Residual life prediction method based on intelligent noise analysis and deep recurrent neural network

Abstract

The invention discloses a residual life prediction method based on intelligent noise analysis and a deep recurrent neural network, and the method comprises the steps: carrying out the intelligent noise analysis of an original data signal, collecting the operation data of a thyristor in a working process, and carrying out the noise reduction of the signal; then, features are extracted from the denoised signals, feature dimension reduction fusion is carried out, and dimension reduction processing is carried out on high-dimensional features; and selecting a deep recurrent neural network to establish a residual life model, and designing a specific structure of the deep recurrent neural network. And finally, comparing a calculation result with an experimental result, and performing model evaluation. The problems that traditional modeling methods such as Weibull distribution life modeling and physical failure model life modeling are often based on specific assumptions and conditions, and the generalization ability is limited are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of remaining life modeling, and in particular to a remaining life prediction method based on intelligent noise analysis and deep recurrent neural network. Background Art

[0002] Thyristors play a vital role in power systems such as converter valves. They are key components for achieving AC / DC conversion. Through their on and off functions, converter valves can efficiently complete the conversion between AC and DC. The high power handling capacity of thyristors enables them to withstand large currents and high voltages, ensuring the stable operation of the power system. Their fast switching speed enables the power system to adjust current and voltage more quickly, improving the system's response speed and dynamic performance. At the same time, thyristors have high reliability and can maintain stable performance even in harsh working environments, thereby reducing the risk of failure in the power system.

[0003] Therefore, it is crucial to establish an accurate thyristor remaining life model. By establishing a remaining life model, the performance status of the thyristor at a certain point in the future can be predicted, potential failure risks can be discovered in time, and corresponding preventive measures can be taken to avoid the adverse effects of sudden failures on the power system. The remaining life model can also provide maintenance personnel with information on the health status and expected life of the thyristor, thereby ensuring the continuous and efficient operation of the power system. Traditional thyristor remaining life modeling techniques mainly include Weibull distribution life modeling, physical failure model life modeling, and empirical formula life modeling.

[0004] Traditional thyristor remaining life modeling technologies each have their own unique advantages, but they also have certain limitations. Weibull distribution life modeling mainly relies on a single characteristic parameter to characterize the health status of the thyristor, ignoring other factors that may affect the life; physical failure model life modeling requires an in-depth understanding of the internal structure and failure process of the thyristor, and the modeling process is complex and time-consuming; empirical formula life modeling is usually based on specific conditions and assumptions and is not applicable to all situations. Therefore, in practical applications, it is necessary to comprehensively consider various factors and select appropriate technical methods for thyristor remaining life modeling. Summary of the invention

[0005] Based on the problems raised by the above background technology, the purpose of the present invention is to provide a remaining life prediction method based on intelligent noise analysis and deep recursive neural network. By performing intelligent noise analysis on the original data signal, the operating data of the thyristor during the working process is collected, and the signal is subjected to noise reduction processing; then, features are extracted from the noise-reduced signal, and feature dimensionality reduction fusion is performed, and high-dimensional features are subjected to dimensionality reduction processing. A deep recursive neural network is selected to establish a remaining life model, and the specific structure of the deep recursive neural network is designed. Finally, the calculation results are compared with the experimental results to evaluate the model; the problem that traditional modeling methods such as Weibull distribution life modeling and physical failure model life modeling are often based on specific assumptions and conditions and have limited generalization capabilities is solved.

[0006] The present invention is achieved through the following technical solutions:

[0007] The first aspect of the present invention provides a method for predicting remaining life based on intelligent noise analysis and deep recurrent neural network, comprising the following steps:

[0008] Step A: obtaining operation data of the thyristor during operation, and performing noise reduction processing on the operation data using a smart noise analysis algorithm;

[0009] Step B: extracting time domain features and frequency domain features from the noise-reduced operating data, and performing dimensionality reduction processing on the time domain features and the frequency domain features using an isometric mapping algorithm to form low-dimensional features;

[0010] Step C: establishing a thyristor remaining life model based on a deep recurrent neural network, normalizing the low-dimensional features, and processing the normalized low-dimensional features using the thyristor remaining life model to obtain a denormalized value;

[0011] Step D: When the denormalized value is greater than a preset threshold, the remaining life of the thyristor is calculated using the denormalized value.

[0012] In the above technical solution, a smart noise analysis algorithm is established to perform noise reduction processing on the collected data signal, remove random noise in the signal, and retain useful fault information. The smart noise analysis algorithm can effectively identify and separate the noise components in the operation data of the thyristor during operation, thereby improving the quality of the signal and providing an accurate data basis for subsequent remaining life prediction.

[0013] The purpose of extracting time domain features and frequency domain features from the noise-reduced operating data is to fully reflect the working state of the thyristor. Since time domain features and frequency domain features are high-dimensional features, this method proposes to use an isometric mapping algorithm to reduce the dimensionality of the high-dimensional features while retaining important information as much as possible, forming a low-dimensional feature set that contains rich information.

[0014] A deep recurrent neural network is selected to construct a thyristor remaining life model. The thyristor remaining life model constructed using a deep recurrent neural network is suitable for processing time series data and can effectively avoid the problem of gradient disappearance. For the prediction of the remaining life of the thyristor, it can capture the long-term dependencies in the thyristor data and complete the prediction of the remaining life of the thyristor by processing the long-term dependencies in the thyristor data. The low-dimensional feature vector is normalized and the normalized low-dimensional feature vector is input into the thyristor remaining life model for denormalization prediction. When the denormalization value is greater than the preset threshold, the number of predicted sampling points multiplied by the sum of the vibration signal interval time and the sampling time is the remaining life of the thyristor.

[0015] In an optional embodiment, the operating data includes: on-state voltage drop, reverse recovery charge, turn-off time and voltage resistance.

[0016] In an optional embodiment, the operation data is subjected to noise reduction processing using a smart noise analysis algorithm, comprising the following steps:

[0017] Step A1: Divide the operation data into several segments on average, construct each segment of the operation data into a Hankel matrix, and perform intelligent noise decomposition on the Hankel matrix;

[0018] Step A2: establishing an intelligent noise difference spectrum using the singular values ​​in the Hankel matrix after the intelligent noise decomposition, and finding the effective singular value positions from the intelligent noise difference spectrum to obtain an effective singular value position sequence;

[0019] Step A3: reconstructing a signal for each valid singular value position in the valid singular value position sequence to obtain a reconstructed signal;

[0020] Step A4: Calculate the clearance factor of the reconstructed signal, and select the reconstructed signal with the largest clearance factor as the denoised signal.

[0021] In an optional embodiment, the intelligent noise difference spectrum is established by using the singular values ​​in the Hankel matrix after the intelligent noise decomposition, including the following steps:

[0022] Step a, sequentially obtaining a first singular value and a second singular value in the Hankel matrix after the intelligent noise decomposition; wherein the first singular value and the second singular value are adjacent;

[0023] Step b, calculating the difference between the first singular value and the second singular value to obtain a difference value;

[0024] Step c: All differential values ​​obtained by repeating steps a to b are formed into a sequence to generate a smart noise differential spectrum.

[0025] In an optional embodiment, searching for effective singular value positions from the smart-noise difference spectrum includes the following steps:

[0026] Determining the maximum peak position in the intelligent noise difference spectrum;

[0027] Calculate the absolute value of the difference between the kth difference value and the k+1th difference value in the smart noise difference spectrum. If the absolute value of the difference is greater than the smart noise difference threshold, k is the valid difference value position.

[0028] All valid differential value positions after the maximum peak position are arranged in sequence to generate a valid singular value position sequence.

[0029] It should be noted that if the latter singular value is much smaller than the former one, there will be a peak value at the corresponding position of the difference spectrum of the intelligent noise, and the maximum peak position in the difference spectrum of the intelligent noise is found.

[0030] In an optional embodiment, the time domain features include: mean, variance, standard deviation and skewness factor.

[0031] In an optional embodiment, the frequency domain features include frequency spectrum features and envelope spectrum features.

[0032] In an optional embodiment, using an isometric mapping algorithm to perform dimensionality reduction processing on the time domain features and the frequency domain features includes the following steps:

[0033] The time domain features and the frequency domain features constitute a data set;

[0034] Randomly select two data points from the data set, calculate the Euclidean distance between the two data points, and store the Euclidean distance in a distance matrix;

[0035] Constructing a neighbor graph using the data points and the distance matrix based on a nearest neighbor algorithm;

[0036] Calculating the shortest paths between all pairs of points in the neighbor graph, and storing the shortest paths as geodesic distances in a geodesic distance matrix;

[0037] Using the geodesic distance matrix to construct an inner product matrix, construct a centralization matrix, and using the centralization matrix to perform a dual-centralization process on the inner product matrix to obtain a dual-centralized matrix;

[0038] Performing eigenvalue decomposition on the dual-centered matrix to obtain eigenvalues ​​arranged in descending order;

[0039] Select the eigenvectors corresponding to the first m eigenvalues ​​from the eigenvalues ​​arranged in descending order to construct an eigenvector matrix;

[0040] A diagonal matrix is ​​constructed using the eigenvalues ​​arranged in descending order, and the diagonal matrix and the eigenvector matrix are calculated to obtain low-dimensional embedding coordinates.

[0041] In an optional embodiment, establishing a thyristor remaining life model based on a deep recurrent neural network includes the following steps:

[0042] Initialize the neural network parameters of the thyristor remaining life model;

[0043] Determining the input dimension and time step of the input layer in the thyristor remaining life model;

[0044] Determine the number of layers, number of units and gate structure of the LSTM hidden layer in the thyristor remaining life model; wherein the gate structure includes a forget gate, an input gate, an output gate, a hidden state, a candidate cell state and a cell state;

[0045] The dimension and activation function of the output layer in the thyristor remaining life model are determined.

[0046] In an optional embodiment, the thyristor remaining life model is used to process the normalized low-dimensional features, comprising the following steps:

[0047] Use a real-time recursive learning algorithm to calculate the hidden state of the normalized low-dimensional features at each time step;

[0048] Calculating the network output of the hidden state using the activation function of the output layer;

[0049] Obtain the target output, calculate the loss of the current time step using the network output and the target data, and update the weight using the current time step.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0051] 1. The prediction method of the present invention is based on data-driven, and by learning a large amount of actual operation data, it can better adapt to the prediction of the remaining life of thyristors under different working conditions and environments, thereby improving the generalization ability of the model;

[0052] 2. Deep recurrent neural networks are particularly suitable for processing time series data and can capture long-term dependencies in the data. The present invention takes advantage of this and designs a deep recurrent neural network structure suitable for predicting the remaining life of thyristors, which effectively avoids the gradient vanishing problem and improves the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0054] Figure 1 A schematic diagram of a flow chart of a method for predicting remaining useful life based on intelligent noise analysis and deep recurrent neural network provided in Example 1 of the present invention;

[0055] Figure 2 A schematic diagram of the process structure of the intelligent noise analysis algorithm provided in Example 1 of the present invention;

[0056] Figure 3 A schematic diagram of the structure of a thyristor remaining life model provided in Embodiment 1 of the present invention;

[0057] Figure 4 A schematic diagram of a specific process of the remaining useful life prediction method based on intelligent noise analysis and deep recurrent neural network provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0058] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0059] Example 1

[0060] Figure 1 A schematic diagram of the flow of the remaining life prediction method based on intelligent noise analysis and deep recurrent neural network provided in Example 1 of the present invention, such as Figure 1 As shown in the figure, the remaining life prediction method based on intelligent noise analysis and deep recurrent neural network includes the following steps:

[0061] Step A: obtaining operation data of the thyristor during operation, and performing noise reduction processing on the operation data using a smart noise analysis algorithm;

[0062] Step B: extracting time domain features and frequency domain features from the noise-reduced operating data, and performing dimensionality reduction processing on the time domain features and the frequency domain features using an isometric mapping algorithm to form low-dimensional features;

[0063] Step C: establishing a thyristor remaining life model based on a deep recurrent neural network, normalizing the low-dimensional features, and processing the normalized low-dimensional features using the thyristor remaining life model to obtain a denormalized value;

[0064] Step D: When the denormalized value is greater than a preset threshold, the remaining life of the thyristor is calculated using the denormalized value.

[0065] It should be noted that the intelligent noise analysis algorithm is established to perform noise reduction processing on the collected data signals, remove random noise in the signals, and retain useful fault information. The intelligent noise analysis algorithm can effectively identify and separate the noise components in the operation data of the thyristor during operation, thereby improving the quality of the signal and providing an accurate data basis for subsequent remaining life prediction.

[0066] The purpose of extracting time domain features and frequency domain features from the noise-reduced operating data is to fully reflect the working state of the thyristor. Since time domain features and frequency domain features are high-dimensional features, this method proposes to use an isometric mapping algorithm to reduce the dimensionality of the high-dimensional features while retaining important information as much as possible, forming a low-dimensional feature set that contains rich information.

[0067] A deep recurrent neural network is selected to construct a thyristor remaining life model. The thyristor remaining life model constructed using a deep recurrent neural network is suitable for processing time series data and can effectively avoid the problem of gradient disappearance. For the prediction of the remaining life of the thyristor, it can capture the long-term dependencies in the thyristor data and complete the prediction of the remaining life of the thyristor by processing the long-term dependencies in the thyristor data. The low-dimensional feature vector is normalized and the normalized low-dimensional feature vector is input into the thyristor remaining life model for denormalization prediction. When the denormalization value is greater than the preset threshold, the number of predicted sampling points multiplied by the sum of the vibration signal interval time and the sampling time is the remaining life of the thyristor.

[0068] In an optional embodiment, the operating data includes: on-state voltage drop, reverse recovery charge, turn-off time and voltage resistance.

[0069] It should be noted that key data signals such as on-state voltage drop, reverse recovery charge, turn-off time and withstand voltage capability are collected from the actual operation data of the thyristor using instruments and equipment such as thyristor dv / dt tester and trigger characteristic tester. These data will be used for subsequent analysis.

[0070] In an optional embodiment, the operation data is subjected to noise reduction processing using a smart noise analysis algorithm, comprising the following steps:

[0071] Step A1: Divide the operation data into several segments on average, construct each segment of the operation data into a Hankel matrix, and perform intelligent noise decomposition on the Hankel matrix;

[0072] Step A2: establishing an intelligent noise difference spectrum using the singular values ​​in the Hankel matrix after the intelligent noise decomposition, and finding the effective singular value positions from the intelligent noise difference spectrum to obtain an effective singular value position sequence;

[0073] Step A3: reconstructing a signal for each valid singular value position in the valid singular value position sequence to obtain a reconstructed signal;

[0074] Step A4: Calculate the clearance factor of the reconstructed signal, and select the reconstructed signal with the largest clearance factor as the denoised signal.

[0075] The flow chart of the intelligent noise analysis algorithm is as follows: Figure 2 As shown, the running data signal x containing noise i The data signal is evenly divided into n segments, and a Hankel matrix A is constructed for each segment of the running data signal, where the Hankel matrix A is shown in the following formula (1):

[0076]

[0077] And use formula (2) to perform intelligent noise decomposition on the Hankel matrix A:

[0078] A=UΣV T (2)

[0079] In the above formula, U and V are orthogonal matrices, Σ is a diagonal matrix whose diagonal elements are singular values.

[0080] In an optional embodiment, the intelligent noise difference spectrum is established by using the singular values ​​in the Hankel matrix after the intelligent noise decomposition, including the following steps:

[0081] Step a, sequentially obtaining a first singular value and a second singular value in the Hankel matrix after the intelligent noise decomposition; wherein the first singular value and the second singular value are adjacent;

[0082] Step b, calculating the difference between the first singular value and the second singular value to obtain a difference value;

[0083] Step c: All differential values ​​obtained by repeating steps a to b are formed into a sequence to generate a smart noise differential spectrum.

[0084] It should be noted that the difference between the kth singular value (the first singular value) and the k+1th singular value (the second singular value) is calculated using formula (3), and all the difference values ​​are combined into a sequence such as formula (4) to generate a noise-to-noise difference spectrum; wherein the noise-to-noise difference spectrum is used to represent the difference between adjacent singular values.

[0085] b k =σ k -σ k+1 (3)

[0086] B=[b1,b2,...,b q-1 ] (4)

[0087] where σ k is the kth singular value.

[0088] In an optional embodiment, searching for effective singular value positions from the smart-noise difference spectrum includes the following steps:

[0089] Determining the maximum peak position in the intelligent noise difference spectrum;

[0090] Calculate the absolute value of the difference between the kth difference value and the k+1th difference value in the smart noise difference spectrum. If the absolute value of the difference is greater than the smart noise difference threshold, k is the valid difference value position.

[0091] All valid differential value positions after the maximum peak position are arranged in sequence to generate a valid singular value position sequence.

[0092] It should be noted that if the latter singular value is much smaller than the former one, there will be a peak value at the corresponding position of the difference spectrum of the intelligent noise, and the maximum peak position in the difference spectrum of the intelligent noise is found.

[0093] Set the intelligent noise difference threshold d and find the value that satisfies |b k -b k+1 |>d and all k values ​​after the maximum peak position of the differential spectrum. And all the effective singular value number k values ​​after the maximum peak position of the differential spectrum are calculated, and the effective singular value number sequence is obtained as shown in formula (5):

[0094] K=[k1,k2,...,k p ] (5)

[0095] According to each k i The signal is reconstructed by using the value and formula (6) to obtain p reconstructed signals. The clearance factor of each reconstructed signal is calculated, and the reconstructed signal with the largest clearance factor is selected as the denoising result of the signal segment. The denoising results of each segment are connected in series to obtain the denoising result of the entire signal.

[0096]

[0097] In an optional embodiment, the time domain features include: mean, variance, standard deviation and skewness factor.

[0098] In an optional embodiment, the frequency domain features include frequency spectrum features and envelope spectrum features.

[0099] It should be noted that the time domain feature reflects the statistical characteristics of the denoised signal in the time domain. For the denoised data signal X = (x1, x2, x3...x N ), including mean, variance, standard deviation and skewness factor.

[0100] The mean value u is the average value of the data signal amplitude over a period of time, and its calculation is shown in formula (7):

[0101]

[0102] Variance σ 2 It reflects the dispersion of the random process of the data signal around its mean value and is a dynamic measure. Its calculation is shown in formula (8):

[0103]

[0104] The standard deviation S shows the dispersion of the random process of the denoised data signal around its mean value. It is also a dynamic measure. Its calculation is shown in formula (9):

[0105]

[0106] Skewness factor is a characteristic index that statistically determines whether the signal amplitude falls on the left and right sides of the mean value within a period of time. Its calculation is shown in formula (10):

[0107]

[0108] In formulas (7) to (10), x i represents the data signal after noise reduction, and N represents the number of data signals.

[0109] The frequency domain characteristics reflect the characteristics of the denoised signal in the frequency domain. For the denoised data signal X = (x1, x2, x3...x N ), usually including frequency spectrum characteristics and envelope spectrum characteristics.

[0110] Among them, the spectrum characteristics and envelope spectrum characteristics are fed back in the form of frequency mean and frequency center. The frequency mean ff1 reflects the average amplitude of the data signal in the frequency domain, and its calculation is shown in formula (11). The frequency center ff2 reflects the distribution position of the main frequency in the spectrum, and its calculation is shown in formula (12):

[0111]

[0112]

[0113] In the above formula, the noise-reduced data signal x iAfter Fourier transform or envelope transform, it becomes y j , K represents the number of data signals after Fourier transform or envelope transform, and its corresponding frequency vector is f j .

[0114] In an optional embodiment, using an isometric mapping algorithm to perform dimensionality reduction processing on the time domain features and the frequency domain features includes the following steps:

[0115] The time domain features and the frequency domain features constitute a data set;

[0116] Randomly select two data points from the data set, calculate the Euclidean distance between the two data points, and store the Euclidean distance in a distance matrix;

[0117] Constructing a neighbor graph using the data points and the distance matrix based on a nearest neighbor algorithm;

[0118] Calculating the shortest paths between all pairs of points in the neighbor graph, and storing the shortest paths as geodesic distances in a geodesic distance matrix;

[0119] Using the geodesic distance matrix to construct an inner product matrix, construct a centralization matrix, and using the centralization matrix to perform a dual-centralization process on the inner product matrix to obtain a dual-centralized matrix;

[0120] Performing eigenvalue decomposition on the dual-centered matrix to obtain eigenvalues ​​arranged in descending order;

[0121] Select the eigenvectors corresponding to the first m eigenvalues ​​from the eigenvalues ​​arranged in descending order to construct an eigenvector matrix;

[0122] A diagonal matrix is ​​constructed using the eigenvalues ​​arranged in descending order, and the diagonal matrix and the eigenvector matrix are calculated to obtain low-dimensional embedding coordinates.

[0123] It should be noted that for each pair of points x in the data set i and x j , calculate the Euclidean distance between them: d ij =∥x i -x j ∥These distances are stored in the distance matrix D, where D ij =d ij . Use the nearest neighbor algorithm to construct the neighbor graph G of high-dimensional data. For each point x i , determine its k nearest neighbors. In the neighbor graph, if x j is x i A neighbor point of x is added to G i to x j edge.

[0124] Calculate the shortest paths between all pairs of points in the neighbor graph and use these shortest paths as the geodesic distance d ij G , and stored in the geodesic distance matrix D G middle.

[0125] Furthermore, for non-neighboring points, initialize d ij G =∞; for neighboring points, initialize d ij G =d ij .

[0126] The inner product matrix B is constructed using the geodesic distance matrix according to formula (13):

[0127]

[0128] And construct the centralization matrix H as shown in formula (14):

[0129]

[0130] The inner product matrix B is double-centered using the centering matrix H, and the matrix K is obtained as shown in formula (15):

[0131]

[0132] Where I is the identity matrix and 1 is the all-ones vector.

[0133] Perform eigenvalue decomposition on the matrix K and obtain the eigenvalues ​​λ1,λ2,…,λ in descending order d , select the eigenvectors corresponding to the first m largest eigenvalues ​​to construct the matrix V = [v1, v2, ..., v m ], calculate the low-dimensional embedding coordinates as shown in formula (16):

[0134]

[0135] Among them, Λ m =diag(λ1,λ2,…,λ m ) is a diagonal matrix consisting of the first m largest eigenvalues.

[0136] In an optional embodiment, establishing a thyristor remaining life model based on a deep recurrent neural network includes the following steps:

[0137] Initialize the neural network parameters of the thyristor remaining life model;

[0138] Determining the input dimension and time step of the input layer in the thyristor remaining life model;

[0139] Determine the number of layers, number of units and gate structure of the LSTM hidden layer in the thyristor remaining life model; wherein the gate structure includes a forget gate, an input gate, an output gate, a hidden state, a candidate cell state and a cell state;

[0140] The dimension and activation function of the output layer in the thyristor remaining life model are determined.

[0141] In an optional embodiment, the thyristor remaining life model is used to process the normalized low-dimensional features, comprising the following steps:

[0142] Use a real-time recursive learning algorithm to calculate the hidden state of the normalized low-dimensional features at each time step;

[0143] Calculating the network output of the hidden state using the activation function of the output layer;

[0144] Obtain the target output, calculate the loss of the current time step using the network output and the target data, and update the weight using the current time step.

[0145] It should be noted that the feature data is normalized to ensure that different features have the same weight in subsequent analysis. In this embodiment, two normalization methods, a linear normalization method and a nonlinear normalization method, are proposed. The linear normalization method is a method of linearly transforming data to change the data range as shown in formula (17).

[0146]

[0147]

[0148] Among them, x norm is the data after linear normalization, min(x) is the minimum value of the data, and max(x) is the maximum value of the data.

[0149] The nonlinear normalization method is generally used in cases where data differences are irregular, and can transform the values ​​into a relatively small range, as shown in formula (18), where x is the data after nonlinear normalization.

[0150]

[0151] Initializing the neural network parameters of the thyristor remaining life model includes threshold initialization and weight matrix initialization, wherein for threshold initialization, the threshold is usually set to 0 or a relatively small constant;

[0152] For weight matrix initialization, the Gaussian distribution initialization method is usually used to initialize the weights to a Gaussian (normal) distribution, w~N(μ,σ 2 ), where μ is the mean, σ2 is the variance.

[0153] The remaining life model of thyristor is as follows: Figure 3 As shown, the gate structure in the thyristor remaining life model includes a forget gate, an input gate, an output gate, a hidden state, a candidate cell state, and a cell state as shown in formulas (19) to (24):

[0154] f t =σ(U f x t +W f h t-1 +b f ) (19)

[0155] i t =σ(U i x t +W i h t-1 +b i ) (20)

[0156] o t =σ(U o x t +W o h t-1 +b o ) (twenty one)

[0157] h t =o t *tanh(c t ) (twenty two)

[0158] c t ′=tanh(U c x t +W c h t-1 +b c ) (twenty three)

[0159] c t =f t *c t-1 +i t *c t ' (twenty four)

[0160] Among them, σ is the sigmoid function, tanh is the hyperbolic tangent function, and U f ,U i ,U o Represents the weight matrix between the input layer and the forget gate, input gate, and output gate, respectively. f ,W i ,W o is the weight matrix between the hidden layer and the forget gate, input gate, and output gate, bf ,b i ,b o It is the threshold of forget gate, input gate and output gate.

[0161] The real-time recursive learning (RTRL) algorithm is used to calculate the output and hidden state of each time step. Based on the network output and the actual label, the loss function is calculated as shown in formula (25).

[0162]

[0163] where y t (k) is the network output, d t (k) is the target output and n is the number of output layer units.

[0164] RTRL updates weights at each time step based on information from the current and previous time steps without waiting for the entire sequence to be processed. The weight update is shown in formula (26).

[0165]

[0166] Among them, E t is the loss of the current time step, η is the learning rate, and a suitable learning rate can ensure the stability and speed of training.

[0167] The deep LSTM neural network is trained using the training dataset. The network weights and thresholds are continuously adjusted through the back-propagation algorithm to minimize the loss function. The loss value and the performance indicators of the validation set are monitored during the training process to ensure that the model is not overfitting.

[0168] The specific implementation process of the remaining life prediction method based on intelligent noise analysis and deep recurrent neural network is as follows: Figure 4 As shown:

[0169] The operation data of the thyristor during the working process with a time length of T is collected at intervals of △t until the thyristor fails; a total of n sections of working data are sampled; and the n sections of working data are denoised using the intelligent noise analysis algorithm.

[0170] The time domain features and the frequency domain features are extracted from the n segments of working data after noise reduction processing, and the eigenvalue matrix X is constructed using the time domain features and the frequency domain features.

[0171] The eigenvalue matrix X1 composed of the first n1 sampling points in the eigenvalue matrix X is selected as the training matrix.

[0172] The eigenvalue matrix X and the eigenvalue matrix X1 are reduced in dimension using an isometric mapping algorithm, and the maximum eigenvalue is selected to construct the eigenvector V and the eigenvector V1; wherein the eigenvector V1 is taken as its main component.

[0173] The eigenvector V and the eigenvector V1 are unified using the least squares method to obtain the vector V'.

[0174] After normalizing the feature vector V1, the first p data are selected and input into the thyristor remaining life model to predict the next data, and the thyristor remaining life model is trained through recursive iteration.

[0175] After the thyristor remaining life model is trained, the last p data are used as input data to predict the output at the next moment. After a certain number of iterations, the thyristor remaining life model outputs a denormalized value.

[0176] The denormalized value is compared with the value corresponding to the vector V'. When the denormalized value is greater than the value, the predicted number of sampling points is multiplied by the sum of the data signal interval time and the sampling time △t+T to obtain the remaining life of the thyristor. The number of sampling points refers to the number of data segments sampled in the process of collecting the thyristor key data signal with a time length of T at an interval of △t until the thyristor fails.

[0177] In order to verify the practicality and reliability of the model, the prediction results of the model can be verified through actual experiments. Experimental data are collected and the experimental results are compared and analyzed with the calculation prediction results of the model. Finally, the average percentage error between the predicted value and the actual value is 5.35%.

[0178] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A remaining life prediction method based on intelligent noise analysis and deep recurrent neural network, characterized in that: The steps include: Step A: obtaining operation data of the thyristor during operation, and performing noise reduction processing on the operation data using a smart noise analysis algorithm; Step B: extracting time domain features and frequency domain features from the noise-reduced operating data, and performing dimensionality reduction processing on the time domain features and the frequency domain features using an isometric mapping algorithm to form low-dimensional features; Step C: establishing a thyristor remaining life model based on a deep recurrent neural network, normalizing the low-dimensional features, and processing the normalized low-dimensional features using the thyristor remaining life model to obtain a denormalized value; Step D: When the denormalized value is greater than a preset threshold, the remaining life of the thyristor is calculated using the denormalized value.

2. The remaining life prediction method based on intelligent noise analysis and deep recurrent neural network according to claim 1 is characterized in that: The operation data include: on-state voltage drop, reverse recovery charge, turn-off time and withstand voltage capability.

3. The remaining life prediction method based on intelligent noise analysis and deep recurrent neural network according to claim 1 is characterized in that: The operation data is subjected to noise reduction processing using a smart noise analysis algorithm, comprising the following steps: Step A1: Divide the operation data into several segments on average, construct each segment of the operation data into a Hankel matrix, and perform intelligent noise decomposition on the Hankel matrix; Step A2: establishing an intelligent noise difference spectrum using the singular values ​​in the Hankel matrix after the intelligent noise decomposition, and finding the effective singular value positions from the intelligent noise difference spectrum to obtain an effective singular value position sequence; Step A3: reconstructing a signal for each valid singular value position in the valid singular value position sequence to obtain a reconstructed signal; Step A4: Calculate the clearance factor of the reconstructed signal, and select the reconstructed signal with the largest clearance factor as the denoised signal.

4. The method for predicting remaining useful life based on intelligent noise analysis and deep recurrent neural network according to claim 3 is characterized in that: The intelligent noise difference spectrum is established using the singular values ​​in the Hankel matrix after intelligent noise decomposition, including the following steps: Step a, sequentially obtaining a first singular value and a second singular value in the Hankel matrix after the intelligent noise decomposition; wherein the first singular value and the second singular value are adjacent; Step b, calculating the difference between the first singular value and the second singular value to obtain a difference value; Step c: All differential values ​​obtained by repeating steps a to b are formed into a sequence to generate a smart noise differential spectrum.

5. The method for predicting remaining useful life based on intelligent noise analysis and deep recurrent neural network according to claim 4 is characterized in that: Finding effective singular value positions from the intelligent noise difference spectrum includes the following steps: Determining the maximum peak position in the intelligent noise difference spectrum; Calculate the absolute value of the difference between the kth difference value and the k+1th difference value in the smart noise difference spectrum. If the absolute value of the difference is greater than the smart noise difference threshold, k is the valid difference value position. All valid differential value positions after the maximum peak position are arranged in sequence to generate a valid singular value position sequence.

6. The method for predicting remaining useful life based on intelligent noise analysis and deep recurrent neural network according to claim 1, characterized in that: The time domain characteristics include: mean, variance, standard deviation and skewness factor.

7. The method for predicting remaining useful life based on intelligent noise analysis and deep recurrent neural network according to claim 1, characterized in that: The frequency domain features include spectrum features and envelope spectrum features.

8. The method for predicting remaining useful life based on intelligent noise analysis and deep recurrent neural network according to claim 1, characterized in that: Using an isometric mapping algorithm to perform dimensionality reduction processing on the time domain features and the frequency domain features includes the following steps: The time domain features and the frequency domain features constitute a data set; Randomly select two data points from the data set, calculate the Euclidean distance between the two data points, and store the Euclidean distance in a distance matrix; Constructing a neighbor graph using the data points and the distance matrix based on a nearest neighbor algorithm; Calculating the shortest paths between all pairs of points in the neighbor graph, and storing the shortest paths as geodesic distances in a geodesic distance matrix; Using the geodesic distance matrix to construct an inner product matrix, construct a centralization matrix, and using the centralization matrix to perform a dual-centralization process on the inner product matrix to obtain a dual-centralized matrix; Performing eigenvalue decomposition on the dual-centered matrix to obtain eigenvalues ​​arranged in descending order; Select the eigenvectors corresponding to the first m eigenvalues ​​from the eigenvalues ​​arranged in descending order to construct an eigenvector matrix; A diagonal matrix is ​​constructed using the eigenvalues ​​arranged in descending order, and the diagonal matrix and the eigenvector matrix are calculated to obtain low-dimensional embedding coordinates.

9. The method for predicting remaining useful life based on intelligent noise analysis and deep recurrent neural network according to claim 1, characterized in that: The remaining life model of thyristor is established in the deep recurrent neural network, including the following steps: Initialize the neural network parameters of the thyristor remaining life model; Determining the input dimension and time step of the input layer in the thyristor remaining life model; Determine the number of layers, number of units and gate structure of the LSTM hidden layer in the thyristor remaining life model; wherein the gate structure includes a forget gate, an input gate, an output gate, a hidden state, a candidate cell state and a cell state; The dimension and activation function of the output layer in the thyristor remaining life model are determined.

10. The method for predicting remaining useful life based on intelligent noise analysis and deep recurrent neural network according to claim 9, characterized in that: The low-dimensional features after normalization are processed by using the thyristor remaining life model, comprising the following steps: Use a real-time recursive learning algorithm to calculate the hidden state of the normalized low-dimensional features at each time step; Calculating the network output of the hidden state using the activation function of the output layer; Obtain the target output, calculate the loss of the current time step using the network output and the target data, and update the weight using the current time step.

Images