A vibration detection system and method for on-load tap changers based on wireless sensing and photonic crystal resonator

Abstract

This invention discloses an on-load tap changer vibration detection system and method based on wireless sensing and a photonic crystal resonator, relating to the field of smart grid equipment condition monitoring technology. The method includes: deploying wireless monitoring nodes to collect vibration signals; constructing an adjacency matrix based on mechanical topology connections; calculating the edge weights of the adjacency matrix using the physical distance and mechanical coupling strength between the wireless monitoring nodes; enhancing the vibration signals through optical-vibration coupling using a photonic crystal resonator to generate an amplified signal; compressing the amplified signal using a quantum compressed sensing algorithm to generate a compressed signal; performing orthogonal matching pursuit sparse reconstruction on the compressed signal to extract time-frequency joint features; and inputting the time-frequency joint features into a local convolutional neural network to generate local convolutional neural network parameters. This invention converts mechanical vibration signals into high-sensitivity modulated optical signals through the optical-vibration coupling effect of the photonic crystal resonator, solving the problem of low signal-to-noise ratio in strong noise environments using traditional sensors.

Description

Technical Field

[0001] This invention relates to the field of smart grid equipment condition monitoring technology, and in particular to an on-load tap changer vibration detection system and method based on wireless sensing and photonic crystal resonant cavity. Background Technology

[0002] With the increasing demand for intelligent power systems, online monitoring technology for on-load tap changers (OLTCs) is gradually developing towards multimodal sensing and intelligent diagnosis. In recent years, wireless sensor networks have been widely used in mechanical vibration signal acquisition due to their deployment flexibility and electromagnetic interference resistance. Their topology modeling capabilities provide a new path for multi-node collaborative analysis.

[0003] Traditional vibration signal enhancement mechanisms rely excessively on improving the signal-to-noise ratio of a single sensing node, failing to effectively integrate mechanical topological correlations and multi-physics coupling effects. This results in weak fault features being easily masked by environmental noise, limiting the effective extraction of high-frequency harmonic components. Secondly, existing distributed diagnostic systems generally employ fixed-weight strategies during federated aggregation, failing to consider the joint impact of mechanical connection tightness and signal transmission quality on model convergence, making it difficult to achieve coordinated optimization of local features and global topology. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator to solve the problems of low signal-to-noise ratio and poor model convergence of weak fault features.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a vibration detection method for on-load tap changers based on wireless sensing and a photonic crystal resonator. The method includes: deploying wireless monitoring nodes to collect vibration signals; constructing an adjacency matrix based on mechanical topology connections; calculating the edge weights of the adjacency matrix using the physical distance and mechanical coupling strength between the wireless monitoring nodes; enhancing the vibration signals through optical-vibration coupling using a photonic crystal resonator to generate an amplified signal; compressing the amplified signal using a quantum compressed sensing algorithm to generate a compressed signal; performing orthogonal matching pursuit sparse reconstruction on the compressed signal to extract time-frequency joint features; inputting the time-frequency joint features into a local convolutional neural network to generate local convolutional neural network parameters; calculating federated aggregation weights based on the edge weights and signal-to-noise ratio of the adjacency matrix; fusing the local convolutional neural network parameters; and constructing a global diagnostic model based on a distributed aggregation protocol within a federated learning framework; calculating the Q-value attenuation coefficient of the photonic crystal resonator; inputting the mechanical coupling strength, time-frequency joint features, and Q-value attenuation coefficient into the global diagnostic model to generate fault classification results and maintenance decisions.

[0008] As a preferred embodiment of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator described in this invention, the specific steps for calculating the edge weights of the adjacency matrix based on the physical distance and mechanical coupling strength between wireless monitoring nodes are as follows:

[0009] Determine the mechanical topology of the on-load tap changer, identify the physical connection of the core components, mark the deployment location of the wireless monitoring nodes, and collect vibration signals in real time;

[0010] An adjacency matrix is ​​constructed based on mechanical topology connections, the physical distance between wireless monitoring nodes is measured, and the mechanical coupling strength between each wireless monitoring node is calibrated according to the mechanical connection type between components.

[0011] The edge weights of the adjacency matrix are calculated based on physical distance and mechanical coupling strength.

[0012] As a preferred embodiment of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator described in this invention, the specific steps for generating the compressed signal are as follows:

[0013] The target frequency band for mechanical vibration is defined based on the photonic crystal resonator.

[0014] Integrating a photonic crystal resonator with a wireless monitoring node, the photonic crystal resonator is excited by a laser to generate an optical-vibration coupling effect that converts mechanical vibration into a modulated optical signal, which is then amplified by a bandgap resonance method;

[0015] The amplified and modulated optical signal is converted into an electrical signal using a photodetector, and the real-time Q value of the photonic crystal resonator is recorded.

[0016] A quantum random projection matrix is ​​constructed to linearly project the electrical signal, compressing the vibration signal into a low-dimensional space and generating a compressed signal.

[0017] As a preferred embodiment of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator described in this invention, the specific steps for generating local convolutional neural network parameters are as follows:

[0018] The orthogonal matching pursuit algorithm is initialized, and a sparse basis dictionary is constructed based on the Fourier transform basis and wavelet basis. The vibration signal is then reconstructed based on the quantum random projection matrix and the sparse basis dictionary.

[0019] The reconstructed vibration signal is decomposed by wavelet packet decomposition to generate a time-frequency energy distribution matrix. The instantaneous frequency gradient, energy centroid offset, and energy proportion of each sub-band of the time-frequency energy distribution matrix are calculated to generate joint time-frequency features.

[0020] A local convolutional neural network is constructed by inputting the joint time-frequency features into the local convolutional neural network to generate convolutional kernel weights and fully connected layer parameters.

[0021] As a preferred embodiment of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator described in this invention, the specific steps for calculating the federated aggregation weight based on the edge weights and signal-to-noise ratio of each wireless monitoring node in the adjacency matrix are as follows:

[0022] Extract the edge weights between each wireless monitoring node from the adjacency matrix;

[0023] By combining the signal-to-noise ratio of each wireless monitoring node and using the weight allocation rules defined by the federated learning framework, the federated aggregation weight of each wireless monitoring node is calculated.

[0024] As a preferred embodiment of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator described in this invention, the specific steps for constructing the global diagnostic model using a distributed aggregation protocol based on a federated learning framework are as follows.

[0025] The parameters of the local convolutional neural network uploaded by each wireless monitoring node are weighted and averaged according to the federated aggregation weights to generate global diagnostic model parameters.

[0026] The federated learning framework loads global diagnostic model parameters based on the structure of a local convolutional neural network to construct a global diagnostic model.

[0027] As a preferred embodiment of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator described in this invention, the specific steps for generating fault classification results and maintenance decisions are as follows:

[0028] The initial Q value is defined based on the frequency response curve of the photonic crystal resonator. The Q value attenuation coefficient is calculated based on the difference between the real-time Q value and the initial Q value. The threshold of the Q value attenuation coefficient is defined by experimental calibration.

[0029] The mechanical coupling strength, time-frequency joint characteristics, and Q-value attenuation coefficient are input into the global diagnostic model to generate fault classification results.

[0030] The fault level is determined based on the fault classification results and the Q-value attenuation coefficient threshold, and a maintenance strategy is generated.

[0031] Secondly, this invention provides an on-load tap changer vibration detection system based on wireless sensing and a photonic crystal resonator, comprising a node deployment module, a signal compression module, a feature extraction module, a model building module, and a maintenance decision module. The node deployment module is used to deploy wireless monitoring nodes to collect vibration signals, construct an adjacency matrix based on mechanical topology connections, and calculate the edge weights of the adjacency matrix using the physical distance and mechanical coupling strength between the wireless monitoring nodes. The signal compression module is used to enhance the vibration signal through optical-vibration coupling via the photonic crystal resonator, generating an amplified signal, and compressing the amplified signal based on a quantum compressed sensing algorithm. The system generates a compressed signal; a feature extraction module performs orthogonal matching pursuit sparse reconstruction on the compressed signal, extracts time-frequency joint features, inputs these features into a local convolutional neural network (CNN), and generates CNN parameters; a model building module calculates federated aggregation weights based on the edge weights and signal-to-noise ratio of the adjacency matrix, fuses the CNN parameters, and builds a global diagnostic model based on the distributed aggregation protocol of the federated learning framework; and a maintenance decision module calculates the Q-value attenuation coefficient of the photonic crystal resonator, inputs the mechanical coupling strength, time-frequency joint features, and Q-value attenuation coefficient into the global diagnostic model, and generates fault classification results and maintenance decisions.

[0032] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator as described in the first aspect of the present invention.

[0033] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator as described in the first aspect of the present invention.

[0034] The beneficial effects of this invention are as follows: by converting mechanical vibration signals into high-sensitivity modulated optical signals through the optical-vibration coupling effect of the photonic crystal resonator, the problem of low signal-to-noise ratio of weak fault features and poor data transmission efficiency of high sampling rate in traditional sensors under strong noise environment is solved; at the same time, by dynamically calculating federated aggregation weights based on physical distance and mechanical coupling strength, the parameter fusion process of the global diagnostic model is optimized, solving the problems of slow convergence speed and poor global consistency of the global diagnostic model. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of an on-load tap changer vibration detection method based on wireless sensing and a photonic crystal resonator.

[0037] Figure 2 This is a schematic diagram of an on-load tap changer vibration detection system based on wireless sensing and a photonic crystal resonator.

[0038] Figure 3 A flowchart for generating parameters for a local convolutional neural network.

[0039] Figure 4 A flowchart for generating maintenance strategies.

[0040] Figure 5 This is a flowchart of the operation of a photonic crystal resonator. Detailed Implementation

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0043] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0044] Reference Figures 1-5 This is one embodiment of the present invention, which provides an on-load tap changer vibration detection method based on wireless sensing and a photonic crystal resonator, comprising the following steps:

[0045] S1. Deploy wireless monitoring nodes to collect vibration signals, construct an adjacency matrix based on the mechanical topology connection relationship, and calculate the edge weights of the adjacency matrix through the physical distance and mechanical coupling strength between the wireless monitoring nodes.

[0046] S1.1 Determine the mechanical topology connection relationship of the on-load tap changer, identify the physical connection method of the core components, mark the deployment location of the wireless monitoring node, and collect vibration signals in real time;

[0047] It should be noted that, based on the technical drawings and 3D structure of the on-load tap changer, the physical connection paths of the gearbox, drive shaft, and contact assembly are analyzed to generate a mechanical topology connection diagram. Based on this diagram, the bolted, welded, or snap-fit ​​assembly methods of core components such as gear meshing points, contact switching mechanisms, and spring energy storage devices are identified, and a physical connection method classification list containing connection types and position coordinates is constructed. Based on this classification list, wireless monitoring nodes are deployed at the locations of maximum radial vibration on the gearbox surface, the impact points of contact assembly operation, and the drive shaft bearing housing, among other core components. These nodes are then used to analyze the mechanical topology connection diagram. The physical connection sequence of the core components (e.g., gearbox → drive shaft → spring mechanism → contact group) is used to assign a node number to each wireless monitoring node; the three-dimensional coordinates and node numbers of the wireless monitoring nodes are marked by a laser locator and recorded in the mechanical topology connection diagram; a photonic crystal vibration sensor group is installed on the flange surface of the transformer top cover, and the 1-5kHz vibration signal is enhanced by the built-in photonic crystal resonant cavity; the triaxial accelerometer of the wireless monitoring node is configured to synchronously collect the vibration signal at a sampling rate of 50kHz, and after compression and downsampling, the vibration signal and the coordinates of the wireless monitoring node are uploaded to the edge computing unit in real time via the LoRa protocol.

[0048] S1.2 Construct an adjacency matrix based on mechanical topology connections, measure the physical distance between wireless monitoring nodes, and calibrate the mechanical coupling strength between each wireless monitoring node according to the mechanical connection type between components.

[0049] It should be noted that, based on the mechanical topology connection diagram of the on-load tap changer (the physical connection path of gear-drive shaft-contact), an N×N adjacency matrix is ​​constructed (N is the total number of wireless monitoring nodes). The row and column indices of the adjacency matrix correspond to the node numbers, and the initial value is zero. According to the mechanical topology connection diagram, the positions of the adjacency matrix corresponding to wireless monitoring nodes with direct mechanical connections are marked as to be filled, while the remaining unconnected positions are kept at zero. Using the three-dimensional coordinates of the wireless monitoring nodes, the physical distance between adjacent wireless monitoring nodes is calculated using the Euclidean distance formula, and a physical distance matrix is ​​generated. Based on the list of physical connection methods, mechanical coupling strength is assigned to each type of mechanical connection—bolted, welded, and snap-fit ​​assembly—forming a mechanical coupling strength mapping table. For example, welding is the highest strength benchmark for rigid connections, with a vibration energy transmission efficiency close to 100% (without gaps), so it is assigned a value of 1.0. Due to preload attenuation and micron-level gaps, the vibration transmission efficiency of bolted connections is about 80% to 85% of that of welding, so an intermediate value of 0.8 is taken. Snap-fit ​​connections rely on elastic deformation for locking, and are prone to fatigue failure under long-term vibration, with a transmission efficiency of about 50% of that of welding, so a value of 0.5 is assigned.

[0050] S1.3 Calculate the edge weights of the adjacency matrix based on physical distance and mechanical coupling strength.

[0051] It should be noted that the edge weight values ​​of all positions to be filled in the adjacency matrix are calculated based on the physical distance matrix and the mechanical coupling strength mapping table, and the expression is as follows:

[0052] ;

[0053] in, These are edge weight values, representing the wireless monitoring nodes. and wireless monitoring nodes The strength of the association; It is the mechanical coupling strength, representing the wireless monitoring node. and wireless monitoring nodes The rigidity of the mechanical connection is determined by the type of connection; It refers to the physical distance, representing the wireless monitoring node. and wireless monitoring nodes The Euclidean distance between them is calculated based on the three-dimensional coordinates of the wireless monitoring nodes; It is the row index in the adjacency matrix, representing the node number; It is a column index in the adjacency matrix, representing the node number.

[0054] S2. The vibration signal is enhanced by optical-vibration coupling through a photonic crystal resonator to generate an amplified signal, and the amplified signal is compressed based on a quantum compressed sensing algorithm to generate a compressed signal.

[0055] S2.1 Define the target frequency band for mechanical vibration based on typical mechanical faults of on-load tap changers;

[0056] It should be noted that the refractive index of the basic material of the photonic crystal resonator is determined according to the material handbook, and the effective refractive index is calculated in combination with the photonic crystal structure. Based on the effective refractive index, the lattice constant is calculated using the photonic bandgap center frequency formula. Based on the on-load tap changer fault detection requirements and Fourier transform analysis, the target frequency band is designed as 1-5kHz, taking into account both the characteristic frequency band of mechanical faults and anti-interference performance. The 1-5kHz frequency band covers the 2-3kHz characteristic range of gear loosening faults (based on the American National Standards Institute's gear fault frequency band division and the harmonic enhancement effect of 3 times the gear meshing frequency) and the 4-5kHz sensitive area of ​​contact wear (derived from the national standard "High Voltage AC Circuit Breakers"). The system analyzes the high-frequency characteristics of arc erosion and the fundamental frequency of stress waves at 1–2 kHz (conforming to the transient impact response characteristics of the national standard "Methods for Fatigue Testing of Mechanical Springs"), while avoiding the electromagnetic noise interference zone <1 kHz and the signal attenuation zone >5 kHz. Based on the lattice constant and effective refractive index of the photonic crystal resonator, the bandgap response frequency band of the photonic crystal resonator is calculated using the formula for the center frequency of the photonic crystal bandgap. The transmission spectrum of the photonic crystal resonator is calculated by scanning the target frequency band of mechanical vibration using the finite element method. If the transmission spectrum shows significant attenuation within the target frequency band of mechanical vibration, it indicates that the bandgap response frequency band completely covers the target frequency band of mechanical vibration.

[0057] It should also be noted that Fourier transform extracts fault features by converting vibration signals into frequency domain energy distributions, specifically including two techniques: vibration frequency domain spectrum analysis and high- and low-frequency envelope consistency analysis. According to test data from the condition-based maintenance test procedure for high-voltage AC circuit breakers, the proportion of high-frequency energy in the 10–20kHz range increases to 57.42% when the contacts of the on-load tap changer are worn, while the proportion of low-frequency energy in the 0–10kHz range reaches as high as 96.15% when the contacts are stuck. Under normal conditions, the waveform shape and amplitude of the high- and low-frequency envelopes are highly consistent, but significant deviations occur during faults. However, traditional Fourier transform has inherent defects such as lack of time-frequency localization capability, sensitivity to strong noise environments (weak features are easily submerged under electromagnetic interference >30dB), and inability to synchronously optimize time-frequency resolution. This results in insufficient diagnostic accuracy when analyzing high-frequency harmonics >5kHz, leading to a false negative rate exceeding 22%.

[0058] By enhancing the target frequency band signal of 1–5 kHz through the bandgap resonance of a photonic crystal resonator, a synergistic application mode is formed with traditional Fourier analysis: the high-energy frequency band is initially screened based on the fast Fourier transform (such as a 6–10 dB energy increase in the 2–3 kHz band indicating a loose gear), then the photonic crystal resonator is activated to enhance the energy abnormal frequency band through optical-vibration coupling (gain > 15 dB), and finally the enhanced vibration signal is compressed through a quantum compressed sensing algorithm.

[0059] S2.2 Integrate the photonic crystal resonator with the wireless monitoring node. Excite the photonic crystal resonator with a laser to generate an optical-vibration coupling effect, convert the mechanical vibration into a modulated optical signal, and amplify the modulated optical signal through the bandgap resonance method.

[0060] It should be noted that the photonic crystal resonator is integrated with the triaxial accelerometer of the wireless monitoring node, and the photonic crystal resonator is irradiated by an excitation light with a wavelength of 1550nm emitted by a laser; mechanical vibration causes periodic deformation of the photonic crystal, and the vibration signal is converted into a wavelength-modulated optical signal through the optical-vibration coupling effect; the bandgap resonance effect of the photonic crystal in the target frequency band is used to selectively amplify the modulated optical signal while suppressing out-of-band noise.

[0061] S2.3. Convert the amplified and modulated optical signal into an electrical signal using a photodetector and record the real-time Q value of the photonic crystal resonator.

[0062] It should be noted that the modulated optical signal is converted into a voltage signal by an InGaAs photodetector, and the voltage amplitude and time series are synchronously recorded by a dynamic signal analyzer at a sampling rate of 100kHz. The timestamp alignment is ensured by a hardware trigger signal. A miniature spectrometer is integrated at the output of the photonic crystal resonator to acquire the transmission spectrum of the photonic crystal resonator in real time. The center wavelength and half-width at half-maximum (FWHM) of the resonance peak are extracted by fast Fourier transform, and the real-time Q value is calculated by the formula for the FWHM of the resonance peak.

[0063] S2.4 Construct a quantum random projection matrix to linearly project the electrical signal, compressing the vibration signal into a low-dimensional space to generate a compressed signal.

[0064] It should be noted that the number of rows m of the quantum random projection matrix is ​​calculated based on the real-time Q value to determine the compression dimension; the time window length is set to n=512 according to the ISO standard to construct an m×n quantum random projection matrix; the electrical signal is sampled and quantized by an ADC analog-to-digital converter according to the time window length n to obtain discrete digital signal segments (essentially a digital representation of the vibration signal); the digital signal segments are multiplied with the quantum random projection matrix to achieve linear projection; and the vibration signal is projected to a low-dimensional space through linear projection to generate a compressed signal.

[0065] S3. Perform orthogonal matching pursuit sparse reconstruction on the compressed signal, extract time-frequency joint features, input the time-frequency joint features into the local convolutional neural network, and generate local convolutional neural network parameters.

[0066] S3.1 Initialize the orthogonal matching pursuit algorithm, construct a sparse basis dictionary based on Fourier transform basis and wavelet basis, and reconstruct the vibration signal based on quantum random projection matrix and sparse basis dictionary;

[0067] It should be noted that the residual threshold is defined as 5% through experimental calibration. A joint sparse basis dictionary is constructed by combining Fourier transform basis and wavelet basis. The Fourier transform basis is used to capture the periodic frequency components of the compressed signal, and the wavelet basis is used to extract transient impact features. The transient impact features are used to capture short-lived high-energy pulse signals in the mechanical vibration signal. The compressed signal and the quantum random projection matrix are input into the orthogonal matching pursuit algorithm. The initial residual is set to the compressed signal itself. The iteration counter is reset to zero, and the residual threshold is used as the stopping condition to complete the initialization of the orthogonal matching pursuit algorithm. The basis vector that contributes the most to the residual in the sparse basis dictionary is selected iteratively until the residual energy is lower than the residual threshold or the maximum number of iterations is reached. When the residual energy is lower than the residual threshold, the vibration signal is considered to have been fully restored. The error between the reconstructed vibration signal and the compressed signal is calculated. The allowable error range is set by the mechanical vibration evaluation standard. If the error exceeds the allowable error range, the wavelet basis type is switched. Finally, the reconstructed vibration signal that meets the accuracy requirements is output.

[0068] S3.2 Perform wavelet packet decomposition on the reconstructed vibration signal to generate a time-frequency energy distribution matrix. Calculate the instantaneous frequency gradient, energy centroid offset, and energy proportion of each sub-band of the time-frequency energy distribution matrix to generate joint time-frequency features.

[0069] It should be noted that the reconstructed vibration signal is decomposed into 8 sub-band nodes by performing a 3-level wavelet packet decomposition. The energy value of each sub-band node is calculated to obtain the energy sequence of each sub-band on the time axis. The energy sequence of each sub-band is arranged in a two-dimensional time-frequency configuration to construct the time points, which is listed as the time-frequency energy distribution matrix of the sub-band. Based on the time-frequency energy distribution matrix, the instantaneous frequency gradient of each sub-band is calculated in the time dimension using the Sobel gradient operator to reflect the frequency fluctuation characteristics of the reconstructed vibration signal. In the frequency dimension, the energy centroid offset is calculated using the centroid formula to characterize the stability of the energy distribution. The proportion of energy in each sub-band to the total energy is statistically analyzed to generate an energy proportion vector. The instantaneous frequency gradient, energy centroid offset, and energy proportion vector are concatenated in time alignment to form a time-frequency joint feature.

[0070] S3.3 Construct a local convolutional neural network by inputting the time-frequency joint features into the local convolutional neural network to generate convolutional kernel weights and fully connected layer parameters.

[0071] It should be noted that the hierarchical structure of the local convolutional neural network is designed based on joint time-frequency features, including an input layer, a first local convolutional layer, a second local convolutional layer, a max pooling layer, a flattening layer, a fully connected layer, and an output layer. The activation functions of the first local convolutional layer, the second local convolutional layer, and the fully connected layer are ReLU, and the activation function of the output layer is Softmax. The parameters of the local convolutional neural network (convolutional kernel weights and fully connected layer parameters) are randomly initialized using a normal distribution (mean 0, variance 0.01), and L2 regularization terms are added to the first and second local convolutional layers to prevent overfitting. The structure of the local convolutional neural network is implemented using the Keras framework.

[0072] The time-frequency joint features are input into a local convolutional neural network for forward propagation. After receiving the time-frequency joint features, the first local convolutional layer extracts the local correlation patterns between adjacent time points and feature dimensions through a 3×3 convolutional kernel sliding, and outputs a feature map. The second local convolutional layer further compresses the feature dimensions and enhances the abstraction ability, generating a high-order feature map. The salient feature map is obtained by downsampling through a max pooling layer, flattened, and input into a fully connected layer to calculate the fault probability distribution. The fault category with the highest fault probability is taken as the predicted label. The fault of the current on-load tap changer is determined according to the equipment inspection report and used as the true label. The error between the predicted label and the true label is calculated through the cross-entropy loss function. The Adam optimizer is used for backpropagation to iteratively update the convolutional kernel weights and fully connected layer parameters until the loss function converges or the maximum number of iterations is reached.

[0073] S4. Calculate federated aggregation weights based on the edge weights and signal-to-noise ratio of the adjacency matrix, fuse the parameters of the local convolutional neural network, and construct a global diagnostic model based on the distributed aggregation protocol of the federated learning framework.

[0074] S4.1 Extract the edge weights between each wireless monitoring node from the adjacency matrix, and calculate the federated aggregation weights of each wireless monitoring node by combining the signal-to-noise ratio of each wireless monitoring node and the weight allocation rules defined by the federated learning framework.

[0075] It should be noted that the adjacency matrix contains off-diagonal and main diagonal elements. Main diagonal elements indicate that the current wireless monitoring node has no communication link of its own. Each element represents the quality of the communication link between wireless monitoring nodes; a value of 1 indicates an ideal channel with no packet loss and low latency, while a value of 0 indicates communication interruption. All off-diagonal elements in the adjacency matrix are extracted and arranged by node number to form an edge weight list. All element values ​​are then arranged into a one-dimensional list to generate an edge weight vector. The real-time signal-to-noise ratio (SNR) measurement of each wireless monitoring node is calculated using the SNR formula to generate the signal-to-noise ratio (SNR) vector. The signal-to-noise ratio (SNR) vector is used. Based on the weight allocation rule defined in the federated learning framework, the federated aggregation weight of each wireless monitoring node is determined by multiplying the mean of the edge weight vectors by the normalized SNR vector. For each wireless monitoring node, the arithmetic mean of the edge weight vectors of the communication links with all neighboring wireless monitoring nodes is calculated. The SNR vectors of all wireless monitoring nodes are linearly mapped to the 0-1 interval to generate a normalized SNR vector. The federated aggregation weight of each wireless monitoring node is calculated by multiplying the arithmetic mean of the edge weight vectors of each wireless monitoring node by the normalized SNR vector.

[0076] S4.2. For the local convolutional neural network parameters uploaded by each wireless monitoring node, perform a weighted average according to the federated aggregation weights to generate global diagnostic model parameters.

[0077] It should be noted that the system receives local convolutional neural network parameters (including convolutional kernel weights and fully connected layer parameters) uploaded by all wireless monitoring nodes and simultaneously obtains federated aggregation weights. Within the federated learning framework, the local convolutional neural network parameters uploaded by each wireless monitoring node are standardized. For convolutional layer weights, if the number of output channels of a wireless monitoring node's convolutional kernel is less than the number of channels of the local convolutional neural network's convolutional kernel, a corresponding number of zero channels are added to the output channel dimension; otherwise, redundant channels are truncated. For fully connected layer parameters, if the number of input or output neurons of a wireless monitoring node is inconsistent, zeros are added or truncated in the corresponding dimension of the fully connected layer of the local convolutional neural network. This ensures that the convolutional kernel weights and fully connected parameters of all wireless monitoring nodes have identical dimensions, thereby supporting layer-by-layer weighted average calculation to generate global diagnostic model parameters. Based on the weight values ​​in the federated aggregation weights, a weighted summation is performed layer by layer on the aligned local convolutional neural network parameters to generate the weighted average global diagnostic model parameters.

[0078] S4.3 The federated learning framework loads global diagnostic model parameters based on the structure of local convolutional neural networks to construct a global diagnostic model.

[0079] It should be noted that, by analyzing the local convolutional neural network structure of each wireless monitoring node, a common hierarchical structure shared by the wireless monitoring nodes is dynamically selected as the baseline for the global diagnostic model skeleton. Input layer, convolutional layer, pooling layer, fully connected layer, and output layer are created according to the skeleton baseline structure. The parameters of the convolutional layer are the kernel weights, the pooling layer type is max pooling, and the parameters of the fully connected layer are weight matrices. The rows of the weight matrix represent the number of neurons in the previous layer, and the columns represent the number of neurons in the current layer. A parameter container is created for each layer, with dimensions consistent with the skeleton baseline, and initial values ​​are empty. The global diagnostic model parameters are loaded into the global diagnostic model skeleton and filled into the parameter containers of the corresponding layers of the global diagnostic model skeleton. The rationality of the global diagnostic model is verified through forward propagation, and the complete global diagnostic model is output.

[0080] The global diagnostic model parameters are initialized and distributed to each wireless monitoring node. Each node performs local training based on local vibration signals, updating its local global diagnostic model parameters (including convolutional kernel weights and fully connected layer weights) using stochastic gradient descent or the Adam optimizer. Gradient clipping and dynamic learning rate adjustment are incorporated during local training. After each training round, the wireless monitoring nodes upload the updated global diagnostic model parameters to the server. Upon receiving the updated parameters, the server performs weighted aggregation using a federated averaging algorithm, where weights are dynamically allocated based on the signal-to-noise ratio of the wireless monitoring node data. Differential privacy protection is added to the aggregated global diagnostic model parameters to generate global diagnostic model parameters that meet privacy constraints. Subsequently, the server distributes the updated global diagnostic model parameters to all wireless monitoring nodes and starts the next round of local training. Simultaneously, global diagnostic model validation is performed, calculating the cross-entropy loss. If the validation loss does not decrease for N consecutive rounds, an early stopping mechanism is triggered. Finally, a converged global diagnostic model is output, with training termination condition being reaching the preset maximum number of iterations. The expression for the global diagnostic model is...

[0081] ;

[0082] in, This represents a fault classification score, used to quantify the probability of a fault in the current state of the equipment; Indicates time From the initial moment until the end time Integral operations; This represents the time-frequency joint characteristic function, describing the vibration signal in time. Time-frequency characteristic value at; Represents a complex exponential function. yes It is an imaginary singular number. Represents angular frequency. Indicates a specific moment in time; It is a time differential variable, representing the time-dependent property. Perform infinitesimal division; Indicates the strength of mechanical coupling; This represents the Q-value attenuation coefficient;

[0083] The mechanical coupling strength directly determines the effective transfer efficiency of vibration energy from the internal structure to the outer shell. Strong coupling ensures that vibration signals in the 1.5-15kHz frequency band can maintain sufficient energy base when passing through the steel shell, avoiding the time-frequency joint features being submerged by background noise. The time-frequency joint features rely on a high-Q photonic crystal resonator to achieve narrowband resonance enhancement, but the Q value is affected by mechanical damping, resulting in an attenuation coefficient. There is a nonlinear balance relationship among the three: when the mechanical coupling strength increases to 0.9, the shell stiffness increases, leading to an increase in the Q value attenuation coefficient. At this time, the frequency resolution loss caused by the decrease in Q value needs to be compensated by the time-frequency ridge tracking algorithm. Conversely, under the condition of loose flange bolts, although the Q value attenuation coefficient drops to 0.01, causing the Q value to recover, the time-frequency joint features decrease by 12dB due to insufficient energy transfer efficiency. Therefore, the time-frequency joint features are reconstructed by wavelet packet energy entropy.

[0084] It should also be noted that the wavelet packet energy entropy is calculated based on the time-frequency energy distribution matrix of each wireless monitoring node generated by wavelet packet decomposition of the vibration signal and the Shannon information entropy formula. The expression is as follows:

[0085] ;

[0086] in, Represents the wavelet packet energy entropy; Indicates the wavelet packet decomposition level; Indicates the total number of sub-bands; This represents the sub-band index, ranging from 1 to... ; Indicates the first Energy probability in each sub-band;

[0087] S5. Calculate the Q-value attenuation coefficient of the photonic crystal resonator, input the mechanical coupling strength, time-frequency joint characteristics and Q-value attenuation coefficient into the global diagnostic model, and generate fault classification results and maintenance decisions.

[0088] S5.1 Define the initial Q value based on the frequency response curve of the photonic crystal resonator, calculate the Q value attenuation coefficient based on the difference between the real-time Q value and the initial Q value, and define the Q value attenuation coefficient threshold through experimental calibration.

[0089] It should be noted that the frequency response curve of the photonic crystal resonator was measured using a vector network analyzer under fault-free conditions, and the ratio of the resonant frequency to the half-power bandwidth was extracted as the initial Q value. The Q value attenuation coefficient was generated based on the percentage difference between the real-time Q value and the initial Q value. Typical faults were artificially simulated and the Q value attenuation coefficient data was recorded. The normal fluctuation range was determined by combining the statistical distribution. The critical value exceeding three times the standard deviation of the mean was selected as the Q value attenuation coefficient threshold. If the Q value attenuation coefficient exceeds the Q value attenuation coefficient threshold range, it indicates that the equipment has a fault risk.

[0090] S5.2 Input the mechanical coupling strength, time-frequency joint characteristics and Q-value attenuation coefficient into the global diagnostic model to generate fault classification results;

[0091] It should be noted that the one-dimensional vector flattened from the mechanical coupling strength and time-frequency joint features is concatenated and merged with the Q-value attenuation coefficient to form a comprehensive input vector, which is then standardized using Z-score. This standardized comprehensive input vector is then input into the global diagnostic model. The global diagnostic model extracts high-order patterns of time-frequency features through convolutional layers and fuses the physical correlation between the mechanical coupling strength and the Q-value attenuation coefficient using fully connected layers. Finally, the output layer generates a fault classification score. The Sigmoid function is used to independently output the probability of each fault type, with the fault probability ranging from 0 to 1. Through 3 The principle is to determine the fault probability threshold. Based on the equipment's historical operating data and known faults, the fault classification scores output by the global diagnostic model are statistically analyzed and converted into fault probabilities. The mean and standard deviation are calculated, and the fault probability threshold is determined as mean + 3 standard deviations. If the probability of a certain type of fault exceeds the fault probability threshold, a fault is determined to exist. For example, gear crack 0.9 and lubrication failure 0.6 (the fault probability thresholds for both are 0.5), so it is determined as "gear crack + lubrication failure".

[0092] S5.3 Determine the fault level based on the fault classification results and the Q-value attenuation coefficient threshold, and generate a maintenance strategy.

[0093] It should be noted that the fault level is determined based on the fault probability and the Q-value decay coefficient threshold. When the fault probability is 30%-70% and the Q-value decay coefficient is <0.2, it indicates a minor fault; when the fault probability is 70%-90% and 0.2≤Q-value decay coefficient<0.4, it indicates a moderate fault; and when the fault probability is >90% and the Q-value decay coefficient is ≥0.4, it indicates a severe fault. Based on the fault level and real-time equipment operating conditions (such as load, temperature, and running time), rules matching the current fault type and level are retrieved from the maintenance strategy knowledge base, such as "severe fault + high load → immediate shutdown" and "moderate fault + low wear → planned maintenance," to generate specific maintenance strategies (such as component replacement, lubrication, and enhanced monitoring). Fault information is generated based on the fault type, fault level, fault location, and fault time. The maintenance strategy is bound to the fault information and pushed to the operation and maintenance management unit through an interface, while the equipment health record is updated synchronously, recording the fault handling trajectory.

[0094] This embodiment also provides an on-load tap changer vibration detection system based on wireless sensing and photonic crystal resonator, including: a node deployment module, a signal compression module, a feature extraction module, a model building module, and a maintenance decision module;

[0095] The node deployment module is used to deploy wireless monitoring nodes to collect vibration signals, construct an adjacency matrix based on the mechanical topology connection relationship, and calculate the edge weights of the adjacency matrix through the physical distance and mechanical coupling strength between wireless monitoring nodes.

[0096] The signal compression module is used to enhance the vibration signal through optical-vibration coupling via a photonic crystal resonant cavity, generate an amplified signal, and compress the amplified signal based on a quantum compressed sensing algorithm to generate a compressed signal.

[0097] The feature extraction module is used to perform orthogonal matching pursuit sparse reconstruction on the compressed signal, extract time-frequency joint features, input the time-frequency joint features into the local convolutional neural network, and generate local convolutional neural network parameters.

[0098] The model building module is used to calculate federated aggregation weights based on the edge weights and signal-to-noise ratio of the adjacency matrix, fuse local convolutional neural network parameters, and build a global diagnostic model based on the distributed aggregation protocol of the federated learning framework.

[0099] The maintenance decision module is used to calculate the Q-value attenuation coefficient of the photonic crystal resonator. It inputs the mechanical coupling strength, time-frequency joint characteristics, and Q-value attenuation coefficient into the global diagnostic model to generate fault classification results and maintenance decisions.

[0100] This embodiment also provides a computer device applicable to the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonator proposed in the above embodiment.

[0101] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0102] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the on-load tap changer vibration detection method based on wireless sensing and a photonic crystal resonator as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0103] In summary, this invention solves the problems of low signal-to-noise ratio of weak fault features and poor data transmission efficiency of high sampling rate in traditional sensors by converting mechanical vibration signals into high-sensitivity modulated optical signals through the optical-vibration coupling effect of photonic crystal resonators. At the same time, by dynamically calculating federated aggregation weights based on physical distance and mechanical coupling strength, the parameter fusion process of the global diagnostic model is optimized, solving the problems of slow convergence speed and poor global consistency of the global diagnostic model.

[0104] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting vibration of an on-load tap changer based on wireless sensing and a photonic crystal resonator, characterized in that: Comprising, The wireless monitoring nodes collect vibration signals, an adjacency matrix is constructed according to the mechanical topology connection relationship, and the edge weight of the adjacency matrix is calculated according to the physical distance and mechanical coupling strength between the wireless monitoring nodes, and the specific steps are as follows, Determine the mechanical topology connection relationship of the on-load tap changer, identify the physical connection mode of the core component, mark the deployment position of the wireless monitoring node, and collect vibration signals in real time; Based on the mechanical topology connection relationship, an adjacency matrix is constructed, the physical distance between the wireless monitoring nodes is measured, and the mechanical coupling strength between the wireless monitoring nodes is calibrated according to the mechanical connection type between the components; Based on the physical distance and the mechanical coupling strength, the edge weight of the adjacency matrix is calculated; The vibration signal is optically coupled and enhanced by the photonic crystal resonant cavity to generate an amplified signal, and the amplified signal is compressed based on the quantum compressed sensing algorithm to generate a compressed signal, and the specific steps are as follows, According to the photonic crystal resonant cavity, the mechanical vibration target frequency band is defined; The photonic crystal resonant cavity is integrated with the wireless monitoring node set, the photonic crystal resonant cavity is excited by laser, the mechanical vibration is converted into a modulated optical signal through the optical- mechanical coupling effect, and the modulated optical signal is amplified through the band gap resonance method; The amplified modulated optical signal is converted into an electrical signal by a photodetector, and the real-time Q value of the photonic crystal resonant cavity is recorded; A quantum random projection matrix is constructed to linearly project the electrical signal, compress the vibration signal into a low-dimensional space, and generate a compressed signal; The compressed signal is subjected to orthogonal matching pursuit sparse reconstruction, time-frequency joint features are extracted, the time-frequency joint features are input into a local convolutional neural network, and local convolutional neural network parameters are generated, and the specific steps are as follows, Initialize the orthogonal matching pursuit algorithm, construct a sparse base dictionary based on the Fourier transform base and the wavelet base, and reconstruct the vibration signal based on the quantum random projection matrix and the sparse base dictionary; The reconstructed vibration signal is subjected to wavelet packet decomposition to generate a time-frequency energy distribution matrix, the instantaneous frequency gradient, energy barycenter offset and energy proportion of each sub-band of the time-frequency energy distribution matrix are calculated, and time-frequency joint features are generated; A local convolutional neural network is constructed, the time-frequency joint features are input into the local convolutional neural network, and convolution kernel weights and fully connected layer parameters are generated; Based on the edge weight of the adjacency matrix and the signal-to-noise ratio, the federal aggregation weight is calculated, and the specific steps are as follows, The edge weight between each wireless monitoring node is extracted from the adjacency matrix; The signal-to-noise ratio of each wireless monitoring node is combined, the federal aggregation weight of each wireless monitoring node is calculated through the weight distribution rule defined by the federal learning framework, and the specific steps are as follows, Fuse the local convolutional neural network parameters, and construct a global diagnostic model based on the distributed aggregation protocol of the federal learning framework; The Q value decay coefficient of the photonic crystal resonant cavity is calculated, the mechanical coupling strength, time-frequency joint features and Q value decay coefficient are input into the global diagnostic model, and fault classification results and maintenance decisions are generated, and the specific steps are as follows, An initial Q value is defined according to the frequency response curve of the photonic crystal resonant cavity, a Q value decay coefficient is calculated according to the difference between the real-time Q value and the initial Q value, and the Q value decay coefficient threshold is defined by experimental calibration method; The mechanical coupling strength, time-frequency joint features and Q value decay coefficient are input into the global diagnostic model to generate fault classification results. According to the fault classification result and the Q value attenuation coefficient threshold, the fault level is determined, and a maintenance strategy is generated.

2. The wireless sensing and photonic crystal resonator cavity based on-load tap changer vibration detection method of claim 1, wherein: The distributed aggregation protocol based on the federated learning framework builds a global diagnostic model, and the specific steps are as follows, The local convolutional neural network parameters uploaded by each wireless monitoring node are weighted and averaged according to the federated aggregation weight to generate global diagnostic model parameters. The federated learning framework loads the global diagnostic model parameters based on the structure of the local convolutional neural network to build a global diagnostic model.

3. A load tap changer vibration detection system based on wireless sensing and photonic crystal resonator, based on the load tap changer vibration detection method based on wireless sensing and photonic crystal resonator in any one of claims 1-2, characterized in that: It comprises, A node deployment module, a signal compression module, a feature extraction module, a model construction module, and a maintenance decision module. The node deployment module is used to deploy wireless monitoring nodes to collect vibration signals, construct an adjacency matrix according to the mechanical topological connection relationship, and calculate the edge weight of the adjacency matrix through the physical distance and mechanical coupling strength between wireless monitoring nodes. The signal compression module is used to perform light-vibration coupling enhancement on the vibration signal through the photonic crystal resonant cavity to generate an amplified signal, and compress the amplified signal based on the quantum compressive sensing algorithm to generate a compressed signal. The feature extraction module is used to perform orthogonal matching pursuit sparse reconstruction on the compressed signal, extract time-frequency joint features, input the time-frequency joint features into a local convolutional neural network, and generate local convolutional neural network parameters. The model construction module is used to calculate the federated aggregation weight based on the edge weight of the adjacency matrix and the signal-to-noise ratio, fuse the local convolutional neural network parameters, and construct a global diagnostic model based on the distributed aggregation protocol of the federated learning framework. The maintenance decision module is used to calculate the Q value attenuation coefficient of the photonic crystal resonant cavity, input the mechanical coupling strength, time-frequency joint features, and Q value attenuation coefficient into the global diagnostic model, and generate fault classification results and maintenance decisions.

4. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor executes the computer program to realize the steps of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonant cavity according to any one of claims 1-2.

5. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the on-load tap changer vibration detection method based on wireless sensing and photonic crystal resonant cavity according to any one of claims 1-2.

Images