UwDAS pipeline intrusion signal recognition method based on ICEEMDAN-FE-AIT and F-ELM

Abstract

The uwDAS pipeline intrusion signal recognition method based on ICEEMDAN-FE-AIT and F-ELM comprises the following steps: using uwDAS to collect pipeline intrusion signals x(t), pre-processing x(t) to obtain pre-processed signals x'(t); performing ICEEMDAN decomposition on x'(t) to obtain multiple IMF modal components; calculating the fuzzy entropy value FE of each order IMF, and determining the IMF in the FE=[FEl, FEh] threshold range as noise dominant ρIMF; performing AIT processing on the noise dominant ρIMF, combining the noise dominant and the remaining useful signal dominant IMF after AIT processing, and reconstructing to obtain a denoising signal x''(t); extracting the time domain and frequency domain features of the denoising signal x''(t) and making a data set, and assigning labels according to different types of pipeline intrusion signals; dividing the data set into a training set and a test set, and inputting the F-ELM model for recognition. DNSN Compared with the traditional denoising method, the method has the best R DNSN The effect of retaining useful signals to the maximum and suppressing noise is the most obvious. The recognition accuracy of the pipeline intrusion signal is obviously improved.

Description

Technical Field

[0001] This invention relates to the field of pipeline safety monitoring, and specifically to a method for identifying uwDAS pipeline intrusion signals based on ICEEMDAN-FE-AIT and F-ELM. Background Technology

[0002] Pipeline transportation is a crucial method for long-distance transport of oil and natural gas. With the rapid development of my country's oil and gas pipeline network, Distributed Optical Fiber Acoustic Sensing (DAS) systems have been widely applied in pipeline safety monitoring. DAS systems detect changes in light intensity caused by vibration through sensing optical fibers, converting these changes into electrical signals which are then processed and analyzed by a host computer. This effectively monitors pipeline vibration and provides early warnings for safety.

[0003] An ultra-weak fiber Bragg grating (uwFBG) based on Fizeau interferometry and an enhanced phase-sensitive optical time domain reflector (PDR) are used. UWDAS, or Active-Launched Detector-Reflector (OLDR) technology, offers higher sensitivity and signal-to-noise ratio compared to DAS systems based on single Rayleigh scattering. However, in practical applications, UWDAS systems are susceptible to environmental influences, potentially leading to a high false alarm rate in subsequent event recognition. Therefore, denoising the signal to improve signal quality is crucial for enhancing the accuracy of subsequent event recognition.

[0004] Current denoising methods applied to DAS systems include Empirical Mode Decomposition (EMD), Complementary Ensemble Empirical Mode Decomposition (CEEMD), Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN), and Variational Mode Decomposition (VMD). EMD and CEEMD are suitable for nonlinear and non-stationary signal analysis, but they suffer from problems such as a high dependence on finding extrema, a lack of mathematical theoretical foundation, and mode aliasing. The effectiveness of VMD is significantly affected by the number of decomposition layers K and the penalty factor alpha.

[0005] Current recognition methods applied to DAS systems include Artificial Neural Networks (ANNs), Convolutional Neural Networks (CNNs), Support Vector Machines (SVMs), and Extreme Learning Machines (ELMs). Neural network-based methods suffer from long recognition times and high time complexity. During recognition, feature selection is crucial to build a more efficient classifier and reduce computational load. Reference [H. Jia, S. Lou, S. Liang, and X. Sheng, "Event Identification by F-ELM Model for..."] [OTDR Fiber-Optic Distributed Disturbance Sensor, IEEE Sens.J., vol.20, no.3, pp.1297-1305, Feb.2020] extracted time-domain and frequency-domain features, selected features using Fisher scores, and combined them with ELM for event recognition, improving recognition accuracy. This paper did not perform signal denoising and only compared three commonly used activation functions when using ELM.

[0006] Defects and shortcomings of existing technology:

[0007] (1): The ICEEMDNA algorithm can decompose any signal into several intrinsic mode components (IMFs) and a residual term (Res). After the pipeline intrusion signal collected by uwDAS is decomposed by ICEEMDNA, a series of IMFs are obtained, with noise mainly distributed in the higher-order IMFs. Current IMF screening strategies are difficult to accurately screen effective IMFs. The correlation coefficient screening method is suitable for signals with low signal-to-noise ratio, but it cannot correctly screen noise-dominated IMFs and useful signal-dominated IMFs; the permutation entropy method depends on the selection of parameters, and it is based on statistical characteristics, which may discard some important information in the IMF. Therefore, a suitable screening method is very important.

[0008] (2): If the noise-dominant IMF is directly removed after decomposition, the signal will be severely distorted; if the noise-dominant IMF is not properly processed, the signal after denoising may still contain obvious noise. The key to the ICEEMDAN algorithm is how to determine the noise-dominant IMF and the useful signal-dominant IMF, and to process the noise-dominant IMF reasonably.

[0009] (3) When extracting features from pipeline intrusion signals, some unimportant features may reduce the accuracy of identification. In addition, most neural network-based methods have long identification times and high time complexity, and time costs need to be considered in practical applications. Therefore, feature selection and choosing a suitable identification method are necessary. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention proposes a method for identifying uwDAS pipeline intrusion signals based on ICEEMDAN-FE-AIT and F-ELM. Compared to traditional denoising methods, this invention's method achieves optimal R-value for pipeline intrusion signals acquired by uwDAS. DNSN This method achieves the most significant noise suppression while preserving useful signals to the greatest extent possible. The accuracy of pipeline intrusion signal identification has been significantly improved.

[0011] The technical solution adopted in this invention is as follows:

[0012] The uwDAS pipeline intrusion signal identification method based on ICEEMDAN-FE-AIT and F-ELM includes the following steps:

[0013] Step 1: Use uwDAS to acquire the pipeline intrusion signal x(t), preprocess the pipeline intrusion signal x(t) to obtain the preprocessed signal x'(t);

[0014] Step 2: Perform ICEEMDAN decomposition on the preprocessed signal x'(t) to obtain multiple IMF mode components;

[0015] Step 3: Calculate the fuzzy entropy value FE for each order of IMF, and determine the IMF in the threshold range of FE = [FEl, FEh] as noise-dominated ρIMF;

[0016] Step 4: Perform AIT processing on the noise-dominant ρIMF, and then process the AIT-processed noise-dominant ρIMF. Combined with the remaining useful signal that dominates the IMF, the signal is reconstructed to obtain the denoised signal x (t);

[0017] Step 5: Extract the time and frequency domain features of the denoised signal x(t) and create a dataset, assigning labels according to different types of pipeline intrusion signals;

[0018] Step 6: Divide the dataset into training and test sets, and input them into the F-ELM model for recognition.

[0019] In step 1, uwDAS is used to collect the pipeline intrusion signal x(t), and x(t) is preprocessed to obtain x'(t), as follows:

[0020] First, x(t) is divided into frames, and then normalized. The normalization formula is as follows:

[0021]

[0022] In the above formula, x'(t) is the preprocessed signal, p is the number of data points; x(1), x(2)...x(p) represent the 1st, 2nd, ..., pth data points of each frame after framing.

[0023] Step 2 includes the following steps:

[0024] Step 2.1: Calculate the first-order residual r1 and the first-order modal component IMF1:

[0025] r1 = <M((x'(t)) i )>;

[0026] IMF1 = x'(t) - r1;

[0027] In the above formula, <·> represents the average of multiple signals; M(·) represents the local mean of the signal; (x'(t)) i =x'(t)+β0E1(ω) i ), i = 1, ..., n1, where n1 is the number of Gaussian white noise particles added during signal construction; ω i This represents the i-th Gaussian white noise with a mean of 0 and a variance of 1; β0 = ε0std(x'(t)) / std(E1(ω)) i std(·) represents the calculated standard deviation, ε0 is the expected signal-to-noise ratio coefficient added at the beginning of the decomposition, and E k (·) represents the k-th modal component after EMD decomposition; E1(ω i ) represents the first-order white noise mode component after EMD decomposition.

[0028] Step 2.2: Calculate the second-order residual r2 and the second-order modal component IMF2:

[0029] r2 = <M(r1+β1E2(ω i ))>;

[0030] IMF2 = r1 - r2;

[0031] In the above formula, β1=ε0std(r1), std(r1) represents the standard deviation of the first-order residual; E2(ω i The second-order white noise mode component after EMD decomposition.

[0032] Step 2.3: For k = 3, ..., K, where K is the number of modal components in the final decomposition, calculate the k-th order residual r. k and the k-th modal component IMFk :

[0033] r k = <M(r k-1 +β k-1 E k (ω i ))>;

[0034] IMF k =r k-1 -r k ;

[0035] In the above formula, β k-1 =ε0std(r k-1 std(r) k-1 E represents the standard deviation of the (k-1)th residual. k (ω i ) represents the k-th order white noise mode component after EMD decomposition, r k-1 This represents the (k-1)th order residual.

[0036] Step 2.4: Repeat step 2.3 until all residuals and IMF modal components are obtained, then stop the decomposition.

[0037] In step 3, the IMF is calculated. k The fuzzy entropy includes the following steps:

[0038] Step 3.1: For a univariate time series of length N: u = {u1, u2, ..., u...} j ,…,u N},u1,u2,…,u j ,…,u N Let each of the following represent the 1st, 2nd, ..., jth, ... Nth values ​​of sequence u, and construct a set of m-dimensional vectors:

[0039]

[0040] In the above formula, u represents IMF k Λ=1,2,…,N-(m-1)d, where m and d represent the embedding dimension and time delay, respectively; Represents the constructed m-dimensional vector; {u Λ ,u Λ+d ,…,u Λ+(m-1)d} represents the m consecutive values ​​of u starting from the Λth point.

[0041] Step 3.2: Define two m-dimensional vectors and Distance between

[0042]

[0043] In the above formula, Λ=1,2,…,N-(m-1)d, λ=1,2,…,N-(m-1)d, Λ≠λ; Glb represents the global average value of u, and Glb represents the global features of the embedding vector. This represents another m-dimensional vector; ChebDist[] represents the Chebyshev distance.

[0044] Step 3.3: Calculate the two vectors using the exponential function. and Similarity between them:

[0045]

[0046] In the above formula, n e θ(Δ) represents the gradient and width of the boundary of the exponential function, respectively. Λ,λ (r) represents two vectors and Similarity between them; △ Λ,λ Represent two vectors and The distance between them;

[0047] Step 3.4: Define the function ψ m,d (n e ,r):

[0048]

[0049] In the above formula, θ(△) Λ,λ (r) represents two vectors and The similarity between them.

[0050] Step 3.5: Fuzzy entropy is defined as:

[0051]

[0052] In the above formula, FuzEn(u,m,r,n) e ,d) represents fuzzy entropy, ψ m+1,d (n e ,r) represents the function obtained by repeating steps 3.1 to 3.4 when the embedding dimension is m+1.

[0053] Step 3.6: Calculate the fuzzy entropy FE for all modal components according to steps 3.1 to 3.5. IMFs within the threshold range of FE = [FEl, FEh] are identified as noise-dominant ρIMFs; FEl represents the lower limit of the threshold range, and FEh represents the upper limit of the threshold range.

[0054] In step 4, AIT processes the data as follows:

[0055]

[0056] In the above formula, This represents the i”th ρIMF after denoising; Represents the i-th ρIMF interval j'. Samples within; Represents the j'th interval Extreme points within; T i” This represents the denoising threshold of the i”th ρIMF, and its value depends on the energy of each intrinsic mode, the number of samples, and the degree of decay of the information entropy. α is an adjustable parameter; Represents the j'th interval of the i” ρIMF; λ' represents the j'th extreme point of the i”-th ρIMF component; i” represents the i”-th ρIMF; j' represents the j'-th interval; λ' represents the denoising threshold of the i”-th ρIMF and the j'-th interval. The ratio of the absolute values ​​of the extreme points within the range

[0057] Perform AIT processing on all noise-dominant ρIMFs, and then process the resulting... Combined with the remaining useful signal that dominates the IMF, the reconstructed signal x(t) is obtained:

[0058]

[0059] In the above formula, M1 and M2 are the noise-dominant values ​​determined based on the fuzzy entropy values. The order range; Res is the r at which the decomposition terminates. k K represents the number of modal components in the final decomposition, and k represents the k-th order modal component.

[0060] In step 5, the time-domain and frequency-domain features of the denoised signal x(t) are extracted and a dataset is created, and labels are assigned according to different types of pipeline intrusion signals.

[0061] In step 6, the dataset is divided into a training set and a test set, and then input into the F-ELM model for recognition.

[0062] The F-ELM model consists of two main parts: calculating the Fisher score for each feature in the feature dataset, and then selecting features with high Fisher scores to be fed into the ELM classifier for recognition. ELM is the commonly used Extreme Learning Machine.

[0063] Step 6 includes the following steps:

[0064] Step 6.1: Calculate the Fisher score for each feature in the dataset. The Fisher score is an evaluation metric for feature selection. The specific formula for calculating the Fisher score is as follows:

[0065]

[0066] In the above formula, c is the number of sample categories. It is the first The number of samples in each class For the first The first feature The mean of the class samples, For the overall sample at the 1st The mean of each feature No. In the class of samples, the first The variance of each feature.

[0067] Step 6.2: Sort the features according to the Fisher score and select the features with high Fisher scores.

[0068] Step 6.3: Train the ELM model using the selected features based on the Fisher score, including:

[0069] Step 6.3.1: The expression for the input matrix IN of the ELM classifier is:

[0070]

[0071] In the above formula, each column represents a training sample, and each row represents a feature; there are a total of Given *n* training samples, each containing *a* features, then the input matrix *IN* has *a* rows. List.

[0072] The actual output matrix Y is expressed as follows:

[0073]

[0074] In the above formula, the actual output matrix Y has c rows. Each column represents the output of a training sample. The output of each training sample contains c output values, and each output value corresponds to a different intrusion signal category.

[0075] Step 6.3.2: Randomly select the connection weight w between the input layer and the hidden layer and the bias b of the hidden layer neurons;

[0076] The connection weight w between the input layer and the hidden layer is expressed as follows:

[0077]

[0078] In the above formula, w 11 w represents the connection weight between the first neuron in the input layer and the first neuron in the hidden layer. 12 w represents the connection weight between the second neuron in the input layer and the first neuron in the hidden layer. 1a The meaning of w is the connection weight between the a-th neuron in the input layer and the first neuron in the hidden layer. 21 w represents the connection weight between the first neuron in the input layer and the second neuron in the hidden layer. 22 w represents the connection weight between the second neuron in the input layer and the second neuron in the hidden layer. 2a w represents the connection weight between the a-th neuron in the input layer and the second neuron in the hidden layer. l1 w represents the connection weight between the first neuron in the input layer and the l-th neuron in the hidden layer. l2 w represents the connection weight between the second neuron in the input layer and the l-th neuron in the hidden layer. la The connection weight between the a-th neuron in the input layer and the l-th neuron in the hidden layer is represented by l, which represents the number of neurons in the hidden layer, and a represents the number of neurons in the input layer.

[0079] The bias b of the hidden layer neurons is expressed as follows:

[0080]

[0081] In the above formula, b1 represents the bias of the first neuron in the hidden layer; b2 represents the bias of the first neuron in the hidden layer; b l This represents the bias of the l-th neuron in the hidden layer.

[0082] Step 6.3.3: Calculate the hidden layer output matrix H;

[0083] Choose an infinitely differentiable function as the activation function g(x) for the hidden layer neurons, and denote the connection weight between the hidden layer and the output layer as β. The expression is:

[0084]

[0085] In the above formula, β 11 β represents the connection weight between the first neuron in the hidden layer and the first neuron in the output layer. 12 β represents the connection weight between the first neuron in the hidden layer and the second neuron in the output layer. 1c β represents the connection weight between the first neuron in the hidden layer and the c-th neuron in the output layer. 21 β represents the connection weight between the second neuron in the hidden layer and the first neuron in the output layer. 22β represents the connection weight between the second neuron in the hidden layer and the second neuron in the output layer. 2c β represents the connection weight between the second neuron in the hidden layer and the c-th neuron in the output layer. l1 β represents the connection weight between the l-th neuron in the hidden layer and the 1-th neuron in the output layer. l2 β represents the connection weight between the l-th neuron in the hidden layer and the 2nd neuron in the output layer. lc This represents the connection weight between the l-th neuron in the hidden layer and the c-th neuron in the output layer. c represents the number of neurons in the output layer.

[0086] The predicted output matrix T of the ELM classifier is obtained as follows:

[0087]

[0088] In the above formula, t1, t2, t q , These represent the prediction output matrix, i.e., the first... One output.

[0089] in,

[0090]

[0091] In the formula, w σ =[w σ1 ,w σ2 ,…,w σa ], w σ1 ,w σ2 ,…,w σa These represent the connection weights between the 1st, 2nd, ..., ath neurons in the input layer and the σth neuron in the hidden layer, respectively, where σ = 1, 2, ..., l.

[0092] in q =[in 1q ,in 2q ,…,in aq ] T ;in 1q ,in 2q ,…,in aq These represent the values ​​in the 1st, 2nd, ..., ath rows and qth columns of the input matrix IN, where T represents the values ​​of [in] . 1q ,in 2q ,…,in aq ] Request transpose;

[0093] t 1q t represents the output of the q-th hidden layer neuron of the 1st sample. 2q Let t represent the output of the q-th hidden layer neuron of the second sample.cq Let represent the output of the hidden layer neuron corresponding to the c-th sample and the q-th sample.

[0094] β σ1 β represents the connection weight between the σ-th neuron in the hidden layer and the 1-th neuron in the output layer. σ2 β represents the connection weight between the σ-th neuron in the hidden layer and the 2nd neuron in the output layer. σc b represents the connection weight between the σ-th neuron in the hidden layer and the c-th neuron in the output layer. σ Let represent the bias of the σ-th neuron in the hidden layer.

[0095] In the above formula, the predicted output matrix T is represented as: Hβ=T', where T' is the transpose of T.

[0096] The hidden layer output matrix H is calculated as follows:

[0097]

[0098] Among them, w1, w2, ..., w l These represent the 1st, 2nd, ..., 1st rows of the connection weights w between the input layer and the hidden layer, respectively.

[0099] These represent the first and second halves of the input matrix IN, respectively. List;

[0100] g(w1·in1+b1) represents the activation function with input w1·in1+b1, g(w2·in1+b2) represents the activation function with input w2·in1+b1, g(w l ·in1+b l ) indicates that the input is w l ·in1+b l The activation function is defined as follows: g(w1·in2+b1) represents the activation function for input w1·in2+b1, g(w2·in2+b2) represents the activation function for input w2·in2+b2, and g(w l ·in2+b l ) indicates that the input is w l ·in2+b l activation function, Indicates that the input is activation function, Indicates that the input is activation function, Indicates that the input is The activation function.

[0101] Step 6.3.4: Calculate the optimal connection weights between the hidden layer and the output layer.

[0102] The value of β is obtained by solving the least-squares solution of the following expression:

[0103]

[0104] Where: β represents the connection weight between the hidden layer and the output layer, and T' represents the transpose of the predicted output matrix T.

[0105] Its optimal solution is The expression is as follows:

[0106]

[0107] In the above formula, H + The Moore-Penrose generalized inverse of the hidden layer output matrix H is given, and the ELM classifier is now trained.

[0108] Step 6.3.5: Test the test set with the trained ELM model to determine the category of the test set.

[0109] This invention discloses a method for identifying uwDAS pipeline intrusion signals based on ICEEMDAN-FE-AIT and F-ELM, with the following technical advantages:

[0110] 1) This invention uses ICEEMDAN to decompose pipeline intrusion signals acquired by uwDAS and uses fuzzy entropy to determine the noise-dominated IMF and the useful signal-dominated IMF. When calculating the fuzzy entropy, global features are considered and a Gaussian membership function is used. As a faster algorithm, it produces the largest Hedges'g effect for long signals with over 500 sample points. The ICEEMDAN decomposition yields components with lower noise, and its algorithm has a solid theoretical foundation and clearer physical meaning. Compared to other discrimination methods, fuzzy entropy can identify more noise-dominated IMFs.

[0111] 2) This invention performs AIT processing on the noise-dominant IMF, avoiding the signal distortion problem caused by directly removing the noise-dominant IMF, and effectively suppresses the noise in the noise-dominant IMF, significantly improving the noise reduction effect.

[0112] 3) This invention extracts time-domain and frequency-domain features, uses Fisher scores to select more suitable features for recognition, and combines Fisher scores and ELM classifiers based on the advantage of short training time, and selects appropriate activation functions and the number of hidden layer neurons, which effectively improves recognition efficiency and accuracy. Attached Figure Description

[0113] Figure 1 This is a flowchart of the present invention.

[0114] Figure 2This is a schematic diagram of the experimental equipment of the present invention.

[0115] Figure 3(1) shows the VMD-PE denoising result of the pipeline intrusion signal collected by uwDAS of the present invention;

[0116] Figure 3(2) is the time-frequency diagram of the pipeline intrusion signal collected by uwDAS of the present invention after VMD-PE denoising.

[0117] Figure 3(3) shows the noise reduction result of the pipeline intrusion signal ICEEMDAN-PE-AIT collected by uwDAS of the present invention;

[0118] Figure 3(4) is the time-frequency diagram of the pipeline intrusion signal ICEEMDAN-PE-AIT after denoising collected by uwDAS of the present invention.

[0119] Figure 3(5) shows the noise reduction result of the pipeline intrusion signal ICEEMDAN-ET-AIT collected by uwDAS of the present invention;

[0120] Figure 3(6) is the time-frequency diagram of the pipeline intrusion signal ICEEMDAN-ET-AIT collected by uwDAS of the present invention after denoising.

[0121] Figure 3(7) shows the noise reduction result of the pipeline intrusion signal ICEEMDAN-FE-WSTD collected by uwDAS of the present invention;

[0122] Figure 3(8) is the time-frequency diagram of the pipeline intrusion signal ICEEMDAN-FE-WSTD after denoising acquired by uwDAS of the present invention.

[0123] Figure 3(9) shows the noise reduction result of the pipeline intrusion signal ICEEMDAN-FE-AIT collected by uwDAS of the present invention;

[0124] Figure 3(10) is the time-frequency diagram of the pipeline intrusion signal ICEEMDAN-FE-AIT collected by uwDAS of the present invention after denoising.

[0125] Figure 4 The diagram shows a comparison of the denoising of the pipeline intrusion signal acquired by uwDAS in this invention with four other denoising methods.

[0126] Figure 5 The denoising of pipeline intrusion signals acquired by uwDAS in this invention is compared with four other denoising methods. DNSN Comparison chart.

[0127] Figure 6 The Fisher score results calculated for 30 time-domain and frequency-domain features are shown in the figure.

[0128] Figure 7Recognition results for ELM with different numbers of neurons and activation functions. Detailed Implementation

[0129] like Figure 1 As shown, the uwDAS pipeline intrusion identification method based on ICEEMDAN-FE-AIT and F-ELM includes the following steps:

[0130] Step 1: Use uwDAS to collect the pipeline intrusion signal x(t), and preprocess x(t) to obtain x'(t);

[0131] Step 2: Perform ICEEMDAN decomposition on x'(t) to obtain multiple IMF mode components;

[0132] Step 3: Calculate the fuzzy entropy value FE for each order of IMF, and determine the IMF in the threshold range of FE = [FEl, FEh] as noise-dominated ρIMF;

[0133] Step 4: Perform AIT processing on the noise-dominant ρIMF, and then process the AIT-processed noise-dominant ρIMF. The denoised signal x(t) is obtained by combining the remaining useful signal with the dominant IMF and reconstructing it.

[0134] Step 5: Extract the time and frequency domain features of signal x(t) and create a dataset, assigning labels according to different types of pipeline intrusion signals;

[0135] Step 6: Divide the dataset into training and test sets, and input them into the F-ELM model for recognition.

[0136] In step 1, the pipeline intrusion signal x(t) is acquired using uwDAS, and x(t) is preprocessed to obtain x'(t). The specific steps are as follows:

[0137] First, x(t) is divided into frames, and then normalized. The normalization formula is as follows:

[0138]

[0139] In the above formula, x'(t) is the preprocessed signal, and p = 20000.

[0140] In step 2, x'(t) is decomposed using ICEEMDAN to obtain multiple IMF mode components. The specific steps of ICEEMDAN are as follows:

[0141] 2.1: Calculate the first-order residual r1 and the first-order IMF1:

[0142] r1 = <M((x'(t)) i )>;

[0143] IMF1 = x'(t) - r1;

[0144] In the above formula, <·> represents the average of multiple signals, and M(·) represents the local mean of the signal; (x'(t)) i =x'(t)+β0E1(ω) i ), i = 1, ..., n1, where n1 is the number of Gaussian white noise particles added during signal construction, ω i Let β0 represent the i-th Gaussian white noise with a mean of 0 and a variance of 1, and β0 = ε0std(x'(t)) / std(E1(ω)). i std(·) represents the calculated standard deviation, ε0 is the expected signal-to-noise ratio coefficient added at the beginning of the decomposition, and E k (·) represents the k-th modal component after EMD decomposition.

[0145] 2.2: Calculate the second-order residual r2 and the second-order IMF2:

[0146] r2 = <M(r1+β1E2(ω i ))>;

[0147] IMF2 = r1 - r2;

[0148] In the above formula, β1 = ε0std(r1).

[0149] 2.3: For k = 3, ..., K, calculate the k-th residual r k and the k-th modal component IMF k :

[0150] r k = <M(r k-1 +β k-1 E k (ω i ))>;

[0151] IMF k =r k-1 -r k ;

[0152] In the above formula, K is the number of modal components in the final decomposition, and β k-1 =ε0std(r k-1 ).

[0153] 2.4: Repeat step 2.3 until all residuals and IMF modal components are obtained, then stop the decomposition.

[0154] In step 3, the fuzzy entropy value FE of each order of IMF is calculated, and IMFs within the threshold range of FE = [FEl, FEh] are determined to be noise-dominant ρIMFs. This is used to calculate the IMF. k Taking fuzzy entropy as an example, the specific steps for calculating fuzzy entropy are as follows:

[0155] 3.1: For a univariate time series of length N: u={u1,u2,…,u j ,…,u N Construct a set of m-dimensional vectors:

[0156]

[0157] In the above formula, u represents IMF k Λ=1,2,…,N-(m-1)d, where m and d represent the embedding dimension and time delay, respectively.

[0158] 3.2: Define two m-dimensional vectors and Distance between

[0159]

[0160] In the above formula, Λ≠λ. Glb represents the global average value of u, and Glb represents the global feature of the embedding vector.

[0161] 3.3: Calculating two vectors using an exponential function and Similarity between

[0162]

[0163] In the above formula, n e , where r are the gradient and width of the boundary of the exponential function, respectively.

[0164] 3.4: Define the function ψ m,d (n e ,r):

[0165]

[0166] 3.5: Fuzzy entropy is defined as:

[0167]

[0168] 3.6: For all IMFs, calculate the fuzzy entropy FE according to steps 3.1-3.5. IMFs within the threshold range of FE = [FEl, FEh] are identified as noise-dominant ρIMFs. FEl represents the lower limit of the threshold range, and FEh represents the upper limit of the threshold range.

[0169] In step 4, the noise-dominant ρIMF is subjected to AIT processing, and the AIT-processed noise-dominant... The denoised signal is obtained by combining the remaining useful signal with the dominant IMF and then reconstructing it.

[0170] 4.1: The AIT formula is as follows:

[0171]

[0172] In the above formula, This represents the i”th ρIMF after denoising. Represents the i-th ρIMF interval j'. Samples inside, Represents the j'th interval Extreme points within, T i” Let represent the denoising threshold of the i”-th ρIMF, and its value depends on the energy of each intrinsic mode, the number of samples, and the degree of decay of information entropy. α is set to 5.

[0173] 4.2: Dominate the noise processed by AIT The signal is reconstructed by combining it with the remaining useful signals that dominate the IMF.

[0174]

[0175] In the above formula, M1 and M2 are the noise-dominant values ​​determined based on the fuzzy entropy values. The order range, where Res is the r at which the decomposition terminates. k .

[0176] In step 6, the dataset is divided into training and test sets in a 7:3 ratio and then input into the F-ELM model for recognition. The F-ELM model calculates the Fisher score for each feature in the dataset and selects features with higher Fisher scores for recognition. The specific formula for calculating the Fisher score is as follows:

[0177]

[0178] In the above formula, c is the number of sample categories. It is the first The number of samples in each class For the first The first feature The mean of the class samples, For the overall sample in the first The mean of each feature No. In the class of samples, the first The variance of each feature. For example, uwDAS was used to collect four types of pipeline intrusion signals and one set of non-intrusion signals, so the number of sample categories c=5, and 30 features were extracted in the time and frequency domains.

[0179] The ELM model is trained using selected features based on the Fisher score, and then tested on the test set to determine the category of the test set.

[0180] Example 1:

[0181] In the embodiments of the present invention, each group is tested 100 times and the average value is taken. The following parameters are set when denoising and identifying pipeline intrusions collected by uwDAS.

[0182] Frame division:

[0183] Each frame consists of 20,000 data points, with a frame interval of 10,000.

[0184] ICEEMDAN parameter settings:

[0185] After multiple experiments, the expected signal-to-noise ratio coefficient ε0 of adding Gaussian white noise in ICEEMDAN was set to 0.2, the number of white noise added n1 was set to 100, and the maximum number of components allowed to be decomposed was set to 20.

[0186] FE parameter settings:

[0187] According to the literature [H. Azami, P. Li, SEArnold, J. Escudero and A. Humeau-Heurtier, “Fuzzy Entropy Metrics for the Analysis of Biomedical Signals: Assessment and Comparison,” IEEE Access., vol. 7, pp. 104833-104847, 2019], the Gaussian membership function is used, and the formula is as follows:

[0188]

[0189]

[0190] In the above formula, C r This represents the centroid of the Gaussian membership function. Other FE parameter settings: m = 2, C r =0.1, n e =2, d=1, threshold range setting when distinguishing between noise-dominant and useful signal-dominant components:

[0191] The fuzzy entropy threshold range is set to FE = [0.1, 1]. Modal components within the threshold range are identified as noise-dominant components, while others are identified as useful signal-dominant components.

[0192] Feature extraction and dataset creation:

[0193] References [H. Jia, S. Lou, S. Liang, and X. Sheng, “Event Identification by F-ELM Model for [OTDR Fiber-Optic Distributed Disturbance Sensor, "IEEE Sens.J., vol.20, no.3, pp.1297-1305, Feb.2020], and added 5 time-domain features. The extracted 20 time-domain features and 10 frequency-domain features are shown in Table 1 and Table 2, respectively.

[0194] Table 1 20 Time-Domain Features

[0195]

[0196] Table 2 10 Frequency Domain Features

[0197]

[0198] Pipe intrusion signals collected using uwDAS include chainsaw, mechanical vibration, excavator digging, manual digging, and no intrusion. Labels are assigned according to different types of pipe intrusion signals, with label numbers ranging from 1 to 5. The dataset distribution is shown in Table 3.

[0199] Table 3 Dataset Distribution

[0200]

[0201] ELM parameter settings:

[0202] Figure 7 The average recognition accuracy of five signals was compared under different activation functions and different numbers of hidden layer neurons, and the results were analyzed. Figure 7 After obtaining the results, the softplus activation function was used, and the number of hidden layer neurons was set to 300.

[0203] The following five methods were used to compare the denoising effects of pipe intrusion signals acquired by uwDAS:

[0204] Method 1: Variational Mode Decomposition-Permutation Entropy (VMD-PE), with VMD penalty factor set to 10000, PE embedding dimension set to 5, and time delay set to 1;

[0205] Method 2: Improved complete set empirical pattern decomposition-permutation entropy-adaptive interval threshold ICEEMDAN-PE-AIT, with the embedding dimension of PE set to 5 and the time delay set to 1.

[0206] Method 3: Improved Complete Set Empirical Mode Decomposition - Energy Density and Mean Periodicity - Adaptive Interval Threshold ICEEMDAN-ET-AIT, where ET is the product of energy density and mean periodicity for each order IMF. When satisfied The first IMF at that time is the dividing point between the noise-dominated IMF and the useful signal-dominated IMF;

[0207] Method 4: Improved complete set empirical mode decomposition-fuzzy entropy-wavelet soft thresholding denoising ICEEMDAN-FE-WSTD, where WSTD uses the sym6 wavelet basis and the number of decomposition layers is set to 5;

[0208] Method 5: The ICEEMDAN-FE-AIT noise reduction method proposed in this invention.

[0209] In the above methods, the ICEEMDAN parameter settings are the same, the PE threshold is set to 0.55, and IMFs with PE greater than or equal to 0.55 are judged as noise-dominant IMFs. The difference between methods 2, 3, and 5 is that different methods are used to distinguish between noise-dominant IMFs and useful signal-dominant IMFs. The difference between methods 4 and 5 is that they handle noise-dominant IMFs differently.

[0210] Using the ratio of noisy signal to noise reduction error R DNSN To evaluate the denoising performance of the above methods, when R DNSN A larger value indicates a smaller denominator, meaning less noise is removed and the denoising effect is less significant; conversely, a smaller value indicates more significant denoising and a better effect. The calculation formula is as follows, and the denoising results are shown in the figure. Figure 4 .

[0211]

[0212] Where x'(t) is the normalized pipeline intrusion signal, x'(t) is the denoised signal, and T" is the number of sampling points.

[0213] T” = 20000;

[0214] Figures 3(1) to 3(10) The presentation shows the time-domain waveforms of manually mined signals after denoising using five different methods, along with the corresponding time-frequency graphs obtained through short-time Fourier transform. (Comparison) Figures 3(1) to 3(10) As shown in Figure 3(9), after denoising using the ICEEMDAN-FE-AIT method, the non-vibration segment of the intrusion signal is smoother, while the vibration segment retains the useful signal to the greatest extent. Combined with the time-frequency diagram, it can be seen that the ICEEMDAN-FE-AIT method has the most significant noise suppression effect, effectively avoiding the loss of useful signal information. Figure 4 A detailed comparison of the five denoising methods also shows that the method proposed in this invention achieves the best denoising effect. From... Figure 5 R DNSN The results show that, compared with the other four methods, the ICEEMDAN-FE-AIT method has the best overall denoising effect on the five real signals, with its R... DNSN The corresponding values ​​are 15.2043dB, 16.7654dB, 14.9815dB, 15.5541dB, and 13.5428dB.

[0215] Figure 6 The Fisher scores for different features are displayed, and the Fisher scores for each feature are significantly different. The number of features is set according to the Fisher scores from high to low, starting with 5 and increasing by 5 at a time until 30. Figure 7 Various activation functions were compared, including hardline, sigmoid, sinusoid, tanh, softmax, softplus, relu, and some variants of relu. Figure 7 The comparison results show that the highest recognition accuracy is achieved when using the softplus activation function and setting the number of hidden layer neurons to 300. Therefore, subsequent ELM experiments will use the softplus activation function and set the number of hidden layer neurons to 300. Table 4 compares the average values ​​of the recognition results for the five signals under different numbers of features.

[0216] Table 4 shows the average values ​​of the five signal recognition results under different numbers of features.

[0217]

[0218] The data in Table 4 shows that as the number of features increases, the recognition accuracy, precision, and recall all show an upward trend. The rate of increase gradually slows down as the number of features increases. When the number of features reaches 25, the recognition accuracy reaches 92.84%, which is the highest accuracy among all feature selections. This demonstrates that feature selection can effectively improve the recognition accuracy.

[0219] Table 5. Average recognition results of five signals before and after noise reduction.

[0220]

[0221] Table 6 shows the recognition accuracy of each of the five signals.

[0222]

[0223] Tables 5 and 6 compare the average recognition results and the individual recognition accuracy of the five signals before and after denoising. The data in the tables show that this method achieves the highest recognition accuracy, with recognition accuracies of 92.11%, 97.50%, 87.86%, 91.97%, and 96.91% for the five signals, respectively, and an average accuracy of 93.27%.

[0224] This invention first performs preprocessing on pipe intrusion signals acquired by uwDAS, including framing and normalization. Then, it uses ICEEMDAN-FE-AIT to denoise the preprocessed signal, extracting 30 features from the time and frequency domains to create a dataset. Finally, the dataset is input into the F-ELM model for identification. Compared to traditional denoising methods, this invention's denoising method achieves the optimal ratio of noisy signal to denoising error R for pipe intrusion signals acquired by uwDAS. DNSN This approach achieves the most significant noise suppression while preserving useful signals to the greatest extent possible. By calculating the Fisher score of each feature in the dataset and selecting appropriate features, as well as choosing suitable activation functions and the number of hidden layer neurons when using the ELM classifier, the accuracy of pipeline intrusion signal recognition is significantly improved.

Claims

1. A method for identifying uwDAS pipeline intrusion signals based on ICEEMDAN-FE-AIT and F-ELM, characterized in that... Includes the following steps: Step 1: Use uwDAS to collect pipe intrusion signals For pipeline intrusion signals Preprocessing is performed to obtain the preprocessed signal. ; Step 2: Process the preprocessed signal Perform ICEEMDAN decomposition to obtain multiple IMF mode components; Step 3: Calculate the fuzzy entropy values ​​of each order of IMF. FE ,Will IMF within the threshold range is determined to be noise-dominated. ; Step 4: Noise Dominance Perform AIT processing to dominate the AIT-processed noise. The signal is combined with the remaining useful signal that dominates the IMF for reconstruction to obtain the denoised signal. ; Step 5: Extract the denoised signal The time-domain and frequency-domain features were analyzed to create a dataset, and labels were assigned according to different types of pipeline intrusion signals. Step 6: Divide the dataset into training and test sets, and input them into the F-ELM model for recognition; In step 3, the IMF is calculated. k The fuzzy entropy includes the following steps: Step 3.1: For a univariate time series of length N: , Representing sequences respectively No. Each value constructs a set m 3D vector: ； In the above formula, u IMF k , ,in, m and d These represent the embedding dimension and latency, respectively. Indicates the construction m 3D vector; Indicates the first Starting from a point, continuously m indivual u The value; Step 3.2: Define two m 3D vector and Distance between : ； In the above formula, ; express u The global average, Glb Represents the global features of the embedding vector; Indicates another m 3D vector; Indicates Chebyshev distance; Step 3.3: Calculate the two vectors using the exponential function. and Similarity between them: ； In the above formula, n e , r These are the gradient and width of the boundary of the exponential function, respectively; Represent two vectors and The similarity between them; Represent two vectors and The distance between them; Step 3.4: Define the function : ； In the above formula, Represent two vectors and The similarity between them; Step 3.5: Fuzzy entropy is defined as: ； In the above formula, Represents fuzzy entropy, Indicates the embedding dimension as m When +1, repeat the functions obtained from steps 3.1 to 3.4; Step 3.6: Calculate the fuzzy entropy for all modal components (IMFs) according to steps 3.1 to 3.

5. FE ,Will IMF within the threshold range is determined to be noise-dominated. ; Indicates the lower limit of the threshold range. This indicates the upper limit of the threshold range.

2. The uwDAS pipeline intrusion signal identification method based on ICEEMDAN-FE-AIT and F-ELM according to claim 1, characterized in that: In step 1, uwDAS is used to collect pipeline intrusion signals. ,right Preprocessing is performed to obtain The details are as follows: First of all The process involves frame segmentation, followed by normalization. The normalization formula is as follows: ； In the above formula, The preprocessed signal p For the number of data points; These represent the data of each frame after framing. Data points.

3. The uwDAS pipeline intrusion signal identification method based on ICEEMDAN-FE-AIT and F-ELM according to claim 1, characterized in that: Step 2 includes the following steps: Step 2.1: Calculate the first-order residual And the first-order modal component IMF1: ； ； In the above formula, This represents the average of multiple signals. Represents the local mean of the signal; , , The number of Gaussian white noise particles added when constructing the signal; Indicates the first i Gaussian white noise with a mean of 0 and a variance of 1; , This represents the calculated standard deviation. The desired signal-to-noise ratio coefficients added at the beginning of the decomposition. This represents the first-order white noise mode component after EMD decomposition; Step 2.2: Calculate the second-order residual Second-order modal component IMF2: ； ； In the above formula, , This represents the standard deviation of the first-order residuals; Second-order white noise mode components after EMD decomposition; Step 2.3: For K is the number of modal components in the final decomposition. Calculate the... k order residual and the k First-order modal component IMF k : ； ； In the above formula, , Indicates the first k The standard deviation of the -1st order residual Indicates the first EMD decomposition. k White noise modal components of order 1 Indicates the first k -1st order residual; Step 2.4: Repeat step 2.3 until all residuals and IMF modal components are obtained, then stop the decomposition.

4. The uwDAS pipeline intrusion signal identification method based on ICEEMDAN-FE-AIT and F-ELM according to claim 1, characterized in that: In step 4, AIT processes the data as follows: In the above formula, Indicates the denoised first... indivual ; Indicates the first indivual No. interval Samples within; Indicates the first interval Extreme points within; Indicates the first indivual The denoising threshold is determined by the energy of each intrinsic mode, the number of samples, and the degree of decay of information entropy. ; It is an adjustable parameter; express indivual No. A range; Indicates the first indivual The first component There are several extreme points; Indicates the first indivual ; Indicates the first interval; Indicates the first indivual Denoising threshold and the first interval The ratio of the absolute values ​​of the extreme points within the range ; For all noise dominance Perform AIT processing to obtain the results. The denoised signal is obtained by combining the remaining useful signal with the dominant IMF and reconstructing the signal. : In the above formula, M 1. M 2 represents the noise-dominant factor determined based on the fuzzy entropy value. The order range; Res is the value at which the decomposition terminates. ; This represents the number of modal components in the final decomposition. Indicates the first k First-order modal components.

5. The uwDAS pipeline intrusion signal identification method based on ICEEMDAN-FE-AIT and F-ELM according to claim 1, characterized in that: In step 5, the denoised signal is extracted. We analyzed the time and frequency domain features and created a dataset, assigning labels based on different types of pipeline intrusion signals.

6. The uwDAS pipeline intrusion signal identification method based on ICEEMDAN-FE-AIT and F-ELM according to claim 1, characterized in that: In step 6, the dataset is divided into a training set and a test set, and then input into the F-ELM model for recognition. The F-ELM model consists of two parts: calculating the Fisher score for each feature in the feature dataset, and then selecting features with high Fisher scores to be fed into the ELM classifier for recognition.

7. The uwDAS pipeline intrusion signal identification method based on ICEEMDAN-FE-AIT and F-ELM according to claim 6, characterized in that: Step 6 includes the following steps: Step 6.1: Calculate the Fisher score for each feature in the dataset. The Fisher score is an evaluation metric for feature selection. The specific formula for calculating the Fisher score is as follows: In the above formula, c is the number of sample categories. It is the first The number of samples in each class For the first The first feature The mean of the class samples, For the overall sample at the 1st The mean of each feature No. In the class of samples, the first The variance of each feature; Step 6.2: Sort the features according to their Fisher scores and select the features with high Fisher scores; Step 6.3: Train the ELM model using the selected features based on the Fisher score, including: Step 6.3.1: The expression for the input matrix IN of the ELM classifier is: In the above formula, each column represents a training sample, and each row represents a feature; there are a total of Each training sample contains _ training samples_. a If there are features, then the input matrix IN has a OK List; The actual output matrix Y is expressed as follows: In the above formula, the actual output matrix Y is c OK Each column represents the output of a training sample, and the output of each training sample includes the output... c Each output value corresponds to a different intrusion signal category. Step 6.3.2: Randomly select the connection weights between the input layer and the hidden layer. w and the bias of hidden layer neurons b ; Connection weights between the input layer and the hidden layer w The expression is as follows: In the above formula, This represents the connection weights between the first neuron in the input layer and the first neuron in the hidden layer. This represents the connection weights between the second neuron in the input layer and the first neuron in the hidden layer. The meaning of the input layer a The connection weights between each neuron and the first neuron in the hidden layer. This represents the connection weights between the first neuron in the input layer and the second neuron in the hidden layer. This represents the connection weights between the second neuron in the input layer and the second neuron in the hidden layer. Indicates the input layer's first... a The connection weights between the first neuron and the second neuron in the hidden layer. This indicates that the first neuron in the input layer and the first neuron in the hidden layer... l The connection weights between neurons This indicates that the second neuron in the input layer and the first neuron in the hidden layer... l The connection weights between neurons Indicates the input layer's first... a The first neuron and the hidden layer l Connection weights between neurons This indicates the number of neurons in the hidden layer. Indicates the number of neurons in the input layer; The bias b of the hidden layer neurons is expressed as follows: In the above formula, This indicates the bias of the first neuron in the hidden layer; This indicates the bias of the first neuron in the hidden layer; Indicates the hidden layer number 1 Bias of each neuron; Step 6.3.3: Calculate the hidden layer output matrix H; Choose an infinitely differentiable function as the activation function g(x) of the hidden layer neurons, and denote the activation function between the hidden layer and the output layer. Connection weight is The expression is: In the above formula, This represents the connection weights between the first neuron in the hidden layer and the first neuron in the output layer. This represents the connection weights between the first neuron in the hidden layer and the second neuron in the output layer. This indicates that the first neuron in the hidden layer and the first neuron in the output layer... c The connection weights between neurons This represents the connection weights between the second neuron in the hidden layer and the first neuron in the output layer. This represents the connection weights between the second neuron in the hidden layer and the second neuron in the output layer. This indicates that the second neuron in the hidden layer and the first neuron in the output layer... c The connection weights between neurons Indicates the hidden layer number 1 l The connection weights between each neuron and the first neuron in the output layer. Indicates the hidden layer number 1 l The connection weights between each neuron and the second neuron in the output layer. Indicates the hidden layer number 1 l The nth neuron and the output layer c The connection weights between neurons; Indicates the number of neurons in the output layer; Obtain the predicted output matrix of the ELM classifier. T for: In the above formula, , , , These represent the prediction output matrix, i.e., the first... One output; in, In the formula, , They represent the input layer's first... The first neuron and the hidden layer The connection weights between neurons ; ; Representing the input matrix IN, respectively Line number The value of the column, here Indicates to Transpose; ; This represents the output of the q-th hidden layer neuron for the 1st sample. Let represent the output of the q-th hidden layer neuron for the second sample. Let represent the output of the q-th hidden layer neuron corresponding to the c-th sample; Indicates the hidden layer number 1 The connection weights between each neuron and the first neuron in the output layer. Indicates the hidden layer number 1 The connection weights between the first neuron and the second neuron in the output layer. Indicates the hidden layer number 1 The nth neuron and the output layer c The connection weights between neurons The hidden layer is represented by the first layer. Bias of each neuron; In the above formula, the predicted output matrix T Represented as: Hβ = T ', T 'for T transpose; The hidden layer output matrix is ​​calculated. H : in, These represent the connection weights between the input layer and the hidden layer, respectively. w The OK; These represent the first and second halves of the input matrix IN, respectively. List; Indicates that the input is activation function, Indicates that the input is activation function, Indicates that the input is activation function, Indicates that the input is activation function, Indicates that the input is activation function, Indicates that the input is activation function, Indicates that the input is activation function, Indicates that the input is activation function, Indicates that the input is Activation function; Step 6.3.4: Calculate the optimal connection weights between the hidden layer and the output layer. ; β The value of is obtained by solving the least-squares solution of the following expression: in: This represents the connection weights between the hidden layer and the output layer. Represents the predicted output matrix T transpose; Its optimal solution is The expression is as follows: In the above formula, H + Hidden layer output matrix H Moore-Penrose generalized inverse, ELM classifier training complete; Step 6.3.5: Test the test set with the trained ELM model to determine the category of the test set.

Images