vibration signal of the pumped storage unit and based on the variational autoencoder.

[0217] The known fault type diagnosis model construction module is based on the fused feature vector samples of the known fault vibration signals of the pumped-storage unit and the support vector machine optimized based on the whale optimization algorithm to construct a known fault type diagnosis model for diagnosing the known fault types of the pumped-storage unit.

[0218] The fault diagnosis module of the pumped-storage unit is used to collect the vibration signal of the pumped-storage unit in real time when the pumped-storage unit is running, and extract the fused feature vector samples of the vibration signal of the pumped-storage unit through the vibration signal multi-dimensional fusion feature vector extraction model; the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the state single classifier to identify the normal state and abnormal state of the pumped-storage unit; when the pumped-storage unit is identified as being in an abnormal state, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the fault classification unit to identify whether the pumped-storage unit is in a known fault or an unknown fault; when the pumped-storage unit is identified as being in a known fault, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the known fault type diagnosis model to diagnose the known fault type of the pumped-storage unit; when the pumped-storage unit is identified as being in an unknown fault, the unknown fault label is output, and the unknown fault type is diagnosed by combining methods including expert experience and spectrum analysis.

[0219] The implementation method of each module in the system is consistent with that described in the method and will not be repeated here.

[0220] The present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the pumped-storage unit fault diagnosis method based on vibration signals as described above, and the pumped-storage unit fault diagnosis system based on vibration signals as described above.

[0221] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of complete hardware embodiment, complete software embodiment, or the embodiment in combination with software and hardware. Moreover, the application can adopt the form of the computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java, C++, Python and literal scripting language JavaScript, etc.

[0222] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0223] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0224] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0225] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0226] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A pumped storage unit fault diagnosis method based on vibration signals, characterized in that: include: Collect normal vibration signals of pumped storage units and known fault vibration signals of pumped storage units; Based on the amplitude domain analysis method, Fourier transform theory and one-dimensional convolutional neural network, a vibration signal multi-dimensional fusion feature vector extraction model is constructed; The normal vibration signal of the pumped-storage unit and the known fault vibration signal of the pumped-storage unit are processed by a vibration signal multi-dimensional fusion feature vector extraction model, and the fusion feature vector samples of the normal vibration signal of the pumped-storage unit and the fusion feature vector samples of the known fault vibration signal of the pumped-storage unit are extracted; According to the fused feature vector samples of the normal vibration signal of the pumped storage unit, a state single classifier for identifying the normal state and abnormal state of the pumped storage unit is constructed based on the variational autoencoder. According to the fused feature vector samples of known fault vibration signals of pumped storage units, a fault classification unit for identifying known and unknown faults of pumped storage units is constructed based on variational autoencoder. According to the fusion feature vector samples of known fault vibration signals of pumped storage units, the support vector machine is optimized based on the whale optimization algorithm to build a known fault type diagnosis model for diagnosing known fault types of pumped storage units. When the pumped-storage unit is in operation, the vibration signal of the pumped-storage unit is collected in real time, and the fused feature vector sample of the vibration signal of the pumped-storage unit is extracted through the vibration signal multi-dimensional fusion feature vector extraction model; the fused feature vector sample of the vibration signal of the pumped-storage unit is input into the state single classifier to identify the normal state and abnormal state of the pumped-storage unit; when the pumped-storage unit is identified as being in an abnormal state, the fused feature vector sample of the vibration signal of the pumped-storage unit is input into the fault classification unit to identify whether the pumped-storage unit is in a known fault or an unknown fault; when the pumped-storage unit is identified as being in a known fault, the fused feature vector sample of the vibration signal of the pumped-storage unit is input into the known fault type diagnosis model to diagnose the known fault type of the pumped-storage unit; when the pumped-storage unit is identified as being in an unknown fault, an unknown fault label is output, and the unknown fault type is diagnosed by combining methods including expert experience and spectrum analysis; The method for constructing the state single classifier comprises: constructing a variational autoencoder model for identifying the normal state and abnormal state of the pumped storage unit according to the fused feature vector samples of the normal vibration signal of the pumped storage unit as the state single classifier; The method for constructing the fault classification unit includes: constructing a variational autoencoder model for identifying the normal state and abnormal state of the pumped storage unit according to the fused feature vector samples of the normal vibration signal of the pumped storage unit as the fault classification unit; In the variational autoencoder model, assuming a given D-dimensional sample time series dataset For each sample x, the encoding process of the variational autoencoder maps complex data to the latent variable space, as shown below: In the formula, is the distribution function of the encoder, z is the potential feature, and x is the input data; The decoding process is to restore the latent variable space data to the original data space, by learning the potential features μ and σ of the data and generating new data samples. The decoding expression is as follows: x'~Dec(z)=p θ (x|z) (8), Where θ is the distribution function of the decoder; x' is the reconstructed data; Variational autoencoder introduces a probability distribution Instead of the true posterior distribution p θ (z|x), then from Sampling from the posterior distribution p θ (z|x) sampling, in order to make and p θ (z|x) are as equal as possible, and KL divergence is used to evaluate the similarity between different distributions. If and only if KL divergence is 0, it means and p θ (z|x) are exactly equal, and by optimizing the parameters Minimize the KL divergence with θ, as follows: make but: In the formula, It is called the variational lower bound, which minimizes By maximizing accomplish; Assume that from Take S samples, variational lower bound Further expressed as: Where z' is the lth sample; p θ (z) is the standard normal distribution; Reparameterization is used to simplify the operation. An auxiliary variable ε is obtained from the standard normal N(0, 1), and then it is calculated using μ and σ. The expression is as follows: When the batch size is large enough, we take S = 1, that is, only one sample data is sampled each time, and formula (11) is expressed as: In the formula, the first term is the approximate posterior distribution With the prior distribution p θ KL divergence of (z); the second term is the reconstruction loss term of the variational autoencoder; the variational autoencoder training process aims to maximize This is equivalent to minimizing the KL divergence and reconstruction loss at the same time, so That is Regularization term that approximates the standard normal; Through the above operations, the variational autoencoder maps the input pumped storage unit normal vibration signal fusion feature data into an ideal Gaussian distribution through the encoder, and then inputs the samples sampled by the Gaussian distribution into the decoder, and the decoder generates reconstructed data. The fault symptom index is constructed by comparing the difference between the input data and the reconstructed data. When the fault symptom index exceeds the threshold, it is judged that the pumped storage unit is in an abnormal state, otherwise it is in a normal state; Through the above operations, the variational autoencoder maps the input known fault vibration signal fusion feature data into an ideal Gaussian distribution through the encoder, and then inputs the samples sampled by the Gaussian distribution into the decoder, and the decoder generates reconstructed data. The fault classification index is constructed by comparing the difference between the input data and the reconstructed data. When the fault classification index exceeds the threshold, it is judged that the pumped storage unit is in an unknown fault, otherwise it is in this type of known fault; for each known fault type of the pumped storage unit, its multi-dimensional vibration signal fusion feature is used as input to independently train the variational autoencoder model, and then a fault classification unit for known and unknown faults of the pumped storage unit composed of multiple variational autoencoders is constructed. Each variational autoencoder focuses on judging whether the corresponding known type of fault occurs in the unit. Through the collaborative work of each variational autoencoder model, the fault classification unit can realize real-time diagnosis of known and unknown faults of the pumped storage unit.

2. The pumped storage unit fault diagnosis method based on vibration signals according to claim 1 is characterized in that: The method for constructing the vibration signal multi-dimensional fusion feature vector extraction model includes: The typical time domain characteristics of the vibration signal of the pumped storage unit are extracted based on the amplitude domain analysis method, and the characteristic parameters of the typical time domain characteristics include: simple statistical characteristic parameters and high-order statistical characteristic parameters; the simple statistical characteristic parameters include mean value, root mean square value, peak-to-peak value; the high-order statistical characteristic parameters include slope, kurtosis, and peak state; Based on Fourier transform theory, typical frequency domain features of the vibration signal of the pumped storage unit are extracted, wherein the typical frequency domain features include: frequency amplitude mean, frequency centroid value, frequency standard deviation, frequency variation coefficient, frequency slope and frequency amplitude root mean square; The implicit features of the vibration signal of the pumped storage unit are mined through a one-dimensional convolutional neural network. The structure of the one-dimensional convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer. The internal spatial features of the data are extracted by alternating the convolutional layer and the pooling layer in a way of local connection and shared weights, so as to obtain the potential deep features of the vibration signal of the pumped storage unit. The specific operations include: The convolution layer extracts feature information by sliding convolution operation of the convolution kernel on the input data. For a given input image I and convolution kernel K, the convolution operation is expressed as: P i,j =∑ m ∑ n I i+m,j+n K m,n (1), Where P i,j is the pixel value of the output feature map at position (i, j), (m, n) represents the row and column index of the convolution kernel. In actual calculation, the vibration signal feature vector is one-dimensional data, and a one-dimensional convolutional neural network is used to perform convolution operation on one dimension of the data matrix; Apply the activation function to the result of the convolution operation and set the ReLU function as the activation function of the one-dimensional convolutional neural network. Its expression is as follows: f ReLU (x)=max(0,x)(2); The pooling layer downsamples the features output by the convolutional layer through a pooling operation, wherein the pooling operation includes maximum pooling and average pooling; In the maximum pooling, the pooling operation selects the maximum value in the input area as the output, and the expression is as follows: In the formula, y l Represents the output value after pooling; Represents the value of the feature map at position (i, j) of the input pooling layer; In the average pooling, the average value in the input area is selected as the output. Assuming the size of the pooling window is m×n, the expression is as follows: The fully connected layer is used to integrate the local information with category distinction in the convolution layer or the pooling layer. Based on the softmax function, the output of the fully connected layer will be converted into the probability distribution of the fault category. The output layer will take the label output corresponding to the maximum probability value as the recognition result of the sample type; When the one-dimensional convolutional neural network performs feature extraction, the model is first pre-trained to obtain the optimal pre-trained model. At this time, sample data is input and the output vector of the fully connected layer is taken out as the multi-dimensional fusion feature vector of the vibration signal of the pumped storage unit.

3. The pumped storage unit fault diagnosis method based on vibration signals according to claim 2 is characterized in that: The method for constructing the known fault type diagnosis model includes: Based on the fused feature vector samples of known fault vibration signals of pumped storage units, the whale optimization algorithm is used to optimize the parameters of the support vector machine. Combining the global optimization ability of the whale optimization algorithm and the classification ability of the support vector machine, a known fault type diagnosis model for diagnosing the known fault types of pumped storage units is constructed. The support vector machine maps the fault data to a high-dimensional space through a nonlinear kernel function, constructs a Lagrangian function to obtain the optimal classification hyperplane; assuming a set of training samples {(x1,y1),(x2,y2),...,(x N ,y N )}, where x i belongs to the input vector, y i is the corresponding output vector. For nonlinear problems, the support vector machine introduces relaxation variables With the penalty coefficient C, the Lagrangian function is defined as follows: In the formula, a i ≥0 is the Lagrange multiplier, w represents the weight, and b is the bias vector; Introducing the radial basis kernel function K(x i ,x j ) = exp(||x i -x j || 2 / (2γ 2 )), optimize the objective function and constraints, the expression is as follows: This problem is a quadratic function optimization problem with a unique solution. The optimal solution a is defined as i * , that is: a i * ≠0 corresponding sample data It is the support vector. Finally, the optimal classification hyperplane of the support vector machine is: Among them, b * is the classification threshold, by solving a i [y i (w·x i )+b]-1=0 to obtain; The whale optimization algorithm is used to optimize the penalty coefficient C and the kernel function parameter γ. The specific operations are as follows: The whale optimization algorithm simulates the hunting behavior of humpback whales, which includes surrounding prey, attacking with a bubble net, and searching for prey; The formula for the encirclement phase is as follows: In the formula, t represents the number of iterations; t max Indicates the maximum number of iterations; coefficient vector; A random number between the range [0, 1]; represents the search particle, Represents the target prey, which is the position vector of the current optimal solution, and is continuously updated in iterations to obtain a better solution; is the convergence factor that decreases linearly from 2 to 0 during the iteration process; The whale optimization algorithm mathematically models the bubble network attack behavior through the shrinking and surrounding mechanism and the spiral update mechanism, which are used alternately according to the probability p, where p is a probability randomly generated from [0, 1]. When p < 0.5, the shrinking and surrounding mechanism is used; and when p ≥ 0.5, the spiral mechanism is used to update the position of the search particle. The expression is as follows: Where d is the shape parameter of the logarithmic spiral, l is a random number between [-1, 1], Represents the distance between the search particle and the current optimal solution; It decreases linearly, resulting in The value of will be Changes between Then the search particles gradually surround the current optimal solution, which belongs to the local optimization stage in the whale optimization algorithm; The distance data will be randomly updated, causing the whale to deviate from the original optimal individual and update the population according to the random position to perform a global search. In the process of searching for prey, its mathematical model is defined as follows: In the formula, represents a search particle randomly selected from the population; Based on the vibration signal feature set of known faults of pumped storage units, combined with the strong global optimization ability of the whale optimization algorithm and the classification advantage of the support vector machine, the known fault types can be identified.

4. A pumped storage unit fault diagnosis system based on vibration signals implementing the method described in any one of claims 1 to 3, characterized in that: It includes a vibration signal collection module, a multi-dimensional fusion feature vector extraction model construction module, a vibration signal fusion feature vector sample extraction module, a state single classifier construction module, a fault classification unit construction module, a known fault type diagnosis model construction module, and a pumped storage unit fault diagnosis module; The vibration signal collection module is used to collect normal vibration signals of the pumped-storage unit and known fault vibration signals of the pumped-storage unit; The multi-dimensional fusion feature vector extraction model construction module constructs a vibration signal multi-dimensional fusion feature vector extraction model based on the amplitude domain analysis method, Fourier transform theory and one-dimensional convolutional neural network; The vibration signal fusion feature vector sample extraction module processes the normal vibration signal of the pumped-storage unit and the known fault vibration signal of the pumped-storage unit through a vibration signal multi-dimensional fusion feature vector extraction model, and extracts the fusion feature vector samples of the normal vibration signal of the pumped-storage unit and the fusion feature vector samples of the known fault vibration signal of the pumped-storage unit; The state single classifier construction module constructs a state single classifier for identifying the normal state and abnormal state of the pumped-storage unit based on the fused feature vector samples of the normal vibration signal of the pumped-storage unit and based on the variational autoencoder; The fault classifier unit construction module constructs a fault classification unit for identifying known faults and unknown faults of the pumped storage unit based on a variational autoencoder according to the fused feature vector samples of the known fault vibration signal of the pumped storage unit; The known fault type diagnosis model construction module constructs a known fault type diagnosis model for diagnosing the types of known faults of the pumped storage unit based on the fused feature vector samples of the known fault vibration signal of the pumped storage unit and the support vector machine optimized based on the whale optimization algorithm; The fault diagnosis module of the pumped-storage unit is used to collect the vibration signal of the pumped-storage unit in real time when the pumped-storage unit is running, and extract the fused feature vector samples of the vibration signal of the pumped-storage unit through the vibration signal multi-dimensional fusion feature vector extraction model; input the fused feature vector samples of the vibration signal of the pumped-storage unit into the state single classifier to identify the normal state and abnormal state of the pumped-storage unit; when the pumped-storage unit is identified as being in an abnormal state, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the fault classification unit to identify whether the pumped-storage unit is in a known fault or an unknown fault; when the pumped-storage unit is identified as being in a known fault, the fused feature vector samples of the vibration signal of the pumped-storage unit are input into the known fault type diagnosis model to diagnose the known fault type of the pumped-storage unit; when the pumped-storage unit is identified as being in an unknown fault, an unknown fault label is output, and the unknown fault type is diagnosed by combining methods including expert experience and spectrum analysis.

5. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the pumped storage unit fault diagnosis method based on vibration signals as described in any one of claims 1 to 3 is implemented.

Images