A fault diagnosis method for on-load tap changer based on lightweight YOLO11

Abstract

A method for fault diagnosis of on-load tap changers based on lightweight YOLO11 is used to solve the problem of low efficiency in mechanical fault diagnosis of traditional on-load tap changers. The method includes the following steps: converting the vibration signals of the motor startup phase and the switching switch action phase into high-resolution time-frequency images through continuous wavelet transform CWT to fully reflect the time and frequency characteristics of the signal; using StarNet to replace the YOLO model backbone network, and enhancing the feature extraction capability while maintaining the compactness of the network through star operations; designing an efficient joint convolution detection head EJCD, and realizing model lightweighting and multi-scale target detection through group normalization and convolution parameter sharing. By combining the time-frequency image representation capability of continuous wavelet transform with the efficient target detection framework of deep learning, the present invention not only overcomes the limitations of traditional methods in non-stationary signal processing, but also significantly improves the diagnosis efficiency and model performance.

Description

Technical Field

[0001] The invention relates to the technical field of on-load tap changer fault diagnosis, and in particular to an on-load tap changer fault diagnosis method based on lightweight YOLO11. Background Art

[0002] The transformer stabilizes the voltage within the normal floating range through voltage regulation to ensure the safe and stable operation of the power system. Its operating status is directly related to the stability and reliability of the power supply and distribution system. At present, on-load voltage regulation technology has gradually become the main voltage regulation method for transformers. Frequent voltage switching puts higher requirements on the reliability of on-load tap changers (OLTCs).

[0003] OLTC needs to rely on the close cooperation of complex components such as tap selectors, switching switches and drive mechanisms to achieve step-by-step action in the switching sequence. Its operating reliability is gradually reduced with the large amount of electrical and mechanical shocks after transportation, assembly and commissioning. According to the statistical analysis of fault types, OLTC faults account for 20% of the total faults of on-load tap-changing transformers, of which mechanical faults account for as high as 70%. OLTC under abnormal working conditions poses a huge safety hazard to the safe and stable operation of the transformer. When OLTC is in a stationary state, it is difficult to find early mechanical hazards or defects in the internal system of the switch based solely on the number of switching times or conventional electrical parameter detection. The vibration signal generated by OLTC during the shifting process contains rich equipment status information. How to extract effective features from complex vibration signals to reflect the operating status of OLTC has attracted widespread attention from researchers. Scholars mainly analyze vibration signals from the time domain, frequency domain and time-frequency domain to extract signal features. Time domain analysis mainly extracts statistical features such as the mean, root mean square, and peak value of vibration signals, which is difficult to reveal the frequency components and energy distribution of the signal, and cannot characterize the periodicity and frequency characteristics in the signal, resulting in information loss. Frequency domain analysis converts the signal into a spectrum through Fourier transform. The vibration signal contains mutations or transient components, which appear as the superposition of multiple frequency components in the frequency domain. Frequency domain analysis cannot reflect the distribution of these components in time. The time-frequency domain can simultaneously reveal the time and frequency characteristics of the signal, associate the frequency components of the signal with the time distribution, and show the dynamic change characteristics of the frequency over time, making up for the insufficiency of frequency domain analysis in reflecting time information. Commonly used time-frequency domain analysis methods include: short-time Fourier transform STFT, short-time Wigner-Ville distribution WVD, empirical mode decomposition EMD and its derivative methods, and wavelet transform and its derivative methods. STFT is simple to use and has high computational efficiency, but its resolution is limited and is not suitable for non-stationary signals. WVD will have cross-interference terms when analyzing multi-component signals, and this defect seriously affects the detection performance. EMD decomposes the original signal into multiple intrinsic mode functions (IMFs) to extract signal information at different frequencies or feature levels. However, this type of method cannot effectively solve the problem of modal aliasing generated during the decomposition process, and all involve iterative calculations and parameter optimization processes, resulting in long calculation times when the amount of data is large, making it difficult to meet the real-time requirements of online monitoring of equipment status. Wavelet transform has good time and frequency localization characteristics, and can accurately represent the local characteristics of the signal in the time and frequency domain. It is suitable for analyzing signals with instantaneous changes or local characteristics. Compared with discrete wavelet transform (DWT), continuous wavelet transform (CWT) has high time and frequency resolution, can capture very subtle changes in the signal, and has better processing capabilities for instantaneous changes and non-stationary signals. CWT converts 1D vibration signals into 2D time-frequency images. The color or intensity in the image represents the energy of the signal at that moment and frequency, reflecting the characteristics of the signal at different times and frequency scales.Compared with other methods, CWT has higher computational efficiency and can process large-scale data in a short time. By applying image recognition technology, key features in time-frequency diagrams can be quickly extracted, thereby achieving rapid judgment of equipment status.

[0004] Driven by big data and intelligent manufacturing, modern industry is accelerating its transformation from traditional manufacturing to intelligent manufacturing. The rapid development of computer vision technology has opened up new research directions for signal analysis and equipment status monitoring. In the field of power equipment status monitoring, there are still few studies based on the time-frequency characteristics of 2D images of vibration signals. As an efficient image recognition algorithm, YOLO has broad application potential due to its strong real-time performance and high precision. In recent years, with the further optimization of computing resources and the promotion of lightweight neural networks, the application of YOLO series models in equipment fault diagnosis has been initially explored. However, how to combine the time-frequency characteristics of vibration signals with the YOLO model for deep fusion and develop lightweight intelligent diagnosis methods for resource-constrained scenarios is still a direction worthy of in-depth research. Summary of the invention

[0005] The object of the present invention is to provide a fault diagnosis method for on-load tap changer based on lightweight YOLO11, so as to solve the problem of low efficiency of mechanical fault diagnosis of traditional on-load tap changer.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a fault diagnosis method of an on-load tap changer based on a lightweight YOLO11, comprising the following steps.

[0007] S1. The vibration signals of the motor startup phase and the switching phase are converted into high-resolution time-frequency images through continuous wavelet transform (CWT) to fully reflect the time and frequency characteristics of the signal. Suppose the signal is , the mother wavelet is , then the continuous wavelet transform of the signal It is expressed as: ; Among them, a is the scale factor, which controls the expansion and contraction of the mother wavelet; b is the translation factor, which controls the translation of the mother wavelet; is the complex conjugate of the mother wavelet function.

[0008] S2. StarNet is used to replace the YOLO model backbone network, and star operations are used to enhance feature extraction capabilities while maintaining network compactness.

[0009] The StarNet architecture uses star operations, that is, element-by-element multiplication to map low-dimensional inputs to high-dimensional nonlinear feature space; based on this, StarNet's core computing unit Star Blocks is built. StarNet uses efficient star operations to achieve efficient representation of information in high-dimensional feature space with minimal network structure and computational effort.

[0010] S3. Design an efficient joint convolutional detection head EJCD to achieve model lightweight and multi-scale target detection through group normalization and convolution parameter sharing.

[0011] First, the batch normalization (BN) in the convolutional layers of the detection head is replaced by group normalization (GN).

[0012] Secondly, through joint convolution, convolution modules are shared among multiple feature layers to further optimize the lightweight effect of the model.

[0013] Finally, in order to deal with the problem of inconsistent scales of targets detected by each detection head, the Scale layer is used to scale the features.

[0014] Furthermore, Cmor wavelet is selected as the mother wavelet, which is the complex form of Morlet wavelet. The real part describes the frequency components of the signal, and the imaginary part provides phase information.

[0015] Furthermore, for a single-layer neural network, the star operation is expressed as: .

[0016] in, , is the combination matrix of weight matrix W1 and bias term B; , is the combination matrix of weight matrix W2 and bias term B; for The transpose of for The transpose of is the input feature matrix The combined matrix after adding a feature dimension; Represents element-by-element multiplication. For a single output channel and a single input element, define ( is the weight vector, is the input vector), the star operation is expanded as: .

[0017] in, and Used to index the input feature channel, that is, the elements in the weight vector; is a coefficient that represents the relationship or weighting between each pair of input feature channels, and is specifically defined as: .

[0018] The star operation expansion will contain The term is the number of output feature dimensions. When d is large enough ; As the number of layers increases, the dimension of the implicit feature space will increase exponentially: .

[0019] After the n-th layer of stacking is completed as shown in formula (5), the feature space dimension of the output of the n-th star operation will reach .

[0020] The beneficial effect of the present invention is that the fault diagnosis method of on-load tap changer based on lightweight YOLO11 proposed by the present invention not only overcomes the limitations of traditional methods in non-stationary signal processing, but also significantly improves the diagnostic efficiency and model performance by combining the time-frequency image representation capability of continuous wavelet transform with the efficient target detection framework of deep learning. The method is suitable for real-time online monitoring in resource-constrained scenarios, provides a new research direction for intelligent diagnosis of on-load tap changers, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is the structural diagram of the on-load tap changer.

[0022] Figure 2 This is a schematic diagram of the time-frequency domain conversion of the vibration signal of a certain operation of the on-load tap changer.

[0023] Figure 3 This is the network structure diagram of the Star-EJCD lightweight YOLO model.

[0024] Figure 4 This is the network structure diagram of StarNet.

[0025] Figure 5 This is the structural diagram of the EJCD detection head.

[0026] Figure 6 This is the vibration signal waveform of the entire process of OLTC shifting gears under normal working conditions.

[0027] Figure 7 The following is a graph showing the accuracy of the four models, YOLOv8n, YOLOv10n, YOLO11n, and Star-EJCD-YOLO, as the number of training rounds changes.

[0028] Figure 8The following is a graph showing the recall rate of the four models, YOLOv8n, YOLOv10n, YOLO11n and Star-EJCD-YOLO, as the number of training rounds changes.

[0029] Fig. 9 This is a graph showing the average precision mAP@0.5 of the four models YOLOv8n, YOLOv10n, YOLO11n and Star-EJCD-YOLO as the number of training rounds changes.

[0030] Fig.10 This is a curve chart showing the average precision mAP@0.5:0.95 of the four models YOLOv8n, YOLOv10n, YOLO11n and Star-EJCD-YOLO as the number of training rounds changes. DETAILED DESCRIPTION

[0031] The present invention converts the vibration signals of the motor startup phase and the switching switch action phase into high-resolution time-frequency images through continuous wavelet transform to fully reflect the time and frequency characteristics of the signal. Subsequently, the StarNet structure is used to replace the YOLO model backbone and an efficient joint convolution detection head EJCD is proposed. While lightweighting the model, the ability to extract key features of the model is enhanced to achieve efficient and accurate mechanical fault diagnosis. Compared with traditional methods, the present invention not only has significant advantages in processing non-stationary signals and instantaneous change characteristics, but also significantly improves the computational efficiency, and can meet the real-time requirements of online monitoring of equipment status. The present invention provides a new idea for the intelligent diagnosis of mechanical faults of on-load tap changers OLTC, and shows broad application prospects in the field of power equipment status monitoring. Figure 1 The figure is the structural diagram of the on-load tap changer. The following is an explanation of the time-frequency domain analysis of typical mechanical faults and vibration signals of the on-load tap changer.

[0032] 1. Typical mechanical failures of on-load tap-changers.

[0033] In the actual operation of the on-load tap changer, frequent gear adjustment operations under high voltage and high current environment cause frequent compression and release of springs and frequent collision of contacts. Long-term operation leads to gradual deterioration of the internal components of the on-load tap changer, and mechanical failures such as contact wear, spring fatigue, and loose fasteners are prone to occur. If not discovered and handled in time, serious consequences will occur. The main mechanical failures of the on-load tap changer are summarized in Table 1.

[0034] Table 1-Classification of main mechanical failures of OLTC

[0035]

[0036] 2. Time-frequency domain analysis of vibration signals.

[0037] Accelerometers are often used to collect vibration signals from equipment. The collected vibration signals are usually expressed as a function of acceleration changing with time. Although time domain analysis can show the change of signals over time, it cannot provide detailed information on frequency components, and it is difficult to effectively process non-stationary signals whose frequencies change with time. Frequency domain analysis reveals the frequency components of signals through Fourier transform, which can give the overall spectrum information of the signal but cannot reflect the changes of frequency components at different time points. Time-frequency domain analysis methods such as short-time Fourier transform and wavelet transform can provide both time and frequency information, capture mutations and fault characteristics in signals, and are suitable for analyzing non-stationary signals whose frequencies change with time. It makes up for the shortcomings of time domain and frequency domain analysis, and has important advantages in complex signal analysis and mechanical fault diagnosis.

[0038] The invention discloses a method for fault diagnosis of an on-load tap changer based on lightweight YOLO11, comprising the following steps: S1, converting vibration signals of a motor startup phase and a switch action phase into high-resolution time-frequency images through continuous wavelet transform (CWT) to fully reflect the time and frequency characteristics of the signals.

[0039] Continuous wavelet transform (CWT) extracts information about the signal at different time scales by convolving the signal with a set of wavelet functions, and is used to analyze the local characteristics of the signal in the time-frequency domain. Compared with short-time Fourier transform (SFT), continuous wavelet transform (CWT) is more suitable for non-stationary signals or signals with multi-scale characteristics to adapt to the non-stationarity of the signal. The basic idea of ​​continuous wavelet transform is to analyze the local characteristics of the signal by changing the scale and displacement of the wavelet function. Wavelet transform can be regarded as convolving the signal with mother wavelets of different scales to obtain the local frequency components of the signal at each scale.

[0040] Let the signal be , the mother wavelet is , then the continuous wavelet transform of the signal It is expressed as: (1).

[0041] Among them, a is the scale factor, which controls the expansion and contraction of the mother wavelet; b is the translation factor, which controls the translation of the mother wavelet; is the complex conjugate of the mother wavelet function.

[0042] The mother wavelet is the basic function used for wavelet transform, which determines the properties and application scope of wavelet transform. Different mother wavelets have different shapes, symmetries, translations and scale characteristics, and are suitable for different types of signal analysis. The present invention selects Cmor wavelet as the mother wavelet, which is the complex form of Morlet wavelet. The real part describes the frequency component of the signal, and the imaginary part provides phase information. Cmor wavelet combines Gaussian window and sine wave, has good time-frequency localization characteristics, good resolution in time and frequency, and is usually used in the fields of time-frequency analysis, feature extraction, and spectrum estimation of signals. Figure 2 This is a schematic diagram of the time-frequency domain conversion of the vibration signal of a certain operation of the on-load tap changer.

[0043] S2. StarNet is used to replace the YOLO model backbone network, and star operations are used to enhance feature extraction capabilities while maintaining network compactness.

[0044] YOLO11 is a neural network that integrates target detection, semantic segmentation, and image classification tasks. It has an improved backbone and neck architecture based on the previous generation, which enhances the feature extraction capability, improves the accuracy of object detection and the performance of complex tasks. YOLO11 introduces a refined architecture design and an optimized training process to achieve faster processing speed while maintaining the best balance between accuracy and performance. YOLO11 can be flexibly deployed in edge devices and cloud platform environments to provide users with maximum adaptability. In response to the need for online rapid diagnosis of mechanical faults of on-load tap changers, the present invention proposes a further lightweight model Star-EJCD-YOLO based on YOLO11n, such as Figure 3 By introducing StarNet to replace the original backbone network and combining it with the proposed efficient joint convolutional detection head EJCD, the model can significantly reduce the number of parameters and computational complexity while still maintaining high target detection accuracy and real-time performance.

[0045] In the field of image recognition, as the depth of neural networks and the number of feature map channels increase, network performance is significantly improved, but the computational cost increases. The accumulation of redundant information makes the consumption of computing resources even more intensive and is not suitable for online monitoring devices with high real-time requirements. The StarNet architecture uses star operations, that is, element-by-element multiplication to map low-dimensional inputs to high-dimensional nonlinear feature spaces. This operation can significantly improve the network's feature expression capabilities and thus improve performance without increasing the network width. StarNet has high computational efficiency and low latency while maintaining network compactness, making it suitable for application scenarios with limited resources.

[0046] For a single-layer neural network, the star operation is expressed as: (2).

[0047] in, , is the weight matrix and the bias term The combination matrix of is the input feature matrix The combined matrix after adding a feature dimension; Represents element-by-element multiplication. For a single output channel and a single input element, define ( is the weight vector, is the input vector), the star operation is expanded as: (3).

[0048] in, and Used to index the input feature channels, i.e. the elements in the weight vector. is a coefficient that represents the relationship or weighting method between each pair of input feature channels, and is specifically defined as: (4).

[0049] The star operation expansion will contain term, that is, the number of output feature dimensions, when d is large enough As the number of layers increases, the dimension of the implicit feature space increases exponentially: (5).

[0050] After the n-th layer of stacking is completed as shown in formula (5), the feature space dimension of the n-th star operation output will reach .

[0051] Based on this principle, Star Blocks, the core computing unit of StarNet, was constructed. StarNet uses efficient star operations to achieve efficient representation of information in high-dimensional feature space with minimal network structure and computation. Using StarNet to replace the backbone network of YOLO can not only effectively reduce the consumption of computing resources, but also improve the network's feature extraction capabilities and enhance the performance of the model. It is especially suitable for application scenarios with limited computing resources and high real-time requirements. Figure 4 This is the network structure diagram of StarNet.

[0052] S3. Design an efficient joint convolutional detection head EJCD to achieve model lightweight and multi-scale target detection through group normalization and convolution parameter sharing.

[0053] In the YOLO series of models, the detection head usually consists of three branches, each of which processes the feature information of the same image at different scales. This design enables the model to detect objects at multiple scales. However, these branches run independently and optimize parameters independently, resulting in an increase in the number of model parameters and low efficiency, which is prone to overfitting. In addition, the lack of effective information sharing and collaborative work also limits the overall performance and generalization ability of the model. To solve this problem, the present invention proposes an efficient joint convolution detection head EJCD.

[0054] First, the batch normalization (BN) in the convolutional layers of the detection head is replaced by group normalization (GN).

[0055] BN relies on the statistical information of batch data for normalization, and needs to additionally calculate the mean and variance of each batch and store these statistical information. In scenarios with small batches or real-time updates, the computational overhead will increase significantly. In contrast, GN reduces the computational burden by dividing the feature channels into multiple groups to independently calculate the mean and variance without relying on the batch size, making it suitable for environments with limited computing resources. In addition, GN can effectively improve the performance of the detection head in target positioning and classification tasks.

[0056] Secondly, through joint convolution, convolution modules are shared among multiple feature layers to further optimize the lightweight effect of the model.

[0057] Parameter sharing greatly reduces the number of model parameters and computational complexity, significantly improving operational efficiency, and is especially suitable for scenarios with limited resources. At the same time, multi-layer features can achieve consistent feature modeling through the unified processing of shared convolution parameters, and strengthen information fusion and collaboration between different feature layers. This design not only helps to find the global optimal solution, but also effectively reduces the risk of overfitting and improves the robustness of the model.

[0058] Finally, in order to deal with the problem of inconsistent scales of targets detected by each detection head, the Scale layer is used to scale the features.

[0059] The Scale layer helps the model better adapt to the scales of different objects by assigning independent scaling factors to feature maps of different scales. In object detection tasks, the scale of feature maps usually represents different levels of spatial information. Lower-level feature maps usually contain information about smaller objects, while higher-level feature maps contain information about larger objects. By introducing different scaling factors for each feature layer, the Scale layer enables the model to more flexibly locate and classify objects at different scales, thereby improving the detection accuracy of multi-scale objects. Figure 5It is an EJCD structure, in which two 3×3 Conv_GNs share parameters, and each regression convolution layer Conv_Reg and each classification convolution layer Conv_Cls share parameters.

[0060] The effectiveness of the on-load tap changer mechanical fault rapid diagnosis method of the present invention is verified in combination with experiments. The experimental object of the present invention is a typical combined KM type on-load tap changer, which is composed of a switching switch and a tap selector. Four working conditions are designed to simulate the common operating conditions and typical mechanical fault conditions of OLTC. The four working conditions are normal state, loose arc plate and contact, loose transmission shaft screws, and transmission gear jamming. The loose arc plate is to loosen a screw ring of the static contact on the A-phase arc plate of the OLTC, and different degrees of loose faults can be achieved by loosening the number of turns of the fixing screws of the corresponding components. The loose contact is achieved by loosening the neutral point contact and the related screws of the main on-off contact; the loose transmission shaft screw is achieved by loosening the screws at the connection of the transmission rod to achieve fault simulation. The transmission gear jam is achieved by plugging wood chips in the gear box of the switch transmission mechanism to increase the friction between the gears, thereby simulating the jam of the transmission gear. The present invention selects three voltage output type acceleration sensors with a sensitivity of 100mV / g, which are respectively installed on the top of the on-load tap changer and evenly spaced at 120° to collect vibration signals. Considering that the impact characteristics and frequency range of the corresponding vibration signal during the operation of the on-load tap changer are about 5-50kHz, the signal sampling rate is set to 25kHz.

[0061] Figure 6The original signal waveform of the vibration of the on-load tap changer during the whole process of switching from the 7th gear to the 8th gear under a normal state is shown. The process can be roughly divided into four stages: motor starting, transmission mechanism transmission, switching action, and motor stop. Different types of mechanical faults will affect the vibration characteristics of the on-load tap changer, resulting in changes in the amplitude and main frequency of the signal. During the motor starting stage, the interaction between the rotor and stator inside the motor produces a strong transient dynamic response, and the mechanical parts of the equipment such as gears and transmission shafts are subjected to a large starting torque. Analyzing the signals during the motor starting stage helps to detect potential faults before the on-load tap changer is actuated, thereby avoiding more serious damage. The switching action stage is the moment when multiple mechanical parts inside the OLTC begin to move and undergo drastic changes. The vibration signal amplitude at this stage is the largest, the amount of information is the richest, and it can fully reflect the dynamic behavior of the equipment. Motor starting and switching action are highly repetitive events during the operation of the OLTC. Different mechanical defects will lead to significant differences in the amplitude and frequency characteristics of the vibration signal. The present invention selects 0.18-0.23s after the sampling trigger as the motor start-up stage and 5.65-5.8s as the switching switch action stage, and takes the signal characteristics of these two stages as the key research objects of fault diagnosis. By extracting these high-information signal fragments in a targeted manner, it is helpful to capture fault characteristics in a complex dynamic environment and improve the accuracy and efficiency of diagnosis.

[0062] The vibration signals of odd-to-even gear switching under four working conditions were collected, and 150 sets of data were collected for each working condition. Subsequently, the vibration signals of the motor startup phase and the switching switch action phase were extracted, and a total of 1,200 corresponding time-frequency graphs were generated through continuous wavelet transform. Finally, the data was divided into training set, validation set, and test set in a ratio of 7:2:1.

[0063] This experiment was conducted in a Windows 11 system environment with a hardware configuration of 12th Intel Core i5-12400F CPU, 32GB RAM and NVIDIA GeForce RTX 4060. The deep learning framework used was PyTorch2.2.1, and was used in conjunction with CUDA 12.1. The training was set to 300 epochs, the batch size was 8, the optimization algorithm was stochastic gradient descent SGD, and the resolution of the input image was 640×640. The purpose of this study is to achieve lightweight models while maintaining high monitoring accuracy. To this end, precision Precision, recall rate Recall, mean average precision mAP@0.5 and mAP@0.5:0.95, model parameter quantity Parameters / M, GFLOPs and FPS were selected as evaluation indicators. The calculation formulas for precision, recall rate and mean average precision are:

[0064] (6).

[0065] (7).

[0066] (8).

[0067] (9).

[0068] Among them, TP is the number of samples correctly predicted as positive examples; FP is the number of samples that are actually negative examples but are incorrectly predicted as positive examples; FN is the number of samples that are actually positive examples but are incorrectly predicted as negative examples. N is the total number of categories. AP is based on the area of ​​the precision-recall curve; Indicates AP-like.

[0069] In order to verify the advantages of the model proposed in this invention, the test results of other models were compared using the same data set and parameter settings. The comparison results are shown in Table 2.

[0070] Table 2-Comparison of experimental results of Star-EJCD-YOLO and other YOLO series models

[0071]

[0072] From the results in the table, it can be seen that the model proposed in the present invention has improved both precision and recall. In terms of mAP index, although Star-EJCD-YOLO's mAP@0.5 is slightly lower than YOLO11n, it reaches 0.808 on the stricter mAP@0.5:0.95, which is better than other models and shows stronger generalization ability and robustness. In terms of model lightweighting, the number of parameters of Star-EJCD-YOLO is only 1.78M, which is 40.7%, 21.6% and 31.0% less than YOLOv8n, YOLOv10n and YOLO11n respectively; the model size is only 3.8MB, which is 38.7%, 30.9% and 30.9% less than YOLOv8n, YOLOv10n and YOLO11n respectively. In terms of computational complexity, Star-EJCD-YOLO has a GFLOPs of 4.2, which is 48.1% and 35.4% less than YOLOv8n and YOLOv10n, respectively. In terms of inference speed, Star-EJCD-YOLO reaches 499.5FPS, which is about 4.3% faster than YOLO11n and about 1.1% faster than YOLOv10n. Figure 7 The curves of the accuracy of the four models YOLOv8n, YOLOv10n, YOLO11n and Star-EJCD-YOLO as the training rounds change. Figure 8 The curves of the recall rate of the four models YOLOv8n, YOLOv10n, YOLO11n and Star-EJCD-YOLO as the training rounds change. Fig. 9 The curves of the mean average precision mAP@0.5 of the four models YOLOv8n, YOLOv10n, YOLO11n and Star-EJCD-YOLO changing with the training rounds are shown below. Fig.10 This is the curve showing the average precision mAP@0.5:0.95 of the four models YOLOv8n, YOLOv10n, YOLO11n and Star-EJCD-YOLO as the number of training rounds changes.

[0073] The ablation results with YOLO11n as the baseline are shown in Table 3. From the ablation results in the table, it can be seen that Star-EJCD-YOLO has significantly optimized the model complexity and computational cost while maintaining high detection accuracy by introducing StarNet and EJCD components. After the introduction of StarNet, the number of parameters is reduced from 2.58M to 1.94M, the amount of calculation is reduced from 6.3 to 5.0GFLOPs, and the model size is reduced from 5.5MB to 4.2MB, while mAP@0.5:0.95 remains at 0.806. This shows that StarNet can effectively reduce model complexity and maintain detection performance. After the introduction of EJCD, mAP@0.5:0.95 is increased from 0.806 to 0.809, and the number of parameters and the amount of calculation are reduced to 2.42M and 5.6GFLOPs respectively, proving the effectiveness of EJCD in improving generalization ability. Star-EJCD-YOLO combines the advantages of both, further reducing the number of parameters to 1.78M, the amount of calculation to 4.2GFLOPs, the model size to 3.8MB, and the mAP@0.5:0.95 to 0.808. Although mAP@0.5 has dropped slightly, this small loss has brought significant lightweight and efficiency improvements, fully verifying the model's ability to balance accuracy and efficiency, and is suitable for applications in resource-constrained scenarios such as real-time online monitoring of the status of on-load tap changer equipment.

[0074] Table 3 - Ablation experiment results

[0075]

[0076] The present invention proposes a fast diagnosis method for mechanical faults of on-load tap changers (OLTCs) based on the lightweight YOLO11 of Star-EJCD to address the deficiencies of existing diagnostic methods in processing non-stationary signals, computational efficiency, and real-time monitoring capabilities of equipment. By integrating the time-frequency domain characteristics of vibration signals and deep learning technology, the present invention demonstrates significant advantages and provides an efficient and intelligent solution for the field of fault diagnosis of power equipment. First, for typical mechanical fault types of OLTC, including loose arc plates and contacts, loose transmission shaft screws, and stuck transmission gears, rich vibration signal data were collected through simulation experiments. The vibration signal has significant dynamic characteristics and high-frequency impact characteristics in different action stages of OLTC, such as motor start-up and switching action, which are difficult to fully characterize through traditional time domain and frequency domain analysis methods. The present invention uses continuous wavelet transform (CWT) to perform time-frequency domain analysis on the signal, converting the vibration signal into a high-resolution time-frequency image, so that the local characteristics of the non-stationary signal can be fully displayed. Secondly, in order to improve the feature extraction capability and lightweight level of the model, it is proposed to replace the backbone network structure of YOLO11 with StarNet. StarNet significantly enhances the model's feature expression capability while maintaining the compactness of the network through star operations. This architecture can achieve efficient feature extraction and modeling for the complex dynamic characteristics of OLTC vibration signals under limited computing resources. In addition, an efficient joint convolution detection head EJCD is designed, and the parameter sharing and feature fusion mechanism of the model is further optimized by replacing batch normalization BN with group normalization GN. Experimental results show that EJCD not only effectively reduces the number of parameters and computational complexity of the model, but also significantly improves the model's multi-scale target detection capability. Finally, by comparing the performance of different YOLO series models in the experiment, the superiority of the proposed method is verified. In terms of precision Precision, recall rate Recall, average precision mean mAP@0.5 and mAP@0.5:0.95 indicators, this model performs well, especially in mAP@0.5:0.95, reaching 0.808, which is better than other lightweight YOLO models. At the same time, the number of parameters of Star-EJCD-YOLO is only 1.78M, which is 31.0% less than YOLO11n, and the inference speed is increased to 499.5FPS, which fully demonstrates the effectiveness of its lightweight design.

[0077] In summary, the on-load tap changer fault diagnosis method based on lightweight YOLO11 proposed in this invention not only overcomes the limitations of traditional methods in non-stationary signal processing, but also significantly improves the diagnostic efficiency and model performance by combining the time-frequency image representation capability of continuous wavelet transform with the efficient target detection framework of deep learning. This method is suitable for real-time online monitoring in resource-constrained scenarios, provides a new research direction for intelligent diagnosis of power equipment such as OLTC, and has broad application prospects.

Claims

1. A fault diagnosis method for on-load tap changer based on lightweight YOLO11, characterized in that: The following steps are involved: S1. The vibration signals of the motor startup stage and the switching action stage are converted into high-resolution time-frequency images through continuous wavelet transform (CWT) to fully reflect the time and frequency characteristics of the signal; Assume the signal is x(t) and the mother wavelet is ψ(t), then the continuous wavelet transform of the signal W is x (a,b) is expressed as: Among them, a is the scale factor, which controls the expansion and contraction of the mother wavelet; b is the translation factor, which controls the translation of the mother wavelet; is the complex conjugate of the mother wavelet function; S2, using StarNet to replace the YOLO model backbone network, and enhancing the feature extraction capability while maintaining the network compactness through star operations; The StarNet architecture uses star operations, i.e., element-by-element multiplication, to map low-dimensional inputs to high-dimensional nonlinear feature spaces. Based on this, StarNet's core computing unit, Star Blocks, is built. StarNet uses efficient star operations to achieve efficient representation of information in high-dimensional feature spaces with minimal network structure and computational effort. S3. Design an efficient joint convolution detection head EJCD to achieve model lightweight and multi-scale target detection through group normalization and convolution parameter sharing; First, the batch normalization BN in the convolutional layer of the detection head is replaced by group normalization GN; Secondly, through joint convolution, convolution modules are shared among multiple feature layers to further optimize the lightweight effect of the model; Finally, to address the problem of inconsistent scales of the targets detected by each detection head, the Scale layer is used to scale the features; For a single-layer neural network, the star operation is expressed as: in, It is the combination matrix of weight matrix W1 and bias term B; It is the combination matrix of weight matrix W2 and bias term B; for The transpose of for The transpose of The combined matrix after adding a feature dimension to the input feature matrix X; * represents element-by-element multiplication; For the scenario of a single output channel and a single input element, define w1, w2, x∈R (d+1)×1 , w1, w2 are weight vectors, x is the input vector, and the star operation is expanded as follows: Among them, i and j are used to index the input feature channels, that is, the elements in the weight vector; α(i,j) is a coefficient that represents the relationship or weighting method between each pair of input feature channels, which is specifically defined as: Star operation expansion includes term, that is, the number of output feature dimensions, when d is large enough As the number of layers increases, the dimension of the implicit feature space will increase exponentially: After the n-th stacking layer shown in the above formula is completed, the feature space dimension of the output of the n-th star operation will reach 2. The on-load tap changer fault diagnosis method based on lightweight YOLO11 according to claim 1 is characterized in that: Cmor wavelet is selected as the mother wavelet, the real part describes the frequency component of the signal, and the imaginary part provides phase information.

Images