Iron core reactor epoxy material aging detection method and system based on noise characteristics

Abstract

The invention discloses an iron core reactor epoxy material aging detection method and system based on noise characteristics. The method comprises the steps that noise signals of a reactor in a normal state and a reactor in an epoxy material aging state are collected through a microphone noise array; preprocessing the collected signal to remove noise, and converting the signal into an image as a sample set; constructing an image feature extraction model based on a residual neural network, namely Resnet, and performing training; hOG feature extraction is carried out on the image; and combining the extracted features of each image with the HOG features as a basic sample, converting the basic sample into one-dimensional sequences, carrying out Euclidean distance comparison on different one-dimensional sequences, setting a threshold value, and judging the aging condition of the reactor epoxy material according to the Euclidean distance comparison result. According to the method, the aging condition of the iron core reactance epoxy material can be detected from the two aspects of features extracted by Resnet and features extracted by HOG, and the detection accuracy and practicability are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of power equipment detection, relates to insulation detection of a dry-type iron core reactor, and in particular to epoxy material degradation detection of a dry-type iron core reactor based on noise feature merging. Background Art

[0002] Dry-type iron core reactors are important power equipment for long-distance transmission and distribution of electric energy in power grids, and their stability is crucial to the safety of power grids. As an important mechanical component of the reactor, the winding has the highest proportion of mechanical failures. The winding vibrates in the leakage magnetic field due to the alternating electromagnetic force. Combined with the long-term vibration effect of harmonic superposition, the epoxy material of the fixed winding is irreversibly degraded, which eventually leads to mechanical degradation and insulation degradation of the reactor winding. In severe cases, production and living electricity consumption may be affected, causing significant economic losses. During the operation of the reactor, the soundprint morphology transmitted from the inside to the outside changes with the mechanical stability of the body. Therefore, the noise signal received from the outside of the reactor contains rich mechanical state information.

[0003] The noise array is used to obtain the external noise distribution and dynamic response law of the reactor, which can not only capture the mechanical state of the reactor winding in time, but also meet the needs of live detection, avoiding the inconvenience caused by the shutdown of the reactor. The noise array receives noise signals of multiple channels, and then uses the beamforming processing algorithm to calculate the location information of the sound source. It is composed of multiple microphones arranged according to a certain spatial geometric law. Compared with the test method using only a single microphone, the noise array has a series of obvious advantages in the signal reception and processing process, mainly in the anti-reverberation, noise removal, and speech separation capabilities. The higher spatial resolution, the multi-channel signal processing method of the microphone array enriches the processing domain of the sound signal, and adds the spatial domain in addition to the time domain and frequency domain, making the sound source positioning more accurate and more adaptable. Multiple noise sensors are used to collect sound synchronously, and the noise signals obtained by different sensors are reproduced in the sound pressure level distribution on the corresponding plane, and represented by a pseudo-color map, where dark colors represent positions with relatively large sound pressure levels, and light colors represent positions with relatively small sound pressure levels. See papers [1]-

[19] . In traditional methods, hand-designed features are often used to carry out defect detection, including color features, texture features, shape features, etc. However, these features cannot fully reflect the essential characteristics of the target in the image and cannot meet the requirements of image classification. Although image features extracted by deep learning can better meet the requirements of image classification, convolutional neural networks are prone to filter out the associated details of the entire image and its parts after multiple pooling, resulting in low classification accuracy of convolutional neural networks when distinguishing objects of similar categories.

[0004] [1] Wang Hongfang, Wang Naiqing, Li Tongsheng Effect of transformer winding axial preload on winding axial vibration characteristics

[0005] [J]. Power System Technology, 1999, 23(9): 8-11.

[0006] [2] Ji Shengchang, Wang Junde, Li Yanming. Research on axial vibration characteristics of transformer windings under steady-state conditions[J].

[0007] New Technology of Electrical and Electrical Engineering, 2006(01):35-38+48.

[0008] [3] Ji Shengchang, Zhang Fan, Qian Guochao, Zhu Yeye, Dong Hongkui, Zou Dexu. Axial direction of transformer winding under steady-state conditions.

[0009] Vibration characteristics and its influencing factors[J]. High Voltage Engineering, 2016, 42(10): 3178-3187.

[0010] [4] Ma Hongzhong, Geng Zhihui, Chen Kai, Wang Chunning, Li Kai, Li Yong. Winding deformation of power transformer based on vibration.

[0011] New method of fault diagnosis[J]. Automation of Electric Power Systems, 2013, 37(08):89-95.

[0012] [5] Zhao Hongfei, Ma Hongzhong, Chen Kai, Wang Chunning, Yu Ziyu. Transformer winding loosening simulation based on vibration signal.

[0013] Experimental research[J].China Electric Power,2014,47(01):13-16+22.

[0014] [6] Ma Hongzhong, Zhou Yu, Li Kai, Xu Honghua, Wang Taoyun. Vibration-based evaluation method for transformer winding compression state.

[0015] [J]. Automation of Electric Power Systems, 2015, 39(18): 127-132.

[0016] [7]BERKER Z, GOLUBEV A, RUSOV V, et al.Vibro-acoustic method of transformer clamping pressure monitoring[C] / / Proceedings ofIEEEInternational Symposium on Electrical Insulation.Piscataway:

[0017] IEEE, 2000:63-266.

[0018] [8] Guo Jie. Research on vibration and fault feature extraction of power transformers[D] Zhejiang University, 2011

[0019] [9] Zhang Bin, Xu Jianyuan, Chen Jiangbo, Li Hui, Lin Xin, Zang Zhuang. Winding shape analysis based on vibration information of power transformer.

[0020] Transformation diagnosis method[J]. High Voltage Engineering, 2015, 41(07): 2341-2349.

[0021]

[10] Wang Fenghua, Hu Xuming, Qian Yong, Chen Peilong, Ma Xiaohong. A review of research on transformer winding vibration monitoring technology.

[0022] [J]. Guangdong Electric Power, 2018, 31(08): 52-61.

[0023]

[11] Luo Yuan, Zhuang Hua, Qin Langkuan, Zhao Jieying, Yin Hao, Liu Dongquan, Wu Yuting, Liu Ke, Hu Hanchuan.

[0024] Method for grading the heterogeneity of rectal tumor ultrasound contrast imaging based on biomatrix[J].

[0025] Zhi, 2019, 36(06): 964-968. References

[0026]

[12] Lu Wenbo. Research on mechanical fault diagnosis method and its application based on the spatial distribution characteristics of the sound field[D].

[0027] Shanghai Jiaotong University, 2012.

[0028]

[13] Cao Xia. Research on content-based lace retrieval method and system implementation [D]. Jiangnan University, 2017.

[14] Yue Yingjuan, Wang Xu, Cai Yanping. Recognition and diagnosis of texture features of internal combustion engine variational modal Rihaczek spectrum [J].

[0029] Chinese Journal of Scientific Instrumentation, 2017, 38(10): 2437-2445.

[0030]

[15] Song Benqin, Li Peijun. Classification of high-resolution remote sensing images with improved LBP texture[J].

[0031] Remote Sensing, 2010(04):40-45.

[0032]

[16] Li Huanli, Guo Lihong, Li Xiaoming, Wang Xinzui, Dong Yuefang. Central symmetric local binary based on statistical features.

[0033] Iris recognition based on patterns[J]. Optics and Precision Engineering, 2013, 21(08): 2129-2136.

[0034]

[17] Gong Luming, Xu Meihua, Liu Dongjun, Zhang Fayu. Pedestrian detection model based on mixed Gaussian and HOG+SVM.

[0035] Model[J]. Journal of Shanghai University (Natural Science Edition), 2018, 24(03): 341-351.

[0036]

[18] Xiang Lei, Xu Jun. Cell detection in breast pathology images based on HOG features and sliding window[J]. Shandong

[0037] Journal of University(Engineering Edition),2015,45(01):37-44+63.

[0038]

[19] Bu Qingzhi, Qiu Jun, Hu Chao. Detection of driver distraction behavior based on HOG feature extraction and SVM.

[0039] Research on measurement methods[J].Integration Technology,2019,8(04):69-75.

[0040] Albiol A,Monzo D,Martin A,etal.Face Recogition Using HOG-EBGM[J].PatternRecognition Letters,2008,29(10);1537-1543. Summary of the invention

[0041] In order to solve the deficiencies in the prior art, the present invention provides a method and system for detecting the aging of epoxy materials of iron core reactors based on noise features. The method can detect the aging of epoxy materials of iron core reactors from two aspects: image features extracted by deep learning and HOG extraction features. It can distinguish the aging conditions from the depth features, and can also enhance the distinction of aging conditions from the detail features, thereby improving the accuracy and practicability of detection.

[0042] To achieve the above-mentioned object of the invention, the present invention adopts the following technical solution.

[0043] A first aspect of the present invention provides a method for detecting aging of epoxy materials of an iron core reactor based on noise characteristics, comprising the following steps:

[0044] Step 1: Use a microphone noise array to collect noise signals from the reactor in a normal state and in an epoxy material aging state;

[0045] Step 2: Preprocess the noise signal collected in step 1 to remove the noise interference signal, use the beamforming algorithm to extract the vibration signal of the epoxy material of the iron core reactor from the preprocessed signal, and convert it into an image as a sample set, and divide it into a normal set, a slight set, a moderate set, and a severe set according to the aging condition of the epoxy material of the iron core reactor;

[0046] Step 3: Build an image feature extraction model based on the residual neural network, i.e., Resnet, and train the normal set, mild set, moderate set, and severe set obtained in step 2 to extract N-dimensional deep features respectively;

[0047] Step 4: Extract HOG features from the images of the normal set, mild set, moderate set, and severe set;

[0048] Step 5: Merge the N-dimensional deep features of each image with the HOG features and use them as the basic samples;

[0049] Step 6: Convert the basic samples into one-dimensional sequences, and perform distance comparison calculations on the one-dimensional sequences of the mild set, moderate set, and severe set respectively with the one-dimensional sequence of the normal set, and set the threshold according to the distance comparison results;

[0050] Step 7: Collect the noise signal of the reactor in the period to be detected, pre-process it to remove the noise interference signal, use the beamforming algorithm to extract the vibration signal of the epoxy material of the iron core reactor in the period to be detected, and convert it into an image, perform feature extraction and HOG feature extraction based on Resnet, fuse the N-dimensional deep features and HOG features and convert them into a one-dimensional sequence, perform distance comparison calculation with the one-dimensional sequence of the normal set, and judge the aging condition of the epoxy material of the reactor according to the distance comparison result.

[0051] Preferably, in step 1, the sensor distribution form of the microphone noise array may be S-shaped or linear;

[0052] During the acquisition process, the excitation voltage was adjusted under the normal state of the reactor and the state of aging of the epoxy material, starting from 90% of the rated voltage and with a step size of 5%, and then adjusted to 110% of the rated voltage to acquire noise array signals outside the reactor.

[0053] Preferably, when setting the aging state of the epoxy material, the epoxy pressure plate is loosened using a torque wrench, and the degrees of loosening are 4 N·m, 8 N·m, and 12 N·m, respectively, thereby forming three corresponding aging fault states, namely, slight aging fault, moderate aging fault, and severe aging fault. In each fault state, the excitation voltage is adjusted in sequence according to the above-mentioned voltage step, and the noise signal is collected respectively.

[0054] Preferably, in step 2, the noise signal is subjected to double-tree complex wavelet decomposition, and the noise variance of each layer of wavelet coefficients after double-tree complex wavelet decomposition is obtained by using double-tree complex wavelet threshold denoising method based on Bayesian maximum a posteriori estimation, and the noise variance is used as the denoising threshold. After thresholding the wavelet coefficients of each layer after double-tree complex wavelet decomposition of the noise signal, the sound signal of the epoxy material of the core reactor after filtering out the high-frequency noise is obtained by inverse double-tree complex wavelet transform.

[0055] Preferably, after the signal distribution processed by the beamforming algorithm is interpolated by using the cosine interpolation method, the Gram angular field method (GAF) is used to generate array signal images of the same size for the signal.

[0056] Preferably, in step 3, the input of the image feature extraction model is the i-th array signal image A in the normal set i , the learned features are output as H(x), and the residual unit of the model is shown in the following formula:

[0057]

[0058] Among them, y n represents the nth residual unit, and They represent the input and output of the nth residual unit respectively; F() is the residual function, which represents the learned residual result, and W n is the convolution kernel of the nth layer of the model, and f is the ReLU activation function;

[0059] The feature extraction results from the 1st layer to the Lth layer of the model are shown in the following formula:

[0060]

[0061] Among them, A L n It is the feature result extracted from the 1st layer to the Lth layer.

[0062] Preferably, the training content includes: using the initial convolution layer of the model to extract edge information of the image through convolution calculation; based on the aforementioned edge information, using the intermediate convolution layer to continue to use convolution operation to extract local information of the image; using the convolution layer to extract high-level semantic features of the image, based on the aforementioned high-level semantic features, saving the model network weight parameters with the above features, and using the model with the weight parameters as a model that can extract N-dimensional deep features of the normal condition of the epoxy material of the iron core reactor;

[0063] The model with weight parameters is used to input a slight set, a moderate set, and a severe set respectively, and the N-dimensional deep features of slight aging faults, moderate aging faults, and severe aging faults of epoxy materials of iron core reactors are obtained.

[0064] Preferably, in step 4, the images of the normal set, the mild set, the moderate set, and the severe set are grayscaled, the images are color and space normalized using a gamma correction method, the gradient size and direction of each pixel in the image are calculated, the image is divided into cells according to a set size, the gradient histogram in each cell is calculated, a block is formed according to a set number of cells, the gradient features in each block are calculated, and the gradient features of all blocks in the image are combined to obtain the HOG features of the image.

[0065] Preferably, in step 6, the distance comparison calculation is shown in the following formula:

[0066]

[0067] Among them, when X=1, that is, D1 qs H represents the Euclidean distance between the qth one-dimensional sequence of the mild set and the sth one-dimensional sequence of the normal set. s1 Represents the first value in the sth one-dimensional sequence of the normal set, H sB Represents the last value in the s-th one-dimensional sequence of the normal set, L1 q1 Represents the first value in the qth one-dimensional sequence of the minor set, L1 qB The last value in the qth one-dimensional sequence represents a mild set, that is, each sequence has B values, X=2 represents a moderate set, and X=3 represents a severe set.

[0068] Preferably, the images in the normal set are compared and calculated using Euclidean distances to obtain n! Euclidean distance comparison results, and the distance result with the largest value is selected as the threshold 1, i.e., α1;

[0069] Secondly, the C distance comparison results between the mild set and the normal set are normalized, and the mapping range is [0, 1]. The normalized comparison results are classified according to [0, 0.2], (0.2, 4], (0.4, 0.6], (0.6, 0.8], (0.8, 1], and then the probability of each interval is obtained. The interval with the largest probability is selected and classified.

[0070] The distance comparison result corresponding to the maximum value of the normalized comparison result is taken as α2;

[0071] The C distance comparison results between the moderate set and the normal set are processed in the same way as α3;

[0072] Finally, the C distance comparison results between the severe set and the normal set are processed in the same way as α4;

[0073] Threshold 2, i.e., α2, and threshold 3, i.e., α3, are expressed as follows:

[0074]

[0075]

[0076] Preferably, in step 7, the aging condition of the epoxy material of the reactor is determined according to the distance comparison result between the one-dimensional sequence of the epoxy material characteristics of the core reactor in the detection period and the one-dimensional sequence of the normal set characteristics and the set threshold, as shown in the following formula:

[0077]

[0078]

[0079] Where: result represents the aging of the epoxy material of the reactor, c represents the average value of the Euclidean distance comparison between the one-dimensional sequence converted by merging the N-dimensional deep features of the image with the HOG features and the normal set samples, J i The distance comparison result between the one-dimensional sequence converted by merging the N-dimensional deep features and HOG features of the image converted from the vibration signal of the epoxy material of the core reactor during the detection period and the one-dimensional sequence converted by merging the N-dimensional deep features and HOG features of the i-th image in the normal set.

[0080] The present invention provides an iron core reactor epoxy material aging detection system based on noise characteristics, and the iron core reactor epoxy material aging detection method based on noise characteristics is implemented, comprising:

[0081] Reactor epoxy material signal acquisition module, signal processing module, signal to image module, image feature extraction model building module, HOG feature extraction module, reactor epoxy material state judgment module;

[0082] The reactor epoxy material signal acquisition module collects noise signals of the reactor in normal state and epoxy material aging state respectively through the microphone noise array;

[0083] The signal processing module is used to pre-process the collected noise signal to remove the noise, and use the beam forming algorithm to extract the relevant sound signal of the epoxy material of the iron core reactor from the denoised signal;

[0084] The signal-to-image module is used to convert the processed sound signal into an image as a sample set;

[0085] The image feature extraction model construction module is used to construct an image feature extraction model using Resnet and train the normal set, mild set, moderate set, and severe set;

[0086] The HOG feature extraction module is used to extract HOG features from the images of the normal set, mild set, moderate set, and severe set;

[0087] The reactor epoxy material status judgment module is used to merge the N-dimensional deep features of each image with the HOG features, and convert them into a one-dimensional sequence as a basic sample. The one-dimensional sequences of the mild set, moderate set, and severe set are respectively compared with the one-dimensional sequence of the normal set for distance calculation, and the threshold is set. The aging condition of the reactor epoxy material is judged according to the distance comparison results between the one-dimensional sequence of the iron core reactor epoxy material characteristics of the time period to be detected and the one-dimensional sequence of the normal set characteristics.

[0088] The beneficial effect of the present invention lies in that, compared with the prior art, the present invention generates an acoustic imaging image based on a delayed-sum beamforming algorithm, and then extracts eigenvalues ​​according to a noise-sound pressure co-occurrence matrix, thereby realizing the detection of a loose core fault of a reactor; compared with the existing method for detecting a loose core fault of a reactor, the present invention has the following advantages: the noise array simplifies the on-site operation process of the operator and greatly reduces the risk of being interfered with; the deep features and HOG features are comprehensively considered, and the merged features affect the judgment accuracy, and can accurately identify the degradation of the epoxy material; there is no physical contact with the reactor during the measurement process, and the operation of the reactor is not affected. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a schematic flow chart of the method for detecting aging of epoxy materials of iron core reactor based on noise characteristics of the present invention;

[0090] Figure 2 A schematic diagram of an acoustic imaging device used in the present invention;

[0091] Figure 3 forming a sample graph for the collected noise signal;

[0092] Figure 4 It is a schematic diagram of the framework of the iron core reactor epoxy material aging detection system based on noise characteristics of the present invention; DETAILED DESCRIPTION

[0093] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0094] like Figure 1 As shown, the present invention provides a method for detecting aging of epoxy materials of iron core reactors based on noise characteristics, comprising the following steps:

[0095] Step 1: Use a microphone noise array to collect noise signals from the reactor in a normal state and in an epoxy material aging state.

[0096] In a preferred but non-limiting embodiment, Figure 2 As shown, in step 1, the sensor distribution form of the microphone noise array can be S-shaped or linear;

[0097] During the acquisition process, the excitation voltage was adjusted under the normal state of the reactor and the state of aging of the epoxy material, starting from 90% of the rated voltage and with a step size of 5%, and then adjusted to 110% of the rated voltage to acquire noise array signals outside the reactor.

[0098] In a preferred but non-restrictive implementation, when setting the aging state of the epoxy material, it is necessary to use a torque wrench to loosen the epoxy pressure plate, and the degrees of loosening are 4 N·m, 8 N·m, and 12 N·m, respectively, thereby forming three corresponding aging fault states, namely, slight aging fault, moderate aging fault, and severe aging fault. In each fault state, the excitation voltage is adjusted in sequence according to the above-mentioned voltage step, and the noise signal is collected respectively.

[0099] Step 2: Preprocess the noise signal collected in step 1 to remove the noise interference signal, and use the beamforming algorithm to extract the vibration signal of the epoxy material of the iron core reactor from the preprocessed signal, such as Figure 3 As shown, it is converted into images as a sample set, and is divided into a normal set, a slight set, a moderate set, and a severe set according to the aging conditions of the epoxy material of the iron core reactor. The normal set is set to contain S images, and the slight set, the moderate set, and the severe set are set to contain Q images.

[0100] In a preferred but non-restrictive implementation, the noise signal is subjected to dual-tree complex wavelet decomposition, and a dual-tree complex wavelet threshold denoising method based on Bayesian maximum a posteriori estimation is used to obtain the noise variance of each layer of wavelet coefficients after the dual-tree complex wavelet decomposition, and the noise variance is used as a denoising threshold. After thresholding the wavelet coefficients of each layer of the noise signal after the dual-tree complex wavelet decomposition, the sound signal of the epoxy material of the core reactor is obtained after filtering out the high-frequency noise through an inverse dual-tree complex wavelet transform.

[0101] In a preferred but non-limiting embodiment, after the signal distribution processed by the beamforming algorithm is interpolated by using the cosine interpolation method, the Gram angular field method (GAF) is used to generate array signal images of the same size for the signal.

[0102] Step 3: Build an image feature extraction model based on the residual neural network, namely Resnet, and train the normal set, mild set, moderate set, and severe set obtained in step 2 to extract N-dimensional deep features respectively.

[0103] In a preferred but non-limiting embodiment, the input of the image feature extraction model is the i-th array signal image A in the normal set. i , the learned features are output as H(x), and the residual unit of the model is shown in the following formula:

[0104]

[0105] Among them, y n represents the nth residual unit, and They represent the input and output of the nth residual unit respectively; F() is the residual function, which represents the learned residual result, and W n is the convolution kernel of the nth layer of the model, and f is the ReLU activation function;

[0106] The feature extraction results from the 1st layer to the Lth layer of the model are shown in the following formula:

[0107]

[0108] Among them, A L n It is the feature result extracted from the 1st layer to the Lth layer.

[0109] In a preferred but non-limiting embodiment, the training content includes: using the initial convolution layer of the model to extract edge information of the image through convolution calculation; based on the aforementioned edge information, using the intermediate convolution layer to continue to use convolution operation to extract local information of the image; using the convolution layer to extract high-level semantic features of the image, based on the aforementioned high-level semantic features, saving the model network weight parameters with the above features, and using the model with the weight parameters as a model that can extract N-dimensional deep features of the normal condition of the epoxy material of the core reactor;

[0110] The model with weight parameters is used to input a slight set, a moderate set, and a severe set respectively, and the N-dimensional deep features of slight aging faults, moderate aging faults, and severe aging faults of epoxy materials of iron core reactors are obtained.

[0111] Step 4: Perform HOG feature extraction on the images of the normal set, mild set, moderate set, and severe set.

[0112] In a preferred but non-limiting embodiment, the images of the normal set, the mild set, the moderate set, and the severe set are grayed, the images are color- and spatially normalized using a gamma correction method, and then the gradient size and direction of each pixel in the image are calculated, the image is divided into cells according to a set size, the gradient histogram in each cell is calculated, a block is formed according to a set number of cells, the gradient features in each block are calculated, and the gradient features of all blocks in the image are combined to obtain the HOG features of the image;

[0113] Since the size of each image is the same, the number of HOG feature dimensions extracted from each image is also the same.

[0114] Step 5: Merge the N-dimensional deep features of each image with the HOG features and use them as the base sample.

[0115] Step 6: Convert the basic samples into one-dimensional sequences, and perform distance comparison calculations on the one-dimensional sequences of the mild set, moderate set, and severe set respectively with the one-dimensional sequence of the normal set, and set the threshold according to the distance comparison results.

[0116] In a preferred but non-limiting embodiment, the distance comparison is calculated as shown in the following formula:

[0117]

[0118] Among them, when X=1, that is, D1 qs H represents the Euclidean distance between the qth one-dimensional sequence of the mild set and the sth one-dimensional sequence of the normal set. s1 Represents the first value in the sth one-dimensional sequence of the normal set, H sB Represents the last value in the s-th one-dimensional sequence of the normal set, L1 q1 Represents the first value in the qth one-dimensional sequence of the minor set, L1 qB The last numerical value in the qth one-dimensional sequence represents the mild set, that is, each sequence has B numerical values, X=2 represents the moderate set, X=3 represents the severe set, and finally the Euclidean distance comparison results of the mild set, moderate set, severe set and normal set are Q*S=C results respectively.

[0119] In a preferred but non-limiting embodiment, distance calculation is performed on the images in the normal set to obtain n! distance comparison results, and the distance result with the largest value is selected as the threshold 1, i.e., α1;

[0120] Secondly, the C distance comparison results between the mild set and the normal set are normalized, and the mapping range is [0, 1]. The normalized comparison results are classified according to [0, 0.2], (0.2, 4], (0.4, 0.6], (0.6, 0.8], (0.8, 1], and then the probability of each interval is obtained. The interval with the largest probability is selected and classified.

[0121] The distance comparison result corresponding to the maximum value of the normalized comparison result is taken as α2;

[0122] The C distance comparison results between the moderate set and the normal set are processed in the same way as α3;

[0123] Finally, the C distance comparison results between the severe set and the normal set are processed in the same way as α4;

[0124] Threshold 2, i.e., α2, and threshold 3, i.e., α3, are expressed as follows:

[0125]

[0126]

[0127] Step 7: Collect the noise signal of the reactor during the period to be detected, pre-process it to remove the noise interference signal, use the beamforming algorithm to extract the vibration signal of the epoxy material of the iron core reactor during the period to be detected, and convert it into an image. Perform feature extraction and HOG feature extraction based on Resnet, fuse the N-dimensional deep features and HOG features and convert them into a one-dimensional sequence, perform distance comparison calculation with the one-dimensional sequence of the normal set, and judge the aging condition of the epoxy material of the reactor based on the distance comparison result.

[0128] In a preferred but non-limiting embodiment, in step 7, the aging of the epoxy material of the reactor is determined according to the distance comparison result between the one-dimensional sequence of the epoxy material characteristics of the core reactor in the detection period and the one-dimensional sequence of the normal set characteristics and the set threshold, as shown in the following formula:

[0129]

[0130]

[0131] Where: result represents the aging of the epoxy material of the reactor, c represents the average value of the Euclidean distance comparison between the one-dimensional sequence converted by merging the N-dimensional deep features of the image with the HOG features and the normal set samples, J iThe distance comparison result between the one-dimensional sequence converted by merging the N-dimensional deep features and HOG features of the image converted from the vibration signal of the epoxy material of the core reactor during the detection period and the one-dimensional sequence converted by merging the N-dimensional deep features and HOG features of the i-th image in the normal set.

[0132] like Figure 4 As shown, the present invention provides an iron core reactor epoxy material aging detection system based on noise characteristics, and the iron core reactor epoxy material aging detection method based on noise characteristics is executed, comprising:

[0133] Reactor epoxy material signal acquisition module, signal processing module, signal to image module, image feature extraction model building module, HOG feature extraction module, reactor epoxy material state judgment module;

[0134] The reactor epoxy material signal acquisition module collects noise signals of the reactor in normal state and epoxy material aging state respectively through the microphone noise array;

[0135] The signal processing module is used to pre-process the collected noise signal to remove the noise, and use the beam forming algorithm to extract the relevant sound signal of the epoxy material of the iron core reactor from the denoised signal;

[0136] The signal-to-image module is used to convert the processed sound signal into an image as a sample set;

[0137] The image feature extraction model construction module is used to construct an image feature extraction model using Resnet and train the normal set, mild set, moderate set, and severe set;

[0138] The HOG feature extraction module is used to extract HOG features from the images of the normal set, mild set, moderate set, and severe set;

[0139] The reactor epoxy material status judgment module is used to merge the N-dimensional deep features of each image with the HOG features, and convert them into a one-dimensional sequence as a basic sample. The one-dimensional sequences of the mild set, moderate set, and severe set are respectively compared with the one-dimensional sequence of the normal set for distance calculation, and the threshold is set. The aging condition of the reactor epoxy material is judged according to the distance comparison results between the one-dimensional sequence of the iron core reactor epoxy material characteristics of the time period to be detected and the one-dimensional sequence of the normal set characteristics.

[0140] The present disclosure may be a system and / or a method.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting aging of epoxy materials of iron core reactor based on noise characteristics, characterized in that: The method comprises the following steps: Step 1: Use a microphone noise array to collect noise signals from the reactor in a normal state and in an epoxy material aging state; Step 2: Preprocess the noise signal collected in step 1 to remove the noise interference signal, use the beamforming algorithm to extract the vibration signal of the epoxy material of the iron core reactor from the preprocessed signal, and convert it into an image as a sample set, and divide it into a normal set, a slight set, a moderate set, and a severe set according to the aging condition of the epoxy material of the iron core reactor; Step 3: Build an image feature extraction model based on the residual neural network, i.e., Resnet, and train the normal set, mild set, moderate set, and severe set obtained in step 2 to extract N-dimensional deep features respectively; Step 4: Extract HOG features from the images of the normal set, mild set, moderate set, and severe set; Step 5: Merge the N-dimensional deep features of each image with the HOG features and use them as the basic samples; Step 6: Convert the basic samples into one-dimensional sequences, and perform distance comparison calculations on the one-dimensional sequences of the mild set, moderate set, and severe set respectively with the one-dimensional sequence of the normal set, and set the threshold according to the distance comparison results; Step 7: Collect the noise signal of the reactor during the period to be detected, pre-process it to remove the noise interference signal, use the beamforming algorithm to extract the vibration signal of the epoxy material of the iron core reactor during the period to be detected, and convert it into an image. Perform feature extraction and HOG feature extraction based on Resnet, fuse the N-dimensional deep features and HOG features and convert them into a one-dimensional sequence, perform distance comparison calculation with the one-dimensional sequence of the normal set, and judge the aging condition of the epoxy material of the reactor based on the distance comparison result.

2. The method for detecting aging of epoxy materials of iron core reactor based on noise characteristics according to claim 1, characterized in that: In step 1, the sensor distribution form of the microphone noise array can be S-shaped or linear; During the acquisition process, the excitation voltage was adjusted under the normal state of the reactor and the state of aging of the epoxy material, starting from 90% of the rated voltage and with a step size of 5%, and then adjusted to 110% of the rated voltage to acquire noise array signals outside the reactor.

3. The method for detecting aging of epoxy materials of iron core reactor based on noise characteristics according to claim 2 is characterized in that: When setting the aging state of the epoxy material, the epoxy pressure plate is loosened using a torque wrench. The degrees of loosening are 4 N·m, 8 N·m, and 12 N·m, respectively. This forms three corresponding aging fault states, namely, slight aging fault, moderate aging fault, and severe aging fault. In each fault state, the excitation voltage is adjusted in turn according to the above-mentioned voltage step, and the noise signals are collected respectively.

4. The method for detecting aging of epoxy materials of iron core reactor based on noise characteristics according to claim 1 is characterized in that: In step 2, the noise signal is subjected to double-tree complex wavelet decomposition, and the noise variance of each layer of wavelet coefficients after double-tree complex wavelet decomposition is obtained by using double-tree complex wavelet threshold denoising method based on Bayesian maximum a posteriori estimation, and the noise variance is used as the denoising threshold. After thresholding the wavelet coefficients of each layer after double-tree complex wavelet decomposition of the noise signal, the sound signal of the epoxy material of the core reactor after filtering out the high-frequency noise is obtained by inverse double-tree complex wavelet transform.

5. The method for detecting aging of epoxy materials of iron core reactor based on noise characteristics according to claim 4 is characterized in that: After the signal distribution processed by the beamforming algorithm is interpolated by using the cosine interpolation method, the Gram angular field method (GAF) is used to generate array signal images of the same size for the signal.

6. The method for detecting aging of epoxy materials of iron core reactor based on noise characteristics according to claim 1, characterized in that: In step 3, the input of the image feature extraction model is the i-th array signal image A in the normal set i , the learned features are output as H(x), and the residual unit of the model is shown in the following formula: Among them, y n represents the nth residual unit, and They represent the input and output of the nth residual unit respectively; F() is the residual function, which represents the learned residual result, and W n is the convolution kernel of the nth layer of the model, and f is the ReLU activation function; The feature extraction results from the 1st layer to the Lth layer of the model are shown in the following formula: in, It is the feature result extracted from the 1st layer to the Lth layer.

7. The method for detecting aging of epoxy materials of iron core reactor based on noise characteristics according to claim 6 is characterized in that: The training content of the image feature extraction model includes: using the initial convolution layer of the model to extract edge information of the image through convolution calculation; based on the aforementioned edge information, using the middle convolution layer to continue to use convolution operation to extract local information of the image; using the convolution layer to extract high-level semantic features of the image, based on the aforementioned high-level semantic features, saving the model network weight parameters with the above features, and using the model with the weight parameters as a model that can extract N-dimensional deep features of the normal condition of the epoxy material of the iron core reactor; The model with weight parameters is used to input a slight set, a moderate set, and a severe set respectively, and the N-dimensional deep features of slight aging faults, moderate aging faults, and severe aging faults of epoxy materials of iron core reactors are obtained.

8. The method for detecting aging of epoxy materials of iron core reactor based on noise characteristics according to claim 1, characterized in that: In step 4, the images of the normal set, mild set, moderate set, and severe set are grayscaled, the images are color and space normalized using the gamma correction method, the gradient size and direction of each pixel in the image are calculated, the image is divided into cells according to the set size, the gradient histogram in each cell is calculated, a block is formed according to the set number of cells, the gradient features in each block are calculated, and the gradient features of all blocks in the image are combined to obtain the HOG features of the image.

9. The method for detecting aging of epoxy materials of iron core reactor based on noise characteristics according to claim 1, characterized in that: In step 6, the distance comparison is calculated as follows: Among them, when X=1, that is, D1 qs H represents the Euclidean distance between the qth one-dimensional sequence of the mild set and the sth one-dimensional sequence of the normal set. s1 Represents the first value in the sth one-dimensional sequence of the normal set, H sB Represents the last value in the s-th one-dimensional sequence of the normal set, L1 q1 Represents the first value in the qth one-dimensional sequence of the minor set, L1 qB The last value in the qth one-dimensional sequence represents a mild set, that is, each sequence has B values, X=2 represents a moderate set, and X=3 represents a severe set.

10. The method for detecting aging of epoxy materials of iron core reactors based on noise characteristics according to claim 9, characterized in that: Perform distance comparison calculation on the images in the normal set to obtain n! distance comparison results, and select the distance result with the largest value as the threshold 1, i.e., α1; Secondly, the C distance comparison results between the mild set and the normal set are normalized, and the mapping range is [0, 1]. The normalized comparison results are classified according to [0, 0.2], (0.2, 4], (0.4, 0.6], (0.6, 0.8], (0.8, 1], and then the probability of each interval is obtained. The distance comparison result corresponding to the largest value of the normalized comparison result in the interval with the largest probability is selected as α′2; The C distance comparison results between the moderate set and the normal set are processed in the same way as α′3; Finally, the C distance comparison results between the severe set and the normal set are processed in the same way as α′4; Threshold 2, i.e., α2, and threshold 3, i.e., α3, are expressed as follows:

11. The method for detecting aging of epoxy materials of iron core reactor based on noise characteristics according to claim 1 or 10, characterized in that: In step 7, the aging of the epoxy material of the reactor is determined according to the distance comparison result between the one-dimensional sequence of the epoxy material characteristics of the core reactor in the detection period and the one-dimensional sequence of the normal set characteristics and the set threshold, as shown in the following formula: Where: result represents the aging of the epoxy material of the reactor, c' represents the average value of the distance comparison result between the one-dimensional sequence converted by merging the N-dimensional deep features of the image with the HOG features and the normal set samples, J i The distance comparison result between the one-dimensional sequence converted by merging the N-dimensional deep features and HOG features of the image converted from the vibration signal of the epoxy material of the core reactor during the detection period and the one-dimensional sequence converted by merging the N-dimensional deep features and HOG features of the i-th image in the normal set.

12. A noise feature-based iron core reactor epoxy material aging detection system for implementing the method according to any one of claims 1 to 11, comprising: Reactor epoxy material signal acquisition module, signal processing module, signal to image module, image feature extraction model building module, HOG feature extraction module, reactor epoxy material state judgment module; characterized by: The reactor epoxy material signal acquisition module collects noise signals of the reactor in normal state and epoxy material aging state respectively through the microphone noise array; The signal processing module is used to pre-process the collected noise signal to remove the noise, and use the beam forming algorithm to extract the relevant sound signal of the epoxy material of the iron core reactor from the denoised signal; The signal-to-image module is used to convert the processed sound signal into an image as a sample set; The image feature extraction model construction module is used to construct an image feature extraction model using Resnet and train the normal set, mild set, moderate set, and severe set; The HOG feature extraction module is used to extract HOG features from the images of the normal set, mild set, moderate set, and severe set; The reactor epoxy material status judgment module is used to merge the N-dimensional deep features of each image with the HOG features, and convert them into a one-dimensional sequence as a basic sample. The one-dimensional sequences of the mild set, moderate set, and severe set are respectively compared with the one-dimensional sequence of the normal set for distance calculation, and the threshold is set. The aging condition of the reactor epoxy material is judged according to the distance comparison results between the one-dimensional sequence of the iron core reactor epoxy material characteristics of the time period to be detected and the one-dimensional sequence of the normal set characteristics.

Images