Unbalanced small sample event rapid classification method and system suitable for phi-OTDR (Optical Time Domain Reflectometer) system

Abstract

The invention provides a method and a system for quickly classifying unbalanced small sample events of a phi-OTDR (Optical Time Domain Reflectometer) system, aiming at obtaining higher classification accuracy and shorter classification time delay. The method comprises the following steps: firstly, in a data set construction stage, expanding an unbalanced small sample set into a sample set meeting requirements by providing a space correlation noise fusion sample enhancement method, so as to ensure sample requirements in model training; then, in order to solve the problem of signal classification of different time scales, a two-dimensional image construction method of a Gramb angle field of time scale transformation is provided, and rapid classification of non-fixed time scale events is solved by converting signals into two-dimensional images and performing interpolation operation. And finally, in order to reduce the model classification time delay and the requirements on computing resources, a high-precision intrusion signal classification model MobileNet-DAS is provided based on a lightweight network, so that the signal classification resource requirements and the classification time delay are reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of distributed optical fiber signal processing, and in particular relates to a fast classification method for unbalanced small sample events applicable to a Φ-OTDR intrusion detection system. Background Art

[0002] Distributed fiber optic sensing technology is a continuous distributed sensing technology that uses optical fiber as a transmission medium and a sensitive element, which can achieve long-distance continuous distributed measurement. Among them, the distributed fiber optic system (Ф-OTDR) based on the principle of phase-sensitive optical time domain reflectometry not only has the advantages of general distributed fiber optic sensing technology, but also has the characteristics of high sensitivity and fast response speed. It has been applied in the field of perimeter security intrusion detection of some important facilities such as railways and ports. In the early distributed fiber optic intrusion detection systems, researchers mainly focused on improving the system signal-to-noise ratio, spatial resolution, perception time, monitoring distance and other indicators. In recent years, significant improvements have been made in the signal-to-noise ratio, monitoring distance and perception events of the Ф-OTDR system. For example, the maximum perception distance can reach 170km. However, in the Ф-OTDR intrusion monitoring system, it is conducive to quickly and accurately distinguishing the event type, which is conducive to timely proposing response strategies and reducing property losses. However, in complex environments, it is difficult to quickly and accurately distinguish the event perception type due to the influence of environmental noise and human actions, which will become a bottleneck for the widespread application of the Ф-OTDR intrusion monitoring system.

[0003] With the rapid development of artificial intelligence technology, data-driven artificial intelligence pattern recognition technology provides a new solution for the classification of Φ-OTDR intrusion detection system events. However, the existing pattern recognition methods can achieve good classification results under the condition of sufficient samples, but for some destructive events, facilities cannot be destroyed multiple times. Under the condition of unbalanced small samples, when the data-driven artificial intelligence method is applied to the classification of intrusion signals in the Φ-OTDR system, it is easy to have underfitting or overfitting of the model, which in turn affects the accuracy of intrusion event separation.

[0004] At present, some scholars have proposed some Ф-OTDR intrusion detection classification methods. In order to reduce the false alarm rate (NAR) of the system, He et al. [1] extracted the spectral entropy and band variance of the signal as frequency domain features, and proposed a railway perimeter intrusion detection method based on the XGBoost algorithm, with an average recognition accuracy of up to 98.5%. Saleh et al. [2] proposed a human activity classification method for Ф-OTDR intrusion detection system, using the gammatone filter cepstral coefficient as input feature and using coexistence support vector machine, the classification accuracy can reach 95%. Jia et al. [3] proposed a method based on NC-SVM (Near Category Support Vector Machines), which first extracted 15 features such as the centroid frequency, mean square frequency and root mean square frequency of the event signal, and then expanded the SVM multi-classifier through the KNN algorithm to achieve the classification of similar events. The neural network-based Ф-OTDR intrusion detection system intrusion signal classification method can automatically extract features from the event signal, thereby achieving higher classification accuracy. Shi et al. [4] designed a Φ-OTDR event classification method based on convolutional neural network (CNN). This method first converts the signal into a grayscale image, and then uses the image classification method to classify 5 types of events, with a classification accuracy of 96.67%. Sun et al. [5] proposed a convolutional neural network event classification method based on pulse scanning images. First, the Φ-OTDR signal is converted into a pulse scanning image, and then a convolutional neural network is designed to classify 5 types of events, with a classification accuracy of 98%. Barantsov et al. [6] proposed a two-stage Φ-OTDR event classification method. This method first designs a bandpass filter to obtain event samples, and then implements event classification based on the AlexNet neural network, with a classification accuracy of 98%. Tian et al. [7] proposed a temporal neural network based on long short-term memory network and double-layer attention mechanism, achieving a 0% false negative rate and a classification accuracy of 93.4%.

[0005] The above method can achieve good classification results under the condition of sufficient samples, but for some destructive events, facilities cannot be destroyed multiple times [8]. Under the condition of unbalanced small samples, when the data-driven artificial intelligence method is applied to the classification of intrusion signals in the Φ-OTDR system, it will be easy to have underfitting or overfitting of the model, which will affect the accuracy of intrusion event separation. Small sample learning is a learning paradigm that can obtain a model with high classification accuracy under the condition of a small number of samples. Data enhancement is an effective way to achieve small sample learning. By enhancing the unbalanced small samples, the model's fit can be improved. Shi et al. [9] proposed a data enhancement method based on time series transformation and cycle generative adversarial network (CycleGAN). By expanding rare samples, the classification accuracy of small samples reached 79.28%. However, this method assumes that two types of events are small samples, while other events have more than 800 samples. In actual scenarios, the number of samples is difficult to meet the model training requirements. However, in the intrusion detection scenario, the surrounding environmental noise is easy to obtain, while it is difficult to collect a large number of intrusion event samples. If a large number of samples are collected, the deployment efficiency of the system will be seriously affected. At the same time, the above methods assume that different events have the same time scale, while for events such as climbing over a fence and destroying a fence, the time scales are very different. In addition, existing methods use traditional neural network models (such as ResNet151, ResNet121, VGG16, VGG19, etc.) in classification. When such models are deployed in intrusion detection systems, the system has limited computing resources, which will increase the model inference delay and make it difficult to meet the real-time detection of intrusion events.

[0006] References:

[0007] [1]H.Meng, S.Wang, C.Gao and F.Liu, "Research on Recognition Method ofRailway Perimeter Intrusions Based on Φ-OTDR Optical Fiber SensingTechnology," in IEEE Sensors Journal, vol.21, no.8, pp.9852-9859, 15April15, 2021, doi:10.1109 / JSEN.2020.3043193.

[0008] [2]Saleh,NurLuqman,et al."Human activities classification based onφ-OTDR system by utilizing gammatone filter cepstrum coefficient envelope usingsupport vector machine."Optics&Laser Technology 164(2023):109417.

[0009] [3]H.Jia,S.Liang,S.Lou and X.Sheng,"Ak-Nearest Neighbor Algorithm-Based Near Category Support Vector Machine Method for Event Identification of -OTDR,"in IEEE Sensors Journal,vol.19,no.10,pp.3683-3689,15May15,2019,doi:10.1109 / JSEN.2019.2891750.

[0010] [4]Shi,Yi,et al."An event recognition method forΦ-OTDR sensingsystem based on deep learning."Sensors 19.15(2019):3421.

[0011] [5]Sun,Qian,et al."Pattern recognition based on pulse scanningimaging and convolutional neural network for vibrational events inΦ-OTDR."Optik 219(2020):165205.

[0012] [6]Barantsov, Ivan A., et al. "Classification of Acoustic InfluencesRegistered Sensors with Phase-Sensitive OTDR Using Pattern RecognitionMethods." 23.2(2023):582.

[0013] [7]Tian, ​​Manling, et al. "Temporal convolution network with a dualattention mechanism for -OTDR event classification."Applied Optics 61.20(2022):5951-5956.

[0014] [8] Ramadan, Salah, et al. "Detection of stress corrosion cracking ofhigh-strength steel used in prestressed concrete structures by acousticemission technique." Applied surface science 254.8 (2008): 2255-2261.

[0015] [9]Shi, Yi, et al. "Event recognition method based on dual-augmentation for anΦ-OTDR system with a few training samples." Optics Express 30.17(2022):31232-31243. Summary of the invention

[0016] To solve the above problems, the present invention proposes a fast classification method for unbalanced small sample events applicable to a Φ-OTDR intrusion detection system, which specifically includes the following steps:

[0017] Step 1, obtaining an intrusion signal sensed by the Φ-OTDR and constructing an unbalanced small sample set, wherein the intrusion signal includes multiple disturbance events and noise signals;

[0018] Step 2: Use the spatial correlation noise fusion data enhancement method to expand the unbalanced small sample set into a sample set that meets the requirements; specifically, the following steps are included:

[0019] Step 2.1, spatial correlation data enhancement: First, prove that the two-dimensional space-time signal has spatial correlation at adjacent distance positions; then perform spatial correlation decomposition on the multi-dimensional disturbance event signal matrix to obtain multiple one-dimensional vibration signals;

[0020] Step 2.2, using the spatial correlation threshold noise reduction method to process the vibration signal with background noise to remove the background noise signal;

[0021] Step 2.3, noise fusion data enhancement; perform wavelet decomposition on the result of step 2.2 to obtain coefficients of different frequencies, perform soft threshold denoising on the frequency coefficients after wavelet decomposition to filter out the high-frequency noise part, reconstruct the denoised frequency coefficients to obtain a one-dimensional time series signal without high-frequency noise, intercept the noise matrix at different positions with the same time scale as the event, perform spatial separation on the noise matrix, and fuse the noise with the denoised signal;

[0022] Step 3, converting the intrusion signal in the sample set into a two-dimensional image through a two-dimensional image construction method based on the Gram angle field of adaptive time scale transformation and performing interpolation operations to obtain an image suitable for network input;

[0023] Step 4: Build the MobileNetV3-DAS network and use the trained network model to quickly classify disturbance events.

[0024] Furthermore, in step 2.1, it is assumed that the phase shift caused by the disturbance event t is represented by the matrix E t According to formula (1), the element x in the matrix ij represents the phase change of point j on the distributed optical fiber at time i, where i∈[1,m], j∈[1,n], m represents the number of one-dimensional signals, and n represents the length of the one-dimensional signal sequence; for event E t Separate in space to obtain multiple time series phase signals e k (k∈[1,n]), and each timing signal e k Both record the time domain characteristics of the disturbance event t;

[0025]

[0026] e k =[x k1 , x k2 , ..., x kn ] T , (k∈[1,m]) (2).

[0027] Furthermore, in step 2.2, define E t The energy matrix is ​​given by R t It is expressed as shown in formula (3), where r i (i∈[1,n]) represents the intrusion disturbance event e i The energy of (i∈[1,n]);

[0028] R t =[r 1 , r 2 ,…,r n ] (3)

[0029] For the event timing signal e k (E t ∈e k , k∈[1, n]), firstly, it is converted from the time domain sequence to the frequency domain sequence f by formula (4) k , and the event signal e is obtained by formula (5) k Energy k , where n represents the length of the signal sequence, e k (i) represents event e k Phase change at time i;

[0030]

[0031] Then, from R t Find the time series signal e with the largest energy t (t∈[1,n]), and the timing signal e is calculated by formula (6) k With signal e k The correlation degree of (k≠i) is s ki , and E t The correlation S of other signals in t As shown in formula (7); finally, the denoised and enhanced event samples are obtained;

[0032]

[0033] S t =[s k1 ,s k2 ,…,s k(n-1 )], (k∈[1,n-1]) (7).

[0034] Furthermore, in step 2.3, let e' k For E' t An element of N t is related to event E' t Noise signals occurring at different locations at the same time and with E' thave the same time scale, as shown in formula (9), n k (k∈[1,l]) represents the same as e' k The noise signal of the same time scale is shown in formula (10);

[0035] e' k =[y k1 ,y k2 , ..., y kη ] T , (k∈[1,n],η≤m) (8)

[0036] N t =[n 1 , n 2 , ..., n l ] (9)

[0037] n η =[b 1 , b 2 , ..., b m ] T , (η∈[1,l]) (10)

[0038] Wavelet soft threshold denoising method is used to remove e' k The original environmental noise in the original event signal e k Decompose; then, the kth coefficient of the jth layer after quantization is obtained by formula (14): Where λ is the threshold value, calculated by formula (12), where idwt() is the wavelet reconstruction function;

[0039]

[0040] z' k =idwt(ω j，k ), (14)

[0041] Finally, the denoised signal z' k With the noise signal n η By superimposing formula (14), we get the new signal α k :

[0042] α k =n η +z' k (15).

[0043] Furthermore, the specific processing process of step 3 is as follows:

[0044] First, for the enhanced phase change sequence α k, obtained by step 2, the phase sequence is normalized to the interval [-1, 1] using formula (16) to obtain the vector Among them, n k (i) represents a one-dimensional time series signal n k The i-th element of ;

[0045]

[0046] Then, by The radius and angle corresponding to each phase element are calculated by formula (17), where the length() function is used to return a vector Length;

[0047]

[0048] Among them, L represents the length of the time series; finally, the matrix G is obtained by using the sum angle relationship GSF or the sum difference relationship ASF and G ADF Each element in the matrix, such as formula (18) and formula (19), is put into the range of 0 to 255 by formula (20), where O(i, j) represents the pixel value at i, j in the image, int(·) is the rounding function, and G(i, j) represents the Gram matrix G ASF or G ADF The element value of the i-th row and i-th column;

[0049]

[0050]

[0051] O(i,j)=int(127.5×(G(i,j)+1)) (20)

[0052] Let the image obtained by Gram's angular field method be image O k×k , and the image scale that satisfies the network input is Q l×l , the scaling ratio ψ is calculated by formula (21), where k is the width and height of the original GAF image. At this time, the position u of each pixel on the coordinate axis in the target image Q is 0 and v 0 As shown in formula (22), the image Q is enlarged by α times in the x-axis direction and β times in the y-axis direction, where α and β can be obtained by the scale ratio between the original image and the target image;

[0053]

[0054] Q(u 0 , v 0 )=O[a(x,y),b(x,y)] (22)

[0055]

[0056] Where a(x, y) and b(x, y) represent the calculation of the coordinates (x, y) using functions a and b, as shown in formula (23); if the scaled u 0 and v 0 When it is not on the coordinate axis, the bilinear interpolation method is used to obtain Q(u 0 , v 0 ), the calculation method is shown in formula (24);

[0057]

[0058] Among them, u 0 and v 0 It represents the result of processing (x, y) by functions a and b, that is, the position of the pixel point on the coordinate axis of the image after scale reconstruction.

[0059] Furthermore, the processing of the MobileNetV3-DAS network is as follows:

[0060] Set the image to X, first perform ordinary two-dimensional convolution C2H 3x3 / 2 feature extraction to obtain X 1 , then X 1 Input into the deep separable convolution BottleNeck based on ECA attention improvement for further feature extraction Figure X 2 , and finally the feature Figure X 2 Perform a maximum pooling MaxPooling, three two-dimensional convolutions and Softmax calculations to obtain the final features. Figure X out (1, 1, k); where C2H 3x3 / 2 indicates that the two-dimensional convolution is composed of a two-dimensional convolution and an h-swish activation function, and the convolution kernel size is 3x3, the stride is 2, and the h-swish activation function is shown in formula (29);

[0061] ReLU(x)=min(max(0,x),6) (28)

[0062]

[0063] Among them, x is the feature map input by the h-swish activation function.

[0064] Furthermore, the three two-dimensional convolutions include one two-dimensional convolution C2H 1x1 / 1 and two two-dimensional convolutions C2NBN 1x1 / 1, where C2H 1x1 / 1 indicates that the module consists of a two-dimensional convolution and an h-swish activation function, and the convolution kernel size is 1x1 and the stride is 1; C2 NBN 1x1 / 1 indicates that a two-dimensional convolution with a convolution kernel of 1x1 and a stride of 1 is used here to calculate the feature map, and the convolution calculation results are not normalized.

[0065] Furthermore, based on the ECA-improved depthwise separable convolution, the structure performs in First, a 1×1 normal volume is used for feature extraction, and then the result is input into a 3×3 channel-by-channel convolution to obtain D' in , the weight vector Φ with the same length as the number of channels is calculated through the ECA attention mechanism, and the weight vector Φ is combined with the feature map D' in Fusion, so that the neural network focuses on the feature map that best represents the signal type Finally, the feature map Through point-by-point convolution calculation, the output feature map D' of a single depth-separable convolution is obtained out ; Among them, the ECA attention mechanism performs average pooling, point-by-point convolution and Sigmoid activation on the input feature map in sequence to obtain a weight vector with the same number as the feature maps.

[0066] The present invention also provides a system for rapid classification of unbalanced small sample events applicable to a Φ-OTDR system, comprising a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a method for rapid classification of unbalanced small sample events applicable to a Φ-OTDR system as described in the above technical solution.

[0067] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0068] The present invention provides a method for rapid classification of unbalanced small sample events suitable for Φ-OTDR intrusion detection systems, in order to obtain higher classification accuracy and shorter classification delay. In the data set construction stage, a spatial correlation noise fusion sample enhancement method is proposed to expand the unbalanced small sample set into a sample set that meets the requirements, thereby ensuring the sample requirements in model training and reducing the sample balance imbalance rate to 1.09. In order to solve the signal classification problem of different time scales, a two-dimensional image construction method of the Gram angle field of time scale transformation is proposed. By converting the signal into a two-dimensional image and interpolating the operation, the rapid classification of events with non-fixed time scales is solved. Finally, in order to reduce the model classification delay and the requirements for computing resources, a high-precision intrusion signal classification model MobileNet-DAS is proposed based on a lightweight network, so that the classification progress of small sample unbalanced samples reaches 97%, and the classification event is reduced to 16.62ms. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 Schematic diagram of spatial correlation of event phase changes in an embodiment of the present invention;

[0070] Figure 2 Schematic diagram of spatial correlation of event signals in an embodiment of the present invention;

[0071] Figure 3 This is the SCDE algorithm flow in the embodiment of the present invention;

[0072] Figure 4 This is the SCNFE algorithm flow in the embodiment of the present invention;

[0073] Figure 5 This is a flow chart of a method for generating a two-dimensional image based on a GAF field in an embodiment of the present invention;

[0074] Figure 6 This is a schematic diagram of the MobileNetV3-DAS network structure in an embodiment of the present invention;

[0075] Figure 7 Schematic diagram of the structure of a depth-wise separable convolution module in an embodiment of the present invention;

[0076] Figure 8 This is a flow chart of an embodiment of the present invention;

[0077] Fig. 9 This is a schematic diagram of an experimental scenario of an embodiment of the present invention;

[0078] Fig.10 This is a schematic diagram of the experimental equipment and optical cable deployment of an embodiment of the present invention;

[0079] Fig.11 is a phase diagram of an intrusion event in an embodiment of the present invention;

[0080] Fig.12 This is a diagram showing the experimental results of the SCE method increase method according to an embodiment of the present invention;

[0081] Fig.13 This is a diagram showing the experimental results of the spatial correlation denoising and enhancement method according to an embodiment of the present invention;

[0082] Fig.14 It is a cosine similarity graph of the results obtained by different data enhancement methods according to the embodiment of the present invention;

[0083] Fig.15 This is a diagram showing the experimental results of the spatial correlation noise fusion enhancement method according to an embodiment of the present invention;

[0084] Fig.16 is a sample quantity diagram of different data enhancement methods according to an embodiment of the present invention;

[0085] Fig.17 This is a graph showing the accuracy of the MobileNetV3-DAS non-enhancement method according to an embodiment of the present invention;

[0086] Fig.18 This is a confusion matrix diagram of the MobileNetV3-DAS experimental results of an embodiment of the present invention;

[0087] Fig.19 Graph showing classification accuracy of different classification models according to an embodiment of the present invention. DETAILED DESCRIPTION

[0088] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0089] like Figure 8 As shown, an embodiment of the present invention provides a method for fast classification of unbalanced small sample events applicable to a Φ-OTDR intrusion detection system, comprising the following steps:

[0090] Step 1, obtaining an intrusion signal sensed by the Φ-OTDR and constructing an unbalanced small sample set, wherein the intrusion signal includes multiple disturbance events and noise signals;

[0091] Step 2, using the spatial correlation noise fusion data enhancement method to expand the unbalanced small sample set into a sample set that meets the requirements; including the following sub-steps:

[0092] Step 2.1, spatial correlation data enhancement: First, prove that the two-dimensional space-time signal has spatial correlation at adjacent distance positions; then perform spatial correlation decomposition on the multi-dimensional disturbance event signal matrix to obtain multiple one-dimensional vibration signals;

[0093] When the Φ-OTDR-based distributed fiber intrusion detection system senses an external intrusion, it will cause the Rayleigh backscattered light inside the fiber to change phase. By extracting the phase change characteristics of different events through a neural network, different event types can be determined. When an intrusion occurs at a certain point, the phase change will produce similar changes in the adjacent space, and gradually weaken from the center of the event to both sides, such as Figure 1 (a) shows the phase change diagram of the intrusion event. Figure 1 (b) is a time-series phase change diagram after spatial expansion, where different colors represent the phase change at different positions.

[0094] Since the signal phase change at 6 meters is the largest, it can be used as a reference signal. There is a correlation signal between this position signal and other position signals. The correlation is calculated by cosine similarity, such as Figure 2 shown.

[0095] According to the above law, it is assumed that the phase change caused by the disturbance event t is represented by the matrix E t According to formula (1), the element x in the matrix ij (i∈[1, m], j∈[1, n]) represents the phase change of point j on the distributed optical fiber at time i, where m represents the number of one-dimensional signals and n represents the length of the one-dimensional signal sequence. t Separate in space to obtain multiple time series phase signals e k (k∈[1,n]), and each timing signal e k The spatial correlation enhancement method (ESC, Enhanced Spatial Correlation) can record the time domain characteristics of the disturbance event E t The number of samples is expanded to n times of the original, thereby increasing the number of samples for deep learning Φ-OTDR intrusion signal classification.

[0096]

[0097] e k =[x k1 ,x k2 ,...,z kn ] T , (k∈[1,m]) (2)

[0098] Step 2.2, use the spatial correlation threshold noise reduction method to process the vibration signal with background noise to remove the background noise signal; the specific processing process of spatial correlation threshold noise reduction enhancement is as follows:

[0099] (1) Extract a disturbance event time series signal from the two-dimensional spatial disturbance signal;

[0100] (2) The disturbance time series signal is converted into the frequency domain by Fourier transform, and the signal energy is calculated according to the frequency spectrum. The specific method is shown in formulas (4) and (5).

[0101] (3) The energy parameters of different sequences are saved, and the signal with the largest energy value is removed as the base signal, as shown in formula (3).

[0102] (4) Calculate the cosine similarity between other related signals and the base signal, and use the similarity threshold γ to remove signals less than the correlation threshold from the original signal. The specific operation is shown in formula (6) and formula (7). The threshold correlation removal process is as follows: Figure 3 shown.

[0103] In practical engineering, it is difficult to accurately cut out the phase change matrix E of the disturbance event t from the ambient noise. t Noise from other spatial locations will be introduced. At this time, if the spatial correlation enhancement method is used to enhance the training samples, the environmental noise will also be marked as an event signal, which will affect the ability of the deep learning model to extract signal features, thereby reducing the classification accuracy of the intrusion signal. Therefore, the present invention proposes a spatial correlation denoising enhancement method (SCDE), which removes the environmental noise samples marked as events in the samples according to the energy of different intrusion disturbance signals and the similarity between disturbance signals. Definition E t The energy matrix is ​​given by R t It is expressed as shown in formula (3), where r i (i∈[1,n]) represents the intrusion disturbance event e i The energy of (i∈[1,n]).

[0104] R t =[r 1 , r 2 ,…,r n ] (3)

[0105] For the event timing signal e k (E t ∈e k , k∈[1,n]), firstly, it is converted from the time domain sequence to the frequency domain sequence f by formula (4) k , and the event signal e is obtained by Fourier transform through formula (5) k Energy k , where n represents the length of the signal sequence, e k (i) represents event e k Phase change at time i.

[0106]

[0107] Then, from R t Find the time series signal e with the largest energy t (t∈[1,n]), and the timing signal e is calculated by formula (6) k With signal e k The correlation degree of (k≠i) is s ki , and E t The correlation S of other signals in t As shown in formula (7). Finally, the denoised and enhanced event samples are obtained through the SCDE algorithm.

[0108]

[0109] S t =[s k1 ,s k2 ,…,s k(n-1 )],(k∈[1,n-1]) (7)

[0110] The SCDE algorithm process proposed by the present invention is as follows Figure 3 As shown, first, a single sample E is taken from the total sample E t , the spatial correlation method is used to t Perform sample enhancement and perform the enhanced event signal e i Perform frequency domain conversion and calculate signal energy r i ; Then, the energy r i Save in R t and from R t Get the signal with the largest energy value e k , returns the subscript k of the signal with the maximum energy; finally, calculate the signal e k With e i The cosine similarity s of (i∈[1,n-1]) ki , and save the signal with similarity threshold greater than γ and the subscript in R t In, from R t Returns the signal index and removes the signal that meets the similarity requirement from E t Copy to E' t As the algorithm returns the result; where max(·) is the function for obtaining the maximum value in the sequence, and index(·) is the function for returning the index corresponding to the parameter.

[0111] Step 2.3: noise fusion data enhancement; the specific processing process is as follows:

[0112] (1) Perform wavelet decomposition on the result of step 2.2 to obtain coefficients of different frequencies. The specific operation is shown in formula (11).

[0113] (2) Perform soft threshold denoising on the frequency coefficients after wavelet decomposition to filter out the high-frequency noise part. The specific operation is shown in formula (12).

[0114] (3) The denoised frequency coefficients are reconstructed to obtain a one-dimensional time series signal without high-frequency noise. The specific operation is shown in formulas (13) and (14).

[0115] (4) The noise matrices at different positions with the same time scale as the event are intercepted, and the noise matrices are spatially separated, and the noise is fused with the denoised signal. The specific operation is shown in formula (15).

[0116] After the above SCDE algorithm, the signal E t The environmental noise samples marked as events are removed from the algorithm. The algorithm can expand the original samples to n times of the original at most, and obtain the enhanced sample set E' t However, it cannot solve the imbalance of different classes in the sample, which will easily lead to overfitting or underfitting of the model in some categories. To solve this problem, the present invention further proposes a spatial correlation noise fusion enhancement method (SCNFE).

[0117] Let e' k For E' t An element of N t is related to event E' t Noise signals occurring at different locations at the same time and with E' t have the same time scale, as shown in formula (9), n k (k∈[1,l]) represents the same as e' k The noise signal of the same time scale is shown in formula (10).

[0118] e' k =[y k1 ,y k2 , ..., y kη ] T , (k∈[1,n],η≤m) (8)

[0119] N t =[n 1 ,n 2 , ..., n l ] (9)

[0120] n η =[b 1 , b 2 , ..., b m ] T, (η∈[1,l]) (10)

[0121] Since the event disturbance signal is mainly distributed in the low frequency range and the noise signal is mainly distributed in the high frequency range, the present invention adopts the wavelet soft threshold denoising method to remove e' k First, the original event signal e k Decompose; then, the kth coefficient of the jth layer after quantization is obtained by formula (14): Where λ is the threshold value, calculated by formula (12), and idwt() is the wavelet reconstruction function.

[0122]

[0123] z' k =idwt(ω j，k ), (14)

[0124] Finally, the denoised signal z' k With the noise signal n η By superimposing formula (14), we get the new signal α k The SCNFE algorithm process proposed by the present invention is as follows Figure 4 shown.

[0125] SCNFE algorithm first event e' k The internal environmental noise is removed by wavelet threshold denoising method to obtain the signal z' k , and then compared with the background noise signal n on the same time scale η Fusion generates a new signal α k The SCNFE algorithm is used to expand small sample events, thus solving the problem of small sample quantity and imbalance of samples of different categories in the small sample recognition process.

[0126] Step 3, by proposing a two-dimensional image construction method of the Gram angle field with adaptive time scale transformation, the intrusion signal in the sample set is converted into a two-dimensional image to obtain an image suitable for network input;

[0127] Gramian Angular Field (GAF) is a method that can convert a one-dimensional time series signal into a two-dimensional image. During the conversion process, the method can retain all the information in the original one-dimensional intrusion signal and transfer it to an image with rich feature information. While ensuring the integrity of the time series signal, the time series dependency problem is solved. Therefore, the method is suitable for signal pattern recognition. However, in engineering applications, the length of the phase change signal of the event is not fixed, which makes it impossible for images with inconsistent sizes to be used as input to the network. Therefore, the present invention proposes an improved Gramian Angular Field image conversion method, which obtains an image scale that meets the requirements of the classification network by enlarging or shrinking GAF.

[0128] First, for the enhanced phase change sequence α k , the phase sequence needs to be normalized to the interval [-1, 1] using formula (16) to obtain the vector Among them, n k (i) represents a one-dimensional time series signal n k The ith element of .

[0129]

[0130] Then, by The radius and angle corresponding to each phase element are calculated by formula (17), where the length() function is used to return a vector Length;

[0131]

[0132] Finally, the matrix G is obtained by using the sum angle relationship GSF or the sum difference relationship from formula (18) and formula (19): AsF and GA DF Each element in the matrix is ​​put into the range of 0 to 255 by formula (20), where O(i, j) represents the pixel value at i, j in the image, int(·) is the rounding function, and G(i, j) represents the Gram matrix G ASF or G ADF The element value of the i-th row and i-th column of .

[0133]

[0134] O(i,j)=int(127.5×(G(i,j)+1))(20)

[0135] Suppose the image obtained by the GAF method is image O k×k , and the image scale that satisfies the network input is Q l×l, to meet the requirements of the MobileNetV3-DAS network of the present invention, l is set to 224, and the scaling ratio ψ is calculated by formula (21), where k is the width and height of the original GAF image, and l is the width and height of the scaled image. At this time, the position u of each pixel on the coordinate axis in the target image Q 0 and v 0 It can be expressed by formula (22), then the image Q is enlarged by α times in the x-axis direction and β times in the y-axis direction, where α and β can be obtained by the scale ratio of the original image to the target image.

[0136]

[0137] Q(u 0 , v 0 )=O[a(x,y),b(x,y)] (22)

[0138]

[0139] Where a(x, y) and b(x, y) represent the calculation of the coordinates (x, y) using functions a and b, as shown in formula (23); if the scaled u 0 and v 0 When it is not on the coordinate axis, the bilinear interpolation method can be used to obtain Q(u 0 , v 0 ), and the calculation method is shown in formula (24).

[0140]

[0141] Among them, u 0 and v 0 It represents the result of processing (x, y) by functions a and b, that is, the position of the pixel point on the coordinate axis of the image after scale reconstruction.

[0142] In step 4, a MobileNetV3-DAS network was designed to complete model training and fast classification of intrusion signals.

[0143] DAS devices are generally resource-constrained devices, and their computing resources are mainly used for sensor signal demodulation and visualization. Traditional convolutional neural network (such as VGG, ResNet, DenseNet, GoogleNet, etc.) model inference also requires a lot of computing resources. Therefore, a lightweight model is needed for real-time signal classification on resource-constrained devices. MobileNetV3 is a lightweight network proposed by Google in 2019. Its core idea is to use deep separable convolution instead of traditional convolution operations, and introduce SE attention mechanism, inverted residual module and activation function Hardswish on the basis of MobileNetV2 to fully extract image features.

[0144] The depthwise separable convolution used in the MobileNet network first calculates the channel-by-channel convolution and then the point-by-point convolution. Compared with ordinary convolution, the depthwise separable convolution extracts the same feature results as ordinary convolution with less calculation, but it has fewer training parameters. Suppose the number of channels contained in the network input feature map is M, and the size of the convolution kernel acting on the feature map is D in ×D in , the dimension of the output feature map is D out ×D out ×N, the computational cost λ of the traditional feature map is shown in formula (25). The computational cost ρ of the depthwise separable convolution is expressed by formula (26), and the ratio of the computational cost of the depthwise separable convolution to the conventional convolution is expressed by formula (27). When the convolution kernel is 3×3, the ratio of the computational cost is approximately 1 / 9.

[0145] λ=D i ×D i ×D o ×D o ×M×N (25)

[0146] ρ=D i ×D i ×D o ×D o ×M+D o ×D o ×M×N (26)

[0147]

[0148] However, when MobileNetV3 is applied to DAS devices to classify perception signals, the SE mechanism used by the network reduces the complexity of the model while also destroying the direct correspondence between the channel and its weight, resulting in low accuracy in classifying similar vibration signals. In addition, the SE attention-based MobileNetV3 model requires a large amount of training data sets, otherwise it may lead to overfitting

[30] . Therefore, based on the MobileNetV3 network, the present invention proposes a new and improved lightweight DAS signal classification network MobileNetV3-DAS. By adopting the ECA (Efficient Channel Attention)

[31] attention mechanism to improve MobileNetV3, the channel features of the feature map are enhanced, and the dependency between channels is extracted, thereby improving the accuracy of the network classification signal and reducing the number of model parameters and inference delay.

[0149] The MobileNetV3-DAS network structure proposed in the present invention is as follows Figure 6 As shown in Figure 1, the input OTDR phase signal is converted into a 224×224×3 RGB image x by GAF. The image is input into the network and firstly extracted by ordinary two-dimensional convolution C2H 3x3 / 2 feature to obtain X 1 (112, 112, 16), then X 1 Input into the improved deep separable convolution BottleNeck based on ECA attention of the present invention to further extract features Figure X 2 (7, 7, 96), and finally the feature Figure X 2 Perform a maximum pooling (MaxPooling), three two-dimensional convolutions and Softmax calculations to obtain the final features. Figure X out(1, 1, k). C2H 3x3 / 2 indicates that the module consists of two-dimensional convolution and h-swish activation function, and the convolution kernel size is 3x3 and the stride is 2; C2H 1x1 / 1 indicates that the module consists of two-dimensional convolution and h-swish activation function, and the convolution kernel size is 1x1 and the stride is 1, and the h-swish activation function is shown in formula (29); Dwise NL and Pwise NL represent 3x3 channel-by-channel convolution and point-by-point convolution, respectively, and use nonlinear activation h-swish; C2 NBN 1x1 / 1 indicates that a two-dimensional convolution with a convolution kernel of 1x1 and a stride of 1 is used to calculate the feature map, and no normalization is used for the convolution calculation result. The network's depthwise separable convolution parameter configuration is shown in Table 1, where Input represents the input BNeck feature map size; Operator represents the convolution operation on the feature map, and the size of the convolution kernel is given for different feature maps; exp size represents the number of point-by-point convolution kernels; out represents the number of depthwise separable convolution output feature maps; the ECA column indicates whether the layer uses the ECA attention mechanism; NL represents the activation function used by the layer, where RE is the ReLU activation function and HS is the h-swish activation function; S represents the convolution operation step size. When the scale of the input feature map and the output feature map are different in the network, the convolution step size is 12, otherwise the step size is 1.

[0150] ReLU(x)=min(max(0,x),6) (28)

[0151]

[0152] Table 1 Depthwise separable convolution parameter configuration

[0153]

[0154]

[0155] like Figure 7 (a) shows the original depth-separable convolution structure of MobileNetV3, in which the SE attention mechanism is used. Figure 7 (b) shows the improved depth-separable convolution based on ECA of the present invention. This structure is used for the input feature map D in First, a 1×1 normal volume is used for feature extraction, and then the result is input into a 3×3 channel-by-channel convolution to obtain D' in , the weight vector Φ with the same length as the number of channels is calculated through the ECA attention mechanism, and the weight vector Φ is combined with the feature map D' in Fusion, so that the neural network focuses on the feature map that best represents the signal type Finally, the feature map Through point-by-point convolution calculation, the output feature map D' of a single depth-separable convolution is obtained out Among them, the ECA attention mechanism performs average pooling, point-by-point convolution and Sigmoid activation on the input feature map in turn to obtain the same number of weight vectors as the feature maps. Compared with the original depthwise separable convolution, EAC reduces two full connection calculations and one activation function calculation, thereby reducing the amount of depthwise separable convolution calculations and the number of parameters.

[0156] An embodiment of the present invention also provides a system for rapid classification of unbalanced small sample events applicable to a Φ-OTDR system, comprising a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a method for rapid classification of unbalanced small sample events applicable to a Φ-OTDR system as described in the above technical solution.

[0157] The beneficial effects achieved by the present invention are described below by a specific example:

[0158] The schematic diagram of the experimental scene of the present invention is as follows Fig. 9 , deploy a 2km long single-core optical cable on the fence around an important facility as shown in the figure, and fix it with a rolling tape every 1 meter. The deployment construction diagram is as follows Fig.10 As shown in (b) in the figure. Fig.10 The device shown in (a) is the Ada-5042s device developed by a certain company, which collects phase change signals of six types of intrusion events: knocking on optical cables, knocking on fences, climbing over fences, shaking optical cables, pedestrians passing by, and vehicles passing by.

[0159] In the experiment, the pulse frequency of DAS was set to 1000Hz, the laser pulse width was 32us, and the spatial resolution was 1m. The acquired fiber vibration amplitude signal was demodulated into phase data using the special IQ demodulation method proposed by our research group in the data preprocessing stage. The phase data was obtained as follows: Fig.11 The duration, sample quantity and signal category description of the collected events are shown in Table 2.

[0160] Table 2 Experimental data description

[0161]

[0162] The CPU of the computer used in the experiment of the present invention adopts Intel Core I9 13900H, the memory is 16GB, and the operating system used is Windwos11. First, the Python3.7 language is used to implement the data enhancement algorithm in the data enhancement task; then, the Pytorch 1.13.1 deep learning framework is used to implement the improved MobileNetV3-DAS network of the present invention in the signal classification task; finally, the training weights of MobileNetV3-small are used as the initial weights of the model in the model training stage, and the training set and the test set are divided into 9:1, and the BatchSize is set to 64, and the learning rate is 0.002 to complete the model training.

[0163] The present invention uses the number of samples of different categories, the category balance rate and the cosine similarity as the measurement index of the data enhancement method in the data enhancement task, wherein the number of samples represents the number of samples in different lists; the category balance rate mainly measures the balance of samples; the cosine similarity mainly measures the similarity of the phase changes of two signals; in the intrusion signal classification task, the accuracy rate, precision rate, recall rate, specificity, F1, model parameters and reasoning time are used to evaluate the model of the present invention. The accuracy rate represents the proportion of all correctly predicted samples; the precision rate represents the proportion of true positives in the samples predicted as positive; the recall rate represents the ratio of the number of correctly predicted true positives to the total positive data; the specificity refers to the accuracy of the classifier when identifying negative samples; the F1 value is an index used to measure the accuracy of the classification model, which takes into account the accuracy and recall of the classification model at the same time; the model parameter represents the size of the generated weight file; the reasoning time represents the time required for the model to classify a single time series signal.

[0164] (1) Data augmentation method experiments and results analysis

[0165] Fig.12 Shown by Fig.10 After the phase matrix of a single intrusion event is enhanced by the SCE method, multiple one-dimensional phase samples are obtained. According to the number of one-dimensional time series signals obtained, the number of samples is expanded to 8-10 times of the original. This is because the SCE method utilizes the spatial correlation of distributed optical fibers, that is, the phase changes caused by event disturbances show certain similarities at adjacent positions in space, and the phase changes at multiple adjacent positions can all represent this type of event.

[0166] Depend on Fig.14 (a) SCE method cosine similarity and Figure 6It can be seen that the cosine similarity of points 2-6 in the SCE method enhanced samples is above 70%. After analysis, it is found that the data in this range is close to the incident location point 3, so the phase disturbance caused by the phase is very similar. The cosine similarity of points 0-1 and 7-9 is between -20% and 69. This is because the SCE method does not judge the correlation between the disturbance events while enhancing the samples, and marks the surrounding environmental noise as an event. Therefore, the SCE enhancement method will reduce the quality of the data set, thereby affecting the model's classification accuracy for intrusion events; from Table 3 and Fig.16 It can be seen that the sample imbalance rate of the SCE method obtained by the imbalance rate is 4.83, but the data set enhanced by the SCE method is still an unbalanced data set.

[0167] Fig.13 The figure shows the event phase sequence obtained by using the SCDE method proposed in the present invention. In the experiment, the similarity threshold is set to 0.7, that is, the signal with a similarity less than 0.7 to the main characteristic signal is removed. Fig.13 It can be seen that the environmental noise marked as events is significantly reduced in the phase sequence obtained by using the SCDE method, thereby improving the quality of the sample data.

[0168] Depend on Fig.14 From (b) in Fig. 3, we can see that the cosine similarity of the signals is above 70%. This is because the SCDE method first locates the characteristic signal position according to the maximum energy value from the single event matrix, and then removes the environmental noise signal in the event sample through the similarity threshold between other signals and the main characteristic signal. Fig.13 It can be seen that the imbalance rate of the samples obtained by the SCDE method is 4.83. This is because although the SCDE method removes the environmental noise in the samples, the samples obtained by the SCDE method are still an unbalanced sample set.

[0169] Fig.15 The experimental results of the SCNFE method proposed in the present invention are shown in FIG. 3. From the figure, it can be seen that the SCNFE method proposed in the present invention generates multiple new signals through noise base fusion, thereby expanding the number of small samples. This is because the method first removes the environmental noise in the current signal from the original signal through the wavelet denoising method, and then generates a new signal of the same type by adding environmental noise from other locations. From Table 3 and Fig.16 Quantitative analysis of the sample data shows that the imbalance rate of the samples obtained by the SCNFE method is 1.09. This is because new samples are generated through the SCNFE method and environmental noise, thereby turning the unbalanced data set into a balanced data set.

[0170] Table 3 Sample size of different data augmentation methods

[0171]

[0172] (2) Experimental verification of event classification method

[0173] In order to verify the influence of different data enhancement methods on the classification results of intrusion signals, the MobileNetV3 network improved by the ECA mechanism is used to classify six types of intrusion signals, and the classification results are shown in Table 4. Table 4 shows that with the increase of the number of samples, the overall classification accuracy of the model shows an upward trend, and the classification accuracy of the SCNFE method and SCDE method proposed in the present invention is improved by 18.29% and 12.20% respectively compared with the SCE method, which shows that the number of data samples and unbalanced samples have a great influence on the average classification accuracy of intrusion signals. For the SCE method, the precision, recall rate and F1 index of the sixth type of event are lower than 0.8, and the other indicators are all above 0.8. This is because the part of the environmental noise contained in the sample in the enhanced data of the SCE method is marked as an event, thereby reducing the ability of the model to extract features. For the SCNFE and SCDE methods, their classification indicators are all above 0.8, and the classification indicators of the SCDE method for the 4th and 6th types of events are both 1.0, while the other classification indicators of the SCNFE method are all above 0.95. In addition, by Fig.17 It can be seen that as the number of data set samples continues to increase, the classification accuracy is also increasing. Fig.18 The confusion matrix shows that the SCNFE method proposed in this invention has small errors only in the second, third and fifth categories of events, and the classification accuracy of other categories is 1.00; the SCDE method has classification errors in all categories except the first and sixth categories of events, with a classification accuracy of 1.00; the SCE method has classification accuracy of less than or equal to 0.9 for the six types of signals. In summary, in signal classification, unbalanced small sample signals have a great influence on the classification results.

[0174] Table 4 Experimental results of different categories of MobileNetV3-DAS

[0175]

[0176]

[0177] Table 5 shows the classification results of different classification networks in the SCNFE method, SCDE method and SCE method. The results show that the classification accuracy of all methods increases with the increase of sample number. Compared with the non-lightweight network VGG, the classification accuracy of the MobileNetV3-DAS network proposed in the present invention is above 0.8 in all enhanced samples of the SCE method, among which the various indicators of the VGG16 network are above 0.92, but the model file generated by the VGG16 network is 512MB, which is 128 times that of the MobileNetV3-DAS network, and the inference time is 155.79ms; compared with the non-lightweight networks AlexNet and ResNet34, although the classification index of the ResNet34 method on the data set generated by the SCNFE method is higher, the model parameters are 10-20 times that of the MobileNetV3-DAS network, so it is difficult to run the traditional neural network classification model on the resource-constrained DAS device. Compared with the lightweight networks MoibleNetV3 and MobileNetV2, MobileNetV3-DAS reduces the model parameters by 29.80% and 52.29% respectively. This is because the ECA module used in the MobileNetV3-DAS network proposed in the present invention has two fewer fully connected layers than the SE module used in the original MobileNetV3, thereby reducing the number of model parameters and thus reducing the inference time. Compared with ResNet34 and MobileNetV2, the MobileNetV3-Das network proposed in the present invention is smaller than other models in terms of model weight and inference time. In summary, the improved MobileNetV3-DAS network of the present invention is superior to other currently popular models in terms of classification indicators, model weight and inference time, so MobileNetV3-Das is more suitable for real-time classification of signals. Fig.19 It can be seen that the MobileNet-DAS method of the present invention shows good classification effects on the data sets enhanced by the SCNFE, SCDE and SCE methods.

[0178] Table 5 Experimental results of different classification methods

[0179]

[0180] In order to verify the influence of different attention mechanisms on network classification performance, the present invention compares and analyzes the spatial channel attention mechanism (CBAM), coordinate attention mechanism (CA), compressed excitation attention mechanism (SE) of MobileNetV3 and efficient channel attention mechanism (ECA). The results in Table 6 show that the MobileNetV3 network improved by ECA reduces the model weight by 29.80%, 13.84% and 11.65% respectively compared with SE, CBAM and CA; on the intrusion signal samples obtained by the SCNFE method proposed in the present invention, the reasoning time is reduced by 10.26%, 24.69% and 35.76% respectively. This is because the ECA module removes the fully connected layer in the original SE module and directly learns through a 1D convolution on the feature map after global average pooling. Compared with the two-dimensional convolution used by other attention mechanisms, it reduces the computational complexity while maintaining performance, thereby reducing the model reasoning time and the number of model parameters.

[0181] Table 6 Experimental results of improving MobileNetV3 with different attention mechanisms

[0182]

[0183] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A fast classification method for unbalanced small sample events suitable for φ-OTDR system, characterized in that: The steps include: Step 1, obtaining an intrusion signal sensed by the φ-OTDR and constructing an unbalanced small sample set, wherein the intrusion signal includes multiple disturbance events and noise signals; Step 2: Use the spatial correlation noise fusion data enhancement method to expand the unbalanced small sample set into a sample set that meets the requirements; specifically, the following steps are included: Step 2.1, spatial correlation data enhancement: First, prove that the two-dimensional space-time signal has spatial correlation at adjacent distance positions; then perform spatial correlation decomposition on the multi-dimensional disturbance event signal matrix to obtain multiple one-dimensional vibration signals; Step 2.2, using the spatial correlation threshold noise reduction method to process the vibration signal with background noise to remove the background noise signal; Step 2.3, noise fusion data enhancement; The result of step 2.2 is decomposed by wavelet to obtain coefficients of different frequencies. The frequency coefficients after wavelet decomposition are subjected to soft threshold denoising to filter out the high-frequency noise part. The denoised frequency coefficients are subjected to signal reconstruction to obtain a one-dimensional time series signal without high-frequency noise. The noise matrix at different positions with the same time scale as the event is intercepted, and the noise matrix is ​​subjected to spatial separation processing, and the noise is fused with the denoised signal. Step 3, converting the intrusion signal in the sample set into a two-dimensional image through a two-dimensional image construction method based on the Gram angle field of adaptive time scale transformation and performing interpolation operations to obtain an image suitable for network input; Step 4: Build the MobileNetV3-DAS network and use the trained network model to quickly classify disturbance events.

2. A method for rapid classification of unbalanced small sample events applicable to a Φ-OTDR system as claimed in claim 1, characterized in that: In step 2.1, assume that the phase shift caused by the disturbance event t is represented by the matrix E t According to formula (1), the element x in the matrix ij represents the phase change of point j on the distributed optical fiber at time i, where i∈[1,m], j∈[1,n], m represents the number of one-dimensional signals, and n represents the length of the one-dimensional signal sequence; for event E t Separate in space to obtain multiple time series phase signals e k (k∈[1,n]), and each timing signal e k Both record the time domain characteristics of the disturbance event t; e k =[x k1 ,x k2 ,...,x kn ] T ,(k∈[1,m]) (2)。 3. A method for rapid classification of unbalanced small sample events applicable to a Φ-OTDR system as claimed in claim 2, characterized in that: In step 2.2, define E t The energy matrix is ​​given by R t It is expressed as shown in formula (3), where r i (i∈[1,n]) represents the intrusion disturbance event e i The energy of (i∈[1,n]); R t ＝[r1，r2，…，r n ] (3) For the event timing signal e k (E t ∈e k , k∈[1,n]), firstly, it is converted from the time domain sequence to the frequency domain sequence f by formula (4) k , and the event signal e is obtained by formula (5) k Energy k , where n represents the length of the signal sequence, e k (i) represents event e k Phase change at time i; Then, from R t Find the time series signal e with the largest energy t (t∈[1,n]), and the timing signal e is calculated by formula (6) k With signal e k The correlation degree of (k≠i) is s ki , and E t The correlation S of other signals in t As shown in formula (7); finally, the denoised and enhanced event samples are obtained; S t =[s k1 ,s k2 ,…,s k(n-1) ],(k∈[1,n-1]) (7)。 4. A method for rapid classification of unbalanced small sample events applicable to a Φ-OTDR system as claimed in claim 3, characterized in that: In step 2.3, let e' k For E' t An element of N t is related to event E' t Noise signals occurring at different locations at the same time and with E' t have the same time scale, as shown in formula (9), n k (k∈[1,l]) represents the same as e' k The noise signal of the same time scale is shown in formula (10); and' k =[and k1 ,and k2 ,...,and kη ] r ,(k∈[1,n],η≤m) (8) N t =[n1,n2,...,n l ] (9) <h2 style=";text-align:left;direction:ltr">n<h2 style=";text-align:left;direction:ltr"> η <h2 style=";text-align:left;direction:ltr"> (b1, b2,..., b)<h2 style=";text-align:left;direction:ltr"> m <h2 style=";text-align:left;direction:ltr"> ]<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> (η∈[1, l]) (10) Wavelet soft threshold denoising method is used to remove e' k The original environmental noise in the original event signal e k Decompose; then, the kth coefficient of the jth layer after quantization is obtained by formula (14): Where λ is the threshold value, calculated by formula (12), where idwt() is the wavelet reconstruction function; With' k =idwt(in j，k ), (14) Finally, the denoised signal z' k With the noise signal n η By superimposing formula (14), we get the new signal α k : α k =n η +z' k (15)。 5. The unbalanced small sample event fast classification method applicable to the Φ-OTDR system according to claim 1, characterized in that: The specific processing process of step 3 is as follows: First, for the enhanced phase change sequence α k , obtained by step 2, the phase sequence is normalized to the interval [-1, 1] using formula (16) to obtain the vector Among them, n k (i) represents a one-dimensional time series signal n k The i-th element of ; Then, by The radius and angle corresponding to each phase element are calculated by formula (17), where the length() function is used to return a vector Length; Among them, L represents the length of the time series; finally, the matrix G is obtained by using the sum angle relationship GSF or the sum difference relationship AsF and G ADF Each element in the matrix, such as formula (18) and formula (19), is put into the range of 0 to 255 by formula (20), where O(i, j) represents the pixel value at i, j in the image, int(·) is the rounding function, and G(i, j) represents the Gram matrix G ASF or G ADF The element value of the i-th row and i-th column; O(i,j)=int(127.5×(G(i,j)+1)) (20) Let the image obtained by Gram's angular field method be image O k×k , and the image scale that satisfies the network input is Q l×l , the scaling ratio ψ is calculated by formula (21), where k is the width and height of the original GAF image. At this time, the position u0 and υ0 of each pixel point on the coordinate axis in the target image Q are expressed by formula (22), then the image Q is enlarged by α times in the x-axis direction and β times in the y-axis direction, where α and , can be obtained by the scale ratio of the original image to the target image; Q(u0, υ0)=O[a(x, y), b(x, y)] (22) Wherein, a(x, y) and b(x, y) represent the calculation of the coordinates (x, y) using functions a and b, and the calculation method is shown in formula (23); if the scaled u0 and υ0 are not on the coordinate axis, the bilinear interpolation method is used to obtain the result of Q(u0, υ0), and the calculation method is shown in formula (24); Among them, u0 and υ0 represent the results obtained by processing (x, y) by functions a and b, that is, the positions of the pixel points on the coordinate axes of the image after scale reconstruction.

6. A method for rapid classification of unbalanced small sample events applicable to a Φ-OTDR system as claimed in claim 1, characterized in that: The processing of the MobileNetV3-DAS network is as follows: Set the image to X. First, extract the features through the ordinary two-dimensional convolution C2H 3x3 / 2 to get X1. Then input X1 into the improved deep separable convolution BottleNeck based on ECA attention to further extract the feature map X2. Finally, perform a maximum pooling MaxPooling, three two-dimensional convolutions and Softmax calculation on the feature map X2 to get the final feature map X. out (1, 1, k); where C2H 3x3 / 2 indicates that the two-dimensional convolution is composed of a two-dimensional convolution and an h-swish activation function, and the convolution kernel size is 3x3, the stride is 2, and the h-swish activation function is shown in formula (29); ReLU(x)=min(max(0,x),6) (28) Among them, x is the feature map input by the h-swish activation function.

7. A method for rapid classification of unbalanced small sample events applicable to a Φ-OTDR system as claimed in claim 6, characterized in that: The three two-dimensional convolutions include one two-dimensional convolution C2H 1x1 / 1 and two two-dimensional convolutions C2 NBN 1x1 / 1, where C2H 1x1 / 1 means that the module consists of two-dimensional convolution and h-swish activation function, and the convolution kernel size is 1x1 and the stride is 1; C2 NBN1x1 / 1 means that the two-dimensional convolution with a convolution kernel of 1x1 and a stride of 1 is used here to calculate the feature map, and the convolution calculation results are not normalized.

8. A method for fast classification of unbalanced small sample events applicable to a Φ-OTDR system as claimed in claim 6, characterized in that: Based on the improved depth separable convolution of ECA, this structure performs in First, a 1×1 normal volume is used for feature extraction, and then the result is input into a 3×3 channel-by-channel convolution to obtain D' in , the weight vector Φ with the same length as the number of channels is calculated through the ECA attention mechanism, and the weight vector Φ is combined with the feature map D' in Fusion, so that the neural network focuses on the feature map that best represents the signal type Finally, the feature map Through point-by-point convolution calculation, the output feature map D' of a single depth-separable convolution is obtained out ; Among them, the ECA attention mechanism performs average pooling, point-by-point convolution and Sigmoid activation on the input feature map in sequence to obtain a weight vector with the same number as the feature maps.

9. A fast classification system for unbalanced small sample events suitable for Φ-OTDR system, characterized by: The invention comprises a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a method for fast classification of unbalanced small sample events applicable to a Φ-OTDR system as claimed in any one of claims 1 to 8.

Images