Dry-type reactor fault diagnosis method and device based on deep learning

Abstract

The invention provides a dry-type reactor fault diagnosis method and device based on deep learning. The method comprises the following steps: based on a preset three-dimensional finite element model of the dry-type reactor, acquiring simulation fault sound signal data by adopting a simulation method so as to construct a source domain data set; training a preset deep learning network by using the source domain data set to obtain an initial fault diagnosis model; fault sound signals of the target dry type reactor are collected to construct a one-dimensional time sequence data set; performing Grubrum angle field coding on each sample in the one-dimensional time sequence data set to generate a corresponding two-dimensional image set; taking the one-dimensional time sequence data set and the two-dimensional image set as a target domain data set, and dividing the target domain data set into a training set and a test set; freezing parameters of a feature extraction network in the initial fault diagnosis model, and performing fine adjustment on other network layer parameters by adopting the training set to obtain an optimal fault diagnosis model; and inputting the test set into the optimal fault diagnosis model to obtain a diagnosis result.

Description

Technical Field

[0001] The present invention relates to the technical field of reactor fault detection, and in particular to a dry-type reactor fault diagnosis method and device based on deep learning. Background Art

[0002] The voiceprint detection method based on the visualization of acoustic signals can realize the non-stop detection of the operating status of the equipment outside without affecting the normal operation of the system; while safely monitoring the status of the equipment, it can also solve the problems of limited detection features and insufficient detection depth of long-distance imaging methods such as infrared and ultraviolet. With the development of artificial intelligence technology, the signal recognition mode based on deep learning has more layers of nonlinear transformation than the traditional information processing mode, and is more suitable for solving complex nonlinear problems in power equipment; it is more powerful in expression and modeling capabilities, and can flexibly and efficiently process massive complex voiceprint information while providing more accurate and intelligent judgment results.

[0003] However, the current condition inspection and maintenance of dry-type reactors is still based on routine testing and inspection during periodic power outages. However, the existing testing and inspection methods and inspection dimensions cannot respond in a timely manner to equipment mechanical, discharge and other defects directly or indirectly caused by various types of dry-type reactor failures, posing a great risk to the safe operation of dry-type reactors. Summary of the invention

[0004] In view of the problem that existing test detection methods and detection dimensions cannot respond in time to equipment mechanical, discharge and other defects directly or indirectly caused by various types of dry-type inductor faults, the present invention provides a dry-type inductor fault diagnosis method and device based on deep learning, which can quickly, automatically and accurately identify the fault type.

[0005] In a first aspect, the present invention provides a dry-type reactor fault diagnosis method based on deep learning, comprising:

[0006] Step 1: Based on the preset three-dimensional finite element model of the dry-type reactor, a simulation method is used to obtain simulated fault sound signal data to construct a source domain data set;

[0007] Step 2: Using the source domain data set to train a preset deep learning network to obtain an initial fault diagnosis model;

[0008] Step 3: Collect the fault sound signal of the target dry-type reactor to construct a one-dimensional time series data set;

[0009] Step 4: Perform Gram angular field encoding on each sample in the one-dimensional time series data set to generate a corresponding two-dimensional image set;

[0010] Step 5: taking the one-dimensional time series dataset and the two-dimensional image set as a target domain dataset, and dividing the target domain dataset into a training set and a test set;

[0011] Step 6: Freeze the parameters of the feature extraction network in the initial fault diagnosis model, and use the training set to fine-tune the parameters of the remaining network layers to obtain the optimal fault diagnosis model;

[0012] Step 7: Input the test set into the optimal fault diagnosis model to obtain the diagnosis result.

[0013] Furthermore, step 1 specifically includes:

[0014] Taking air as the propagation medium, the motion equation and continuity equation that the sound signal follows during propagation are set as follows:

[0015]

[0016] The state equation is

[0017] The wave equation satisfied by the simple harmonic motion sound wave of the reactor is assumed to be:

[0018]

[0019] The acoustic analysis was performed using COMSOL simulation software, the coupled sound sources and boundary conditions were defined, the sound pressure distribution around the reactor was solved, and the simulated fault sound signal data was obtained.

[0020] Furthermore, the feature extraction network adopts a CNN-ResNet dual-channel feature fusion network; the CNN branch is used to extract two-dimensional image features from a two-dimensional image set, and the ResNet branch is used to extract one-dimensional time series features from a one-dimensional time series data set; wherein each convolution layer in the CNN branch is followed by an efficient channel attention ECA module.

[0021] Furthermore, in step 1, the fault types included in the source domain data set include metal foreign body defects, contamination defects, support insulator breakage and bolt loosening.

[0022] Furthermore, in step 3, before collecting the fault sound signal of the target dry-type reactor, it also includes locating the sound source using an adaptive beamforming algorithm based on MVDR; correspondingly, collecting the sound signal at the sound source as the fault sound signal.

[0023] Furthermore, in step 3, a microphone array is used to collect fault sound signals; correspondingly, the MVDR-based adaptive beamforming algorithm is used to locate the sound source, specifically including:

[0024] Transform the signal model received by the microphone array into the frequency domain, then we have:

[0025] X(ω)=AS(ω)+N(ω)

[0026] Set the number of elements in the microphone array to M S , at time n, there are N S signals x1(n),…,x NS (n) respectively from θ1, θ2, …, θ NS The angle of incidence is incident on the microphone array, and the received signal X(n) of the microphone array at time n is defined as MS×1 for:

[0027]

[0028] in, N S The room frequency response at different sound sources, N S Frequency domain representation of sound source signals from different directions;

[0029] Based on the received signal The total output signal y(n) and the total signal power P(θ) of the microphone array are set to:

[0030] y(n)=w H x(n)

[0031] P(θ)=w H (θ)Rw(θ)

[0032] in, Represent the time domain signal and frequency domain signal respectively;

[0033] The following formula is used to describe the fault sound source location problem:

[0034]

[0035] Among them, θ represents the incident angle of the fault sound source to be located;

[0036] Use the Lagrange multiplier method to calculate the weight vector w:

[0037]

[0038] At this point, the total output power of the microphone array becomes:

[0039]

[0040] Then the output power P ACBF The angle corresponding to the maximum value of (θ) is the incident angle of the fault sound signal.

[0041] In a second aspect, the present invention provides a dry-type reactor fault diagnosis device based on deep learning, comprising:

[0042] A source domain data construction module is used to obtain simulated fault sound signal data by a simulation method based on a preset three-dimensional finite element model of a dry-type reactor to construct a source domain data set;

[0043] A pre-training module, used to train a preset deep learning network using the source domain data set to obtain an initial fault diagnosis model;

[0044] a target domain data construction module, for collecting the fault sound signal of the target dry-type reactor to construct a one-dimensional time series data set, performing Gram angle field encoding on each sample in the one-dimensional time series data set to generate a corresponding two-dimensional image set, and using the one-dimensional time series data set and the two-dimensional image set together as a target domain data set, and dividing the target domain data set into a training set and a test set;

[0045] A fine-tuning module is used to freeze the parameters of the feature extraction network in the initial fault diagnosis model and fine-tune the parameters of the remaining network layers using the training set to obtain the optimal fault diagnosis model;

[0046] The diagnosis module is used to input the test set into the optimal fault diagnosis model to obtain the diagnosis result.

[0047] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the first aspect when executing the program.

[0048] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0049] Beneficial effects of the present invention:

[0050] (1) The deep learning-based dry-type reactor fault diagnosis method and device provided by the present invention use an acoustic detection method to extract the dry-type reactor fault sound signal information. The discharge sound signals and pulse current signals corresponding to the fault sound signals under different defect types have certain differences. The acoustic detection method enables the subtle changes in the local discharge signals under various types of faults to be reflected in the time domain characteristics.

[0051] (2) The present invention uses Gram Angular Field (GAF) coding to visualize one-dimensional time series data, which enables the convolutional neural network (CNN) to fully extract the correlation information between acoustic signals and better retain fault characteristics, which is better than time series grayscale feature coding.

[0052] (3) The feature extraction network CNN-ResNet proposed in the present invention adopts dual-channel input and embeds the ECA module in the CNN branch, which helps to extract global key features and improve the accuracy of fault identification.

[0053] (4) The present invention adopts a transfer learning method, which ensures the diagnostic accuracy while making the diagnostic model converge faster, thereby improving the generalization of the model with a small number of samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 One of the flow charts of a dry-type reactor fault diagnosis and identification method based on acoustic signals provided by an embodiment of the present invention;

[0055] Figure 2 A second flow chart of a method for diagnosing and identifying dry-type reactor faults based on acoustic signals provided in an embodiment of the present invention;

[0056] Figure 3 Schematic diagram of a dry-type reactor simulation model provided by an embodiment of the present invention: (a) a three-dimensional model of a reactor; (b) a 1 / 4 simplified model of a reactor; (c) mesh generation; (d) network mass distribution;

[0057] Figure 4 A structural diagram of a deep learning network provided by an embodiment of the present invention;

[0058] Figure 5 A diagram showing the dry-type reactor test principle and platform construction provided by an embodiment of the present invention;

[0059] Figure 6 Schematic diagram of four types of dry-type reactor insulation defects provided by the embodiment of the present invention: (a) contamination defect; (b) metal foreign body defect; (c) loose bolts; (d) loose support insulators;

[0060] FIG. 7( a ) is a timing data diagram corresponding to four insulation defects provided in an embodiment of the present invention;

[0061] FIG. 7( b ) is an image coding diagram corresponding to four insulation defects provided by an embodiment of the present invention;

[0062] Figure 8 A schematic diagram of Gram angle field coding provided by an embodiment of the present invention;

[0063] Fig. 9 It is a schematic diagram of the principle of conventional beamforming;

[0064] Fig.10 A schematic diagram of the structure of a dry-type reactor fault diagnosis and identification device based on acoustic signals provided by an embodiment of the present invention;

[0065] Fig.11 A structural block diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Combination Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a dry-type reactor fault diagnosis method based on deep learning, comprising the following steps:

[0068] S101: Based on a preset three-dimensional finite element model of a dry-type reactor, a simulation method is used to obtain simulated fault sound signal data to construct a source domain data set;

[0069] Specifically, in order to perform simulation analysis of dry-type reactor, its three-dimensional geometric model is first established, such as Figure 3 (a) is shown. Then, in order to improve the calculation efficiency and simplify the calculation complexity, the following simplified assumptions are made: 1) The model does not consider the influence of the support bars, tie rods and gaskets on the windings and the core, simplifying the effect of these structural components on the overall model; 2) The core is simplified to be composed of a whole piece of silicon steel sheet, ignoring the influence of the lamination structure inside the core to simplify the geometric complexity and calculation requirements of the model; 3) The model does not consider the influence of high-order harmonics and eddy current losses on the core, thereby avoiding the influence of the complex electromagnetic effects caused by these factors on the calculation results. The simplified reactor model is shown in Figure 3 (b) as shown.

[0070] In the meshing process, in order to further improve the calculation speed, the model adopts the sweeping meshing method for the winding part. Specifically, the meshing of the winding part is fixed to 3 layers of units to ensure that the consumption of computing resources is minimized while maintaining the calculation accuracy. For the core part, due to its important position in the simulation analysis, free tetrahedral meshing is adopted, and an ultra-fine unit size is selected to capture the complex electromagnetic and mechanical behaviors inside the core. In addition, conventional meshing is used for the rest of the model to ensure the balance between the calculation efficiency and accuracy of the overall model. Finally, the meshing effect of the entire model is as follows: Figure 3 (c) shows a total of 67712 cells. At the same time, the mesh quality follows a normal distribution, indicating that the overall meshing effect is good. Figure 3 (d) as shown.

[0071] When the core is working normally, the differential equation of the magnetic field of the reactor core is:

[0072]

[0073] Where: A is magnetic potential, μ o is the vacuum permeability, which is 4π×10 -7 H / m,μ γ is the relative magnetic permeability, J e is the external current; the saturation magnetization of the core is 2.5×10 5 A / m, magnetic flux density and magnetic field strength satisfy the following relationship:

[0074]

[0075] Where: B is the magnetic flux density, H is the magnetic field intensity.

[0076] The structural field analysis uses the entity in the structural mechanics module, stress-strain (smsld), and the variables are set to (u, v, w). Solve the equation in the transient damped solution domain:

[0077]

[0078] Where ξ is the damping parameter and k is the stiffness matrix.

[0079] For the acoustic pressure field, the acoustic pressure (acpr) in the acoustic module is used, and the equation in the solution domain is the standard acoustic pressure wave equation:

[0080]

[0081] In the formula, ρ o Density kg / m 3 , p is pressure (Pa), c is speed of sound (m / s). The speed of sound is divided into two aspects, the speed of sound in the core and the speed of sound in the air domain.

[0082] Based on the above model, a simulation method is used to obtain simulated fault sound signal data to construct a source domain data set, specifically including: setting the motion equation and continuity equation followed by the sound signal during propagation as:

[0083]

[0084] The state equation is

[0085] Under ideal conditions, the wave equation satisfied by the simple harmonic motion sound wave of the reactor is assumed to be:

[0086]

[0087] Where ρ0 is the fluid density; ν i Represents the vibration velocity of the x, y, and z axes of the reactor in the harmonic response calculation; p is the sound pressure at each point in space; ρ is the medium density increment; c0 is the speed of sound.

[0088] The acoustic analysis was performed using COMSOL simulation software. The material property values ​​are shown in Table 1. The coupled sound source and boundary conditions were defined, the sound pressure distribution around the reactor was solved, and the simulated fault sound signal data was obtained.

[0089] Table 1 Material properties

[0090]

[0091] In order to construct the source domain data set, the collected simulated fault sound signal data is segmented, assuming that every 300 sampling points are taken as a sample, and the sample is labeled with a fault type label, and finally a one-dimensional simulation time series data set consisting of a large number of samples is formed; then, each sample in the one-dimensional simulation time series data set is encoded by Gram angle field (the specific process will be introduced below and will not be repeated here), and the corresponding two-dimensional simulation image set is generated. The one-dimensional simulation time series data set and the two-dimensional simulation image set together constitute the source domain data set.

[0092] S102: Using the source domain data set to train a preset deep learning network to obtain an initial fault diagnosis model;

[0093] Specifically, in this embodiment, the following Figure 4 The deep learning network shown is used as a framework. The deep learning network mainly includes a feature extraction network and a classifier; wherein the feature extraction network adopts a CNN-ResNet dual-channel feature fusion network, and the classifier adopts a fully connected layer. The feature extraction process of the feature extraction network includes: using the CNN branch to extract two-dimensional image features from a two-dimensional image set, and using the ResNet branch to extract one-dimensional time series features from a one-dimensional time series data set; wherein each convolution layer in the CNN branch is followed by an efficient channel attention ECA module. ECA is an efficient channel attention mechanism. In order to fully explore the key information of the channel in feature fusion, this embodiment adds ECA after the CNN convolution layer. This idea reduces the amount of calculation and speeds up the calculation to a certain extent, while retaining the global features between the channels.

[0094] The convolutional layer and ECA of CNN are used to extract the key features of GAF images. The pooling layer is used to downsample the feature map to reduce the size and computation of the feature map. The fully connected layer records the image feature tensor extracted by CNN as T1. ResNet extracts the temporal features of the fault signal. The residual block can help solve the gradient vanishing problem and accelerate the training process. The Dropout layer is used to improve the generalization ability of the network and prevent overfitting. The temporal feature tensor output by the fully connected layer is recorded as T2. Finally, the feature fusion module fuses the two output tensors of the same dimension to obtain a new feature tensor T, which is input into the last fully connected layer for feature transformation to obtain the final classification result.

[0095] Since there is a certain hidden dependency between the one-dimensional time series data of the fault sound signal, and the CNN network focuses on extracting two-dimensional features and ignores such hidden features, this embodiment uses the ResNet deep residual network to extract one-dimensional hidden features. In the residual block, x is the input and the output is H(x). The error between the input and the output is expressed as follows:

[0096] F(x)=H(x)-x

[0097] If the input to output is a shortcut path, the network learns the mapping from x to H(x). At this time, x is known, which is equivalent to learning the mapping from x to H(x). From the residual structure, we can know that the expression of F(x) is:

[0098] F(x)=W2σ(W1x)

[0099] In the above formula, W1 and W2 represent the weights of the first and second layers in the residual network, respectively, and σ is the ReLU function. The ResNet network solves the problem that as the number of network layers increases, the gradient signal gradually decreases, making training difficult.

[0100] S103: collecting fault sound signals of the target dry-type reactor to construct a one-dimensional measured time series data set;

[0101] Specifically, the reactor studied in this embodiment is a dry-type air-core reactor. The experimental principle and platform construction are as follows: Figure 5As shown. Its parameters are as follows: rated voltage is 10kV; power frequency transformer, capacity is 50kVA, primary voltage is 0.4kV, secondary voltage is 100kV; capacitor voltage divider: rated voltage is 100kV, voltage divider ratio is 1000:1, capacitance value is 415.1pF; current limiting resistor: resistance value is 4kΩ; large current generator: input voltage is 380V, output is 63.5V, input current is 44A, output rated current is 262.4A; the filter used in the test is a low-pass filter, the working frequency is 0~20kHz, high-frequency noise above 20kHz will be suppressed, and environmental noise below 100Hz will be directly filtered out by the program during data processing. The maximum rated current of the laboratory circuit breaker is 100A, which meets the laboratory requirements.

[0102] In order to obtain fault sound signals of different fault types for subsequent model fine-tuning, this embodiment mainly provides the production process of four types of reactor insulation defects. Figure 6 As shown, four common defect types are displayed, where defects 1 to 4 represent metal foreign body defects, contamination defects, support insulator breakage, and bolt loosening, respectively.

[0103] Defect 1: During the installation or maintenance of the dry-type inductor, due to incomplete cleaning, metal foreign matter will remain on the inductor. The presence of metal foreign matter will cause local electric field distortion and thus cause local insulation damage. In this embodiment, multiple small copper wires (radius: 0.075 mm) of different lengths (0.5 mm to 1.5 cm) are adsorbed on the dry-type inductor package to simulate the metal foreign matter defect.

[0104] Defect 2: When wet, the conductive ions in the contaminants will move in a directional manner under high voltage and discharge, which is destructive to the encapsulated epoxy resin insulation. In this embodiment, 18.2 mg of diatomaceous earth and 2.6 mg of NaCl are humidified and stirred to fully mix the NaCl and diatomaceous earth, and evenly coated on the designated test area on the dry-type reactor encapsulation.

[0105] Defect 3: The abnormal sound caused by the fracture of the supporting insulator will change the characteristics of the soundprint; in this embodiment, the supporting insulator crack fault is simulated by destroying the supporting insulator at the encapsulation connection to simulate the supporting insulator fracture fault.

[0106] Defect 4: The abnormal noise caused by the loosening of the bolt will change the characteristics of the sound print; in this embodiment, this test simulates the loosening of the bolt by loosening a bolt that fixes the support insulator and the enclosure.

[0107] After the above-mentioned various defect types are produced, in this embodiment, a MEMS microphone sensor array is used to collect the fault sound signal of the dry-type reactor. Similar to the construction process of the one-dimensional simulation time series data set in the above-mentioned source domain data set, for the collected fault sound signals under various fault types, every 300 sampling points are used as a sample, and the corresponding different fault types are used as sample labels, and a one-dimensional measured time series data set containing M samples can be constructed. As shown in Figure 7(a).

[0108] It should be noted that the above-mentioned insulation defect types are only schematic explanations of the fault diagnosis method of the present invention and do not constitute an improper limitation on the technical solution of the present invention. Technical personnel in this field can select several insulation defects according to the specific circumstances of the implementation method and identify the several insulation defects, without being limited to the above-mentioned four insulation defects.

[0109] S104: performing Gram angle field coding on each sample in the one-dimensional measured time series data set to generate a corresponding two-dimensional measured image set;

[0110] Specifically, the GAF (Gramian Angular Field) matrix is ​​a method for time series data analysis that converts a one-dimensional time series into a two-dimensional image for visualization and feature extraction. It includes the following steps:

[0111] (1) The time series data X = {x1, x2, ..., x n}Normalized to the range of [-1,1], the calculation formula is as follows:

[0112]

[0113] (2) Convert the one-dimensional time series from the rectangular coordinate system to the polar coordinate system. The calculation formula is as follows:

[0114]

[0115] Where i is the time node, and N is the constant factor of the regularized polar coordinate system spanning the space. With φ i It is a one-to-one mapping relationship, through the radius r in polar coordinates i Able to retain all information of time series data.

[0116] (3) By introducing the angle difference field GADF and the angle sum field GASF, the angle change is converted into a multi-dimensional matrix. The calculation formula is shown as follows:

[0117]

[0118] Through the three steps of normalization, coordinate system conversion and triangular matrix transformation, the differences in time series data are mapped to the color changes of the image, so that different values ​​of the data correspond to different colors in the image. By selecting an appropriate color mapping scheme, computer vision technology can be applied to the study of time series data. The schematic diagram of Gram's angle field coding is shown in Figure 8 shown.

[0119] Use G i,j Denote the i-th and j-th matrix elements in the time series data, using X i , X j Indicates the different values ​​of two data points. Expanding the two angles and cosine formulas, we get the following formula:

[0120]

[0121] Where σ is the standard deviation and the scaling factor is 1 / 2. i =X i -X j / σ,φ j = 0, when the difference between time series is small, the cosine function value is close to 1; when the difference is large, the cosine function value is close to 0. This feature makes GASF suitable for retaining the overall characteristics of time series, emphasizing the relative position relationship and overall trend in the sequence. Combined with the actual situation, in order to better reflect the fault characteristics, GASF encoding method is used for data processing. As shown in Figure 7(b).

[0122] S105: taking the one-dimensional measured time series data set and the two-dimensional measured image set as a target domain data set, and dividing the target domain data set into a training set and a test set;

[0123] S106: Freeze the parameters of the feature extraction network in the initial fault diagnosis model, and use the training set to fine-tune the parameters of the remaining network layers to obtain the optimal fault diagnosis model;

[0124] Specifically, back-propagation error is used to update and optimize the model during training.

[0125] S107: Input the test set into the optimal fault diagnosis model to obtain a diagnosis result.

[0126] The dry-type inductor fault diagnosis method based on deep learning provided by an embodiment of the present invention processes the collected fault sound signal into a time series data set and a corresponding image set, thereby utilizing a designed dual-channel feature fusion network to extract the one-dimensional time series features and two-dimensional image features of the fault sound signal and perform feature fusion, thereby effectively improving the accuracy of fault type recognition; at the same time, a transfer learning method is adopted to migrate the source domain model that has learned a large number of simulated fault sample features to the target domain model of the measured fault sound signal while ensuring the diagnostic accuracy, so that the diagnostic model has a faster convergence speed and improves the model generalization under a small number of samples.

[0127] On the basis of the above embodiment, before collecting the fault sound signal of the target dry-type inductor, it also includes using an MVDR-based adaptive beamforming algorithm to locate the sound source to shield the noise interference generated by other equipment; correspondingly, the sound signal at the sound source is collected as the fault sound signal.

[0128] As an implementable method, this embodiment uses a microphone array to collect fault sound signals; correspondingly, an adaptive beamforming algorithm based on MVDR is used to locate the sound source. The basic principle is that the time when the signal reaches each element of the microphone array is different, so there is a time difference between the signals received by each element, which is called the delay difference. According to the size of the delay difference, the signals received by each element are compensated in time to align them in the time domain. Finally, they are weighted and added by weighting coefficients to maximize the signal gain in the target direction. S Taking a uniform linear array as an example, assuming that the signal received by the array is a narrowband signal, the principle of conventional beamforming is as follows: Fig. 9 As shown:

[0129] The output of the microphone array is:

[0130]

[0131] The vector is represented as:

[0132] y(t)=w H (θ)x(t) (12)

[0133] Where:

[0134] w(θ)——weighting coefficient of the array,

[0135] Then the output power of the conventional beamforming algorithm is:

[0136] P CBF (θ)=E[y(t) 2 ]=w H (θ)Rw(θ) (13)

[0137] Where: R is the covariance matrix of the array input signal x(t), R = E[x(t)x H (t)].

[0138] The output power of the microphone array can be obtained from formula (3), and the angle corresponding to the maximum output power is the position of the sound source.

[0139] The establishment condition of the adaptive beamforming algorithm based on MVDR is that the cost function of the LCMV algorithm satisfies the constraint w H a(θ) = 1, at this time, the array signal power is minimum after beamforming, and the system signal-to-noise ratio reaches the maximum value. The specific mathematical principle is as follows:

[0140] Transform the signal model received by the microphone array into the frequency domain, then we have:

[0141] X(ω)=AS(ω)+N(ω) (14)

[0142] The number of elements in the microphone array is M S , there are N in the space S Sound source signals from different directions Respectively from The angle of incidence is incident on the microphone array, then the signal received by the microphone array can be expressed as:

[0143]

[0144] in, N S The room frequency response at different sound sources, N S Frequency domain representation of sound source signals from different directions;

[0145] In order to make the array receive only the signal from a certain sound source direction, it is necessary to shield the signals from other directions, so it is necessary to weight the signal received by each array element. Assuming that the weighting coefficient w exists, the total output signal of the microphone array after weighting is:

[0146] y(n)=w H x(n)(16)

[0147] Among them, x(n) and X(n) MS×1 Represent the time domain signal and frequency domain signal respectively;

[0148] The total signal power is:

[0149] P(θ)=w H (θ)Rw(θ) (17)

[0150] Assume that among the many signals received by the microphone array, only the signal incident in the direction of θ1 is the fault sound source signal, and the signals in other directions are interference signals. At this time, the total output signal after beamforming can be written as:

[0151] y(n)=w H x(n)=w H a(θ1)s1(n)+w H z(n) (18)

[0152] Where:

[0153] a(θ1)——weighting coefficient corresponding to the signal in the θ1 direction;

[0154] z(n)——incident signal from other directions.

[0155] Under ideal conditions, the microphone array only receives signals from the direction of the sound source, and the signals in other directions are completely shielded, then:

[0156]

[0157] If the angle of the sound source signal is extended to any angle θ, the problem becomes:

[0158]

[0159] Using the Lagrange multiplier method to calculate the weighted vector w, first construct the cost function J(w), then:

[0160] J(w)=w H Rw-λ w (1-w H a(θ)) (21)

[0161] Taking the weighted vector w as the variable, find the gradient of J(w) and set it to zero, that is:

[0162] ΔJ(w)=2Rw-2λ w a(θ)=0 (22)

[0163] The coefficient w can be obtained from formula (11):

[0164] w=λ w R -1 a(θ) (23)

[0165] Take the conjugate of both ends of the above equation and replace w H Substituting a(θ)=1, we can get:

[0166]

[0167] The weight vector w can be written as:

[0168]

[0169] At this point, the total output power of the microphone array becomes:

[0170]

[0171] According to the derivation, the total power of the microphone array output signal after beamforming is obtained, and the output power P ACBF The angle corresponding to the maximum value of (θ) is the incident angle of the fault sound signal, and its essence is global optimization.

[0172] like Fig.10 As shown, an embodiment of the present invention also provides a dry-type inductor fault diagnosis device based on deep learning, including: a source domain data construction module, a pre-training module, a target domain data construction module, a fine-tuning module and a diagnosis module.

[0173] Among them, the source domain data construction module is used to obtain simulated fault sound signal data by simulation method based on the preset three-dimensional finite element model of dry-type inductor to construct a source domain data set; the pre-training module is used to use the source domain data set to train the preset deep learning network to obtain an initial fault diagnosis model; the target domain data construction module is used to collect the fault sound signal of the target dry-type inductor to construct a one-dimensional time series data set, perform Gram angle field encoding on each sample in the one-dimensional time series data set, generate a corresponding two-dimensional image set, and use the one-dimensional time series data set and the two-dimensional image set as the target domain data set, and divide the target domain data set into a training set and a test set; the fine-tuning module is used to freeze the parameters of the feature extraction network in the initial fault diagnosis model, and use the training set to fine-tune the parameters of the remaining network layers to obtain the optimal fault diagnosis model; the diagnosis module is used to input the test set into the optimal fault diagnosis model to obtain a diagnosis result.

[0174] It should be noted that the deep learning-based dry-type reactor fault diagnosis device provided in the embodiment of the present invention is for the purpose of implementing the above method. Its specific functions can be referred to the above method embodiments, which will not be repeated here.

[0175] In order to verify the effectiveness of the present invention, based on the above embodiments, the present invention also provides the following experimental data.

[0176] The GASF algorithm is used for image encoding in PyCharm software, and batch image conversion is performed on the time series data under four typical defect fault states: metal foreign body defect, loose support insulator, loose bolt, and contamination defect. The training set and test set are divided according to the ratio of 8:2. The specific data processing picture is shown in Figure 7. It can be observed that the GASF image of the partial discharge signal corresponding to the four typical defects has good discrimination.

[0177] Through the simulation method, a total of 1000 groups of acoustic signal samples were obtained for each fault. The source domain data were divided into training set and test set in proportion according to the numbering order. The maximum number of iterations of the back propagation network was set to 500 times, the learning rate was set to 0.01, and the target error was 1×10 -6 ,After 500 iterations, the optimal error is 0.02, and the optimal model parameters are saved and input into the target domain for verification.

[0178] GAF two-dimensional coded images under four different types of faults were used as inputs of different network models for ablation comparison experiments. The running environment of the experiments was PyCharm, the deep learning framework was PyTorch, and different network models were run on a high-performance computer equipped with an Intel corei9-10900x processor and an NVIDIA RTX A5000 graphics card.

[0179] In order to verify the superiority of GAF coding, this paper conducts comparative tests on the two coding methods of GASF and TDS. The network model adopts the CNN-ResNet network proposed in this invention, as well as the existing network CNN and ResNet50. The hardware configuration and software environment are the same as above. A total of 5 classification tests are carried out, and the average accuracy of the five tests is calculated on the training set and the test set, as shown in Table 2. The improved CNN-ResNet model has the highest recognition accuracy in the test set, which can reach 99.50%, which is higher than other network models. And as can be seen from Table 2, the model has a strong generalization ability and no overfitting occurs.

[0180] Table 2 Average accuracy of different encoding methods

[0181]

[0182] In order to better highlight the effectiveness of the improved network model in acoustic signal feature extraction, this experiment also selected 1D-CNN, CNN-LSTM, and CNN-ViT for complete performance comparison, and the evaluation indicators included accuracy, recall, and F1 score. The hardware configuration and the total number of original samples were kept consistent, and a total of five classification tests were conducted to obtain the average, as shown in Table 3.

[0183] Through experimental comparison, it was found that the overall performance of 1D-CNN was poor. The reason may be that when bolt loosening and supporting insulator fracture occurred, the two fault types were both mechanical faults and the timing characteristics were very similar, which made it difficult for 1D-CNN to effectively distinguish them and the overall performance was poor. CNN-LSTM and CNN-ViT performed well, but were not as good as the CNN-ResNet proposed in the present invention, indicating that the scheme of the present invention is more sensitive in identifying similar fault sound signals and more accurate in identifying different fault types.

[0184] Table 3 Average accuracy of different methods

[0185]

[0186] It can be seen from the above experimental data that the present invention can effectively extract the relevant features of the partial discharge signal and can achieve good performance in pattern recognition rate. It has important guiding significance for the operation and maintenance of dry-type inductors and fault analysis and repair, and has positive significance for the development of intelligent detection systems.

[0187] Fig.11 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.11 As shown, the electronic device may include: a processor (processor) 1101, a communication interface (Communications Interface) 1102, a memory (memory) 1103 and a communication bus 1104, wherein the processor 1101, the communication interface 1102, and the memory 1103 communicate with each other through the communication bus 1104. The processor 1101 can call the logic instructions in the memory 1103 to execute the dry-type inductor fault diagnosis method, which includes: based on a preset three-dimensional finite element model of the dry-type inductor, using a simulation method to obtain simulated fault sound signal data to construct a source domain data set; using the source domain data set to train a preset deep learning network to obtain an initial fault diagnosis model; collecting the fault sound signal of the target dry-type inductor to construct a one-dimensional time series data set; performing Gram angle field encoding on each sample in the one-dimensional time series data set to generate a corresponding two-dimensional image set; using the one-dimensional time series data set and the two-dimensional image set as a target domain data set, and dividing the target domain data set into a training set and a test set; freezing the parameters of the feature extraction network in the initial fault diagnosis model, and using the training set to fine-tune the remaining network layer parameters to obtain an optimal fault diagnosis model; inputting the test set into the optimal fault diagnosis model to obtain a diagnosis result.

[0188] In addition, when the logic instructions in the above-mentioned memory 1103 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0189] The embodiment of the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the methods provided by the above-mentioned method embodiments, for example, including: based on a preset three-dimensional finite element model of a dry-type reactor, using a simulation method to obtain simulated fault sound signal data to construct a source domain data set; using the source domain data set to train a preset deep learning network to obtain an initial fault diagnosis model; collecting the fault sound signal of the target dry-type reactor to construct a one-dimensional time series data set; performing Gram angle field encoding on each sample in the one-dimensional time series data set to generate a corresponding two-dimensional image set; using the one-dimensional time series data set and the two-dimensional image set as a target domain data set, and dividing the target domain data set into a training set and a test set; freezing the parameters of the feature extraction network in the initial fault diagnosis model, using the training set to fine-tune the parameters of the remaining network layers to obtain an optimal fault diagnosis model; inputting the test set into the optimal fault diagnosis model to obtain a diagnosis result. .

[0190] An embodiment of the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the methods provided by the above-mentioned method embodiments are implemented, for example, including: based on a preset three-dimensional finite element model of a dry-type inductor, a simulation method is used to obtain simulated fault sound signal data to construct a source domain data set; a preset deep learning network is trained using the source domain data set to obtain an initial fault diagnosis model; the fault sound signal of the target dry-type inductor is collected to construct a one-dimensional time series data set; Gram angle field encoding is performed on each sample in the one-dimensional time series data set to generate a corresponding two-dimensional image set; the one-dimensional time series data set and the two-dimensional image set are used together as a target domain data set, and the target domain data set is divided into a training set and a test set; the parameters of the feature extraction network in the initial fault diagnosis model are frozen, and the parameters of the remaining network layers are fine-tuned using the training set to obtain an optimal fault diagnosis model; the test set is input into the optimal fault diagnosis model to obtain a diagnosis result.

[0191] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dry-type reactor fault diagnosis method based on deep learning, characterized in that: include: Step 1: Based on the preset three-dimensional finite element model of the dry-type reactor, a simulation method is used to obtain simulated fault sound signal data to construct a source domain data set; Step 2: Using the source domain data set to train a preset deep learning network to obtain an initial fault diagnosis model; Step 3: Collect the fault sound signal of the target dry-type reactor to construct a one-dimensional time series data set; Step 4: Perform Gram angular field encoding on each sample in the one-dimensional time series data set to generate a corresponding two-dimensional image set; Step 5: taking the one-dimensional time series dataset and the two-dimensional image set as a target domain dataset, and dividing the target domain dataset into a training set and a test set; Step 6: Freeze the parameters of the feature extraction network in the initial fault diagnosis model, and use the training set to fine-tune the parameters of the remaining network layers to obtain the optimal fault diagnosis model; Step 7: Input the test set into the optimal fault diagnosis model to obtain the diagnosis result.

2. The dry-type reactor fault diagnosis method based on deep learning according to claim 1 is characterized in that: Step 1 specifically includes: Taking air as the propagation medium, the motion equation and continuity equation that the sound signal follows during propagation are set as follows: The state equation is The wave equation satisfied by the simple harmonic motion sound wave of the reactor is assumed to be: The acoustic analysis was performed using COMSOL simulation software, the coupled sound sources and boundary conditions were defined, the sound pressure distribution around the reactor was solved, and the simulated fault sound signal data was obtained.

3. The dry-type reactor fault diagnosis method based on deep learning according to claim 1 is characterized in that: The feature extraction network adopts a CNN-ResNet dual-channel feature fusion network; the CNN branch is used to extract two-dimensional image features from a two-dimensional image set, and the ResNet branch is used to extract one-dimensional time series features from a one-dimensional time series data set; wherein each convolution layer in the CNN branch is connected to an efficient channel attention ECA module.

4. The dry-type reactor fault diagnosis method based on deep learning according to claim 1 is characterized in that: In step 1, the fault types included in the source domain data set include metal foreign body defects, contamination defects, support insulator breakage, and bolt loosening.

5. The dry-type reactor fault diagnosis method based on deep learning according to claim 1, characterized in that: In step 3, before collecting the fault sound signal of the target dry-type reactor, it also includes using an MVDR-based adaptive beamforming algorithm to locate the fault sound source; correspondingly, collecting the sound signal at the fault sound source as the fault sound signal.

6. The dry-type reactor fault diagnosis method based on deep learning according to claim 5 is characterized in that: In step 3, a microphone array is used to collect fault sound signals; correspondingly, the method of using an MVDR-based adaptive beamforming algorithm to locate the fault sound source specifically includes: Transform the signal model received by the microphone array into the frequency domain, then we have: X(ω)=AS(ω)+N(ω) Set the number of elements in the microphone array to M S , at time n, there are N S Signals in different directions Respectively from The angle of incidence is incident on the microphone array, defining the received signal of the microphone array at time n for: in, N S The room frequency response at different sound sources, N S Frequency domain representation of sound source signals from different directions; Based on the received signal The total output signal y(n) and the total signal power P(θ) of the microphone array are set to: y(n)=w H x(n) P(θ)=w H (θ)Rw(θ) Among them, x(n) and Respectively represent the time domain signal and the frequency domain signal; The following formula is used to describe the fault sound source location problem: Among them, θ represents the incident angle of the fault sound source to be located; Use the Lagrange multiplier method to calculate the weight vector w: At this point, the total output power of the microphone array becomes: Then the output power P ACBF The angle corresponding to the maximum value of (θ) is the incident angle of the fault sound signal.

7. A dry-type reactor fault diagnosis device based on deep learning, characterized in that: include: A source domain data construction module is used to obtain simulated fault sound signal data by a simulation method based on a preset three-dimensional finite element model of a dry-type reactor to construct a source domain data set; A pre-training module, used to train a preset deep learning network using the source domain data set to obtain an initial fault diagnosis model; a target domain data construction module, for collecting the fault sound signal of the target dry-type reactor to construct a one-dimensional time series data set, performing Gram angle field encoding on each sample in the one-dimensional time series data set to generate a corresponding two-dimensional image set, and using the one-dimensional time series data set and the two-dimensional image set together as a target domain data set, and dividing the target domain data set into a training set and a test set; A fine-tuning module is used to freeze the parameters of the feature extraction network in the initial fault diagnosis model and fine-tune the parameters of the remaining network layers using the training set to obtain the optimal fault diagnosis model; The diagnosis module is used to input the test set into the optimal fault diagnosis model to obtain the diagnosis result.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images