Fiber-optic pipeline safety early warning algorithm based on deep learning

Abstract

The application relates to a kind of optical fiber pipeline safety early warning algorithm based on deep learning capable of effectively improving early warning accuracy, high environmental fitness, characterized by including: the optical fiber signal collected by distributed optical fiber vibration sensing system is denoised, and vibration wave ripple signal is obtained;Calculate the characteristics of the vibration wave ripple signal of each spatial position point, and the signal characteristics of all position points are formed into a space-time feature image in time-space dimension;Obtain the local space-time feature image of the invasion position point and the invasion position point;Channel one conv2D network is constructed, channel two conv1D-LSTM network is constructed, the features output by the double-channel network are fused, the one-dimensional vibration wave ripple signal characteristics and two-dimensional local space-time image characteristics of the invasion point are fully utilized, a full connection layer, a Dropout layer and a classification layer are added after network fusion, forming a double-channel deep learning network;Signal classification identification is carried out, the invasion behavior that destroys pipeline safety is monitored, and interference is effectively shielded.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of distributed optical fiber vibration sensing signal processing, in particular to an optical fiber pipeline safety early warning algorithm based on deep learning, which can effectively improve the early warning accuracy and has high environmental adaptability. SUMMARY

[0002] Third-party construction and other damage events often occur around oil and gas pipelines in the petroleum and petrochemical industry. Once oil and gas leaks occur, it will cause serious economic losses to the country and society, and directly threaten the safety of people's life and property. Traditional manual inspection methods cannot meet the current pipeline safety monitoring needs. In recent years, distributed optical fiber vibration sensing systems have been widely used in the field of pipeline safety in the oil and gas industry. Compared with traditional monitoring methods, optical fiber pipeline safety early warning technology has the advantages of long-distance monitoring, accurate positioning and fast response.

[0003] The intrusion alarm algorithm is the key to realizing pipeline safety monitoring by the distributed optical fiber vibration sensing system. Due to the complexity of interference noise and the diversity of deployment environment, accurate optical fiber pipeline safety monitoring still faces challenges. The algorithm aims to accurately locate and quickly respond to human excavation, mechanical construction and other damage behaviors, while shielding various environmental disturbances such as vehicle passing. How to improve the alarm accuracy of the optical fiber pipeline safety monitoring system and improve the environmental adaptability of the algorithm is a key problem that needs to be solved at present.

[0004] In current optical fiber pipeline safety monitoring technology, the threshold method or traditional machine learning algorithm often has poor environmental adaptability and low algorithm accuracy in practical application. In recent years, deep learning algorithms have been increasingly applied to the field of optical fiber pipeline safety monitoring due to their good generalization performance. However, most existing deep learning networks are used to learn signal features, and how to construct an efficient deep learning network to more fully mine and utilize different event behavior features still needs further research. In addition, existing algorithms focus on using deep learning networks to mine mixed time-frequency domain features or feature image information of signals, and the insufficient mining of different event behavior features leads to poor environmental adaptability and low alarm accuracy of the algorithm, especially for the high false alarm rate of passing vehicle interference. SUMMARY

[0006] The present application proposes an optical fiber pipeline safety early warning based on deep learning, which can effectively improve the early warning accuracy and has high environmental adaptability.

[0007] The present application achieves the following measures:

[0008] An optical fiber pipeline safety early warning algorithm based on deep learning, characterized by the following steps:

[0009] Step 1: denoising the optical fiber signal collected by the distributed optical fiber vibration sensing system to obtain a vibration ripple signal; calculating the characteristics of the vibration ripple signal of each spatial position point, and forming a space-time characteristic image of all position points in the time-space dimension;

[0010] Step 2: image detection is performed on the space-time characteristic image to obtain an intrusion position point and a local space-time characteristic image of the intrusion position point;

[0011] Step 3: a channel one conv2D network is constructed, the input of the network is the local space-time characteristic image of the intrusion position, a channel two conv1D-LSTM network is constructed, the input of the network is the vibration ripple signal, the characteristics output by the double-channel network are fused, the one-dimensional vibration ripple signal characteristics and the two-dimensional local space-time image characteristics of the intrusion point are fully utilized, and a full connection layer, a Dropout layer and a classification layer are added after network fusion to form a double-channel deep learning network;

[0012] Step 4: using the model obtained by training the above network to perform signal classification and recognition, monitoring the intrusion behavior that damages the safety of the pipeline, and effectively shielding the interference.

[0013] The original optical fiber signal collected by the distributed optical fiber vibration sensing system in step 1 of the application is represented in the form of a two-dimensional matrix containing time domain information and space information: D t×l =(dij) t×l ,(i=1,2...,t;j=1,2...,l), wherein t represents the time dimension, and l represents the number of spatial position points; first, the wavelet denoising method is used to obtain the vibration ripple signal; then the zero-crossing rate characteristics of the vibration ripple signal of each spatial position point are calculated, and the zero-crossing rate characteristics of all position points are formed into a space-time characteristic image in the time-space dimension.

[0014] In step 3 of the application, the double-channel deep learning network includes a parallel space-time image feature extraction layer and a signal vibration ripple feature extraction layer, a feature fusion layer, a full connection layer, a Dropout layer and a classification layer, wherein the channel one network includes three two-dimensional convolution blocks (conv2D), each convolution block includes a convolution sublayer, a batch normalization sublayer, an activation function sublayer, a convolution sublayer, a batch normalization sublayer, an activation function sublayer and a pooling sublayer connected in sequence; the channel two network includes two one-dimensional convolution blocks (conv1D) and a long short-term memory network (LSTM layer), each convolution block includes a convolution sublayer, an activation function sublayer, a convolution sublayer, an activation function sublayer and a pooling sublayer connected in sequence; the double-channel network is flattened respectively, the features are fused, and a full connection layer, a Dropout layer and a classification layer are connected in sequence.

[0015] In step 3 of this invention, the activation function sublayer of the dual-channel deep learning network uses the ReLU function; the pooling sublayer uses max pooling; the flattening operation uses global average pooling instead; the dropout layer is used to prevent overfitting, randomly deactivating neurons with a probability of 0.5 in each iteration of training to avoid the model relying too much on certain local features and improve the model's generalization performance; the classification layer uses the softmax function.

[0016] Step 3 of this invention also includes training a deep learning model, the specific process of which is as follows:

[0017] Step 3-1: Initialize the parameters of the dual-channel deep learning model, including the weight parameter w and the bias b;

[0018] Step 3-2: Input the training sample features into the dual-channel deep learning model for forward propagation to obtain the sample predicted labels;

[0019] Step 3-3: Calculate the loss values ​​for the predicted label and the true label using the cross-entropy loss function. The calculation formula is as follows: Where x represents a sample, n represents the total number of samples, and a and y represent the predicted label and the true label of the sample, respectively;

[0020] Steps 3-4: Calculate the gradient of each learning parameter using backpropagation based on the loss value, dynamically adjust the learning rate of each parameter using the Adam optimization algorithm, update the model parameters using gradient descent, and minimize the cross-entropy loss function through continuous iteration.

[0021] Step 3-5: Use the loss value to determine whether the dual-channel deep learning model has converged. If it has converged, end the training process; otherwise, go to step 3-2.

[0022] This invention constructs a dual-channel deep learning network, utilizing a Conv1D-LSTM network to mine the vibration ripple features of the signal, and simultaneously using a Conv2D network to mine the spatiotemporal image features of the optical fiber signal, thereby significantly improving the recognition accuracy of different behaviors. The optical fiber pipeline safety monitoring algorithm in this solution can effectively monitor intrusion behaviors that threaten pipeline safety, such as mechanical excavation and human excavation, and can shield against environmental interference such as passing vehicles and water pumps. The model has high universality and strong stability, solving the technical problems of poor generalization and high false alarm rate of current algorithm models in different environments. Attached image description:

[0023] Appendix Figure 1 This is a flowchart of the present invention.

[0024] Appendix Figure 2 This is a diagram of the dual-channel deep learning network structure in this invention.

[0025] Figure 1 is a schematic diagram of the vibration ripples of the different behaviors filtered signals in Example 1 of the present application, wherein Figure 3 (a) is a schematic diagram of the mechanical excavation signal, attached Figure 3 (b) is a schematic diagram of the artificial excavation signal, attached Figure 3 (c) is a schematic diagram of the vehicle passing interference signal, attached Figure 3 (d) is a schematic diagram of the water pump interference signal. Figure 3

[0026] Figure 2 is a spatiotemporal feature image of the intrusion location in Example 1 of the present application, wherein Figure 4 (a) is the spatiotemporal feature image corresponding to the mechanical excavation signal, Figure 4 (b) is the spatiotemporal feature image corresponding to the artificial excavation signal, Figure 4 (c) is the spatiotemporal feature image corresponding to the vehicle passing interference signal, Figure 4 (d) is the spatiotemporal feature image corresponding to the water pump interference signal. Figure 4

[0027] Figure 3 is a schematic diagram of the cross-entropy loss and accuracy change curve in the model training process in Example 1 of the present application, wherein Figure 5 (a) is the cross-entropy loss curve, Figure 5 (b) is the training set and validation set accuracy change curve. Figure 5

[0028] Figure 4 is a schematic diagram of the confusion matrix in the algorithm model test in Example 1 of the present application. Figure 6 DETAILED DESCRIPTION

[0029] The present application will be further described below in conjunction with the drawings and examples.

[0030] Example 1

[0031] This example proposes a fiber pipeline safety early warning algorithm based on deep learning, and the algorithm flow is shown in Figure 1, which specifically includes the following steps: Figure 1

[0032] First, denoise the original signal collected by the distributed optical fiber vibration sensing system to obtain the vibration ripple signal; calculate one or more features of the vibration ripple signal at each spatial location point, and the signal features of all location points are represented as spatiotemporal feature images in the time and space dimensions, and finally one or more spatiotemporal feature images are obtained.

[0033] Second, image detection is performed on the spatiotemporal feature image to obtain the local spatiotemporal feature image of the intrusion location point and the intrusion location.

[0034] ​​​​​Thirdly, a dual-channel deep learning network is constructed, which includes a channel one conv2D network inputting the spatial-temporal feature image of the intrusion position, a channel two conv1D-LSTM network inputting the vibration ripple signal, a feature fusion layer, a full connection layer, a Dropout layer and a classification layer.

[0035] Fourthly, the trained model is used for signal classification and recognition to monitor the intrusion behaviors such as mechanical excavation and human excavation which endanger the safety of the pipeline, and effectively shield the interference such as passing vehicles.

[0036] In the first step of the example, the original signal collected by the distributed optical fiber vibration sensing system can be expressed in the form of a two-dimensional matrix containing time domain information and spatial information: D t×l = (dij) t×l , (i = 1, 2,..., t; j = 1, 2,..., l), where t represents the time dimension and l represents the number of spatial position points. First, the original signal of each position point is denoised by wavelet to filter out the direct current component and high-frequency noise, and the vibration ripple signal is obtained. The vibration ripple signals of mechanical excavation, human excavation, passing vehicle interference and water pump interference are shown in Figs. Figure 3 (a)-4(d) respectively. Figure 3

[0037] The zero-crossing rate of the vibration ripple signal of each spatial position point is calculated, and the zero-crossing rate features of all position points form a spatial-temporal feature image in the time and space dimensions. Image detection is performed on the spatial-temporal feature image to obtain the spatial-temporal image of the intrusion position, and the center point of the spatial-temporal image is taken as the intrusion position point. Figure 4 (a)-4(d) respectively.

[0038] Then a dual-channel deep learning network model is constructed, and the network structure is shown in Fig. Figure 2 .

[0039] The dual-channel deep learning network includes parallel spatial-temporal image feature extraction layers and signal vibration ripple feature extraction layers, a feature fusion layer, a full connection layer, a Dropout layer and a classification layer.

[0040] The channel one network includes three two-dimensional convolution blocks (conv2D), each of which includes a convolution sublayer, a batch normalization sublayer, an activation function sublayer, a convolution sublayer, a batch normalization sublayer, an activation function sublayer and a pooling sublayer connected in sequence.

[0041] ​The two-channel network comprises two one-dimensional convolution blocks (conv1D) and a long short-term memory network (LSTM layer), wherein each convolution block comprises, in sequence, a convolution sublayer, an activation function sublayer, a convolution sublayer, an activation function sublayer and a pooling sublayer; the two-channel network is flattened respectively for feature fusion, and is sequentially connected with a fully connected layer, a Dropout layer and a classification layer; the activation function sublayer uses a Relu function; the pooling sublayer selects a maximum value pooling (MaxPooling); the flattening operation selects a global average pooling instead; the Dropout layer is used to prevent overfitting, and randomly inactivates neurons according to a probability of 0.5 in each iteration training process, avoids that the model excessively depends on some local features, and improves the generalization performance of the model; and the classification layer uses a softmax function.

[0042] Next, a deep learning model is trained, and the specific process is as follows: (1) initializing parameters of the two-channel deep learning model, including weight parameters w and bias b; (2) inputting a training sample feature into the two-channel deep learning model for forward propagation to obtain a sample predicted label; (3) calculating a loss value of the predicted label and a real label by using a cross-entropy loss function, and the calculation formula is as follows:

[0043]

[0044] Wherein, x represents a sample, n represents a total number of samples, a and y are respectively a sample predicted label and a real label. The loss value is used to calculate a gradient of each learning parameter by back propagation, the learning rate of each parameter is dynamically adjusted by using an Adam optimization algorithm, the model parameters are updated in a gradient descent manner, and the cross-entropy loss function is minimized through continuous iteration;

[0045] (4) determining whether the two-channel deep learning model converges by using the loss value, if yes, ending the training process, otherwise, jumping to step (2).

[0046] Figure 5 The figure shows the cross-entropy loss and the accuracy in the training process. Figure 5 The horizontal axis of (a) and (b) represents the number of iterations, and the vertical axis represents the cross-entropy loss and the accuracy respectively, wherein the blue curve is the cross-entropy loss and the accuracy of the training sample, and the orange curve is the cross-entropy loss and the accuracy of the verification sample (10% of the samples are divided as a verification set and do not participate in the training). It can be seen that during the network training process, the training accuracy and the verification accuracy steadily improve, and the cross-entropy loss rapidly and stably converges.

[0047] The trained deep learning model is called to perform real-time signal classification and identification and alarm judgment: the vibration wave of all intrusion center points and the corresponding space-time feature image are input into the double-channel deep learning network model, wherein the vibration wave of the intrusion center point is input into the conv1D-LSTM network, and the space-time feature image is input into the conv2D network; the trained deep learning model is called to perform signal classification and identification, and the model real-time identification result is output.

[0048] Finally, the model real-time identification result is judged, if the identification is mechanical excavation or artificial excavation, the alarm information of the corresponding position is output, otherwise, no alarm output is performed.

[0049] The double-channel deep learning network structure constructed by the application is used for model training, and the model training accuracy and verification accuracy are 99.36% and 98.95%, respectively. 500 test samples of each class are selected for model accuracy test, and the confusion matrix is as shown in Figure 6 The class label 0 is mechanical excavation, 1 is artificial excavation, 2 is vehicle passing interference, and 3 is water pump interference. The test result algorithm model false negative rate is 1%, and the false positive rate is 0.7%.

[0050] The algorithm model of the application can accurately monitor the intrusion behavior of mechanical excavation, artificial excavation and other behaviors that damage the safety of the pipeline, effectively shield the interference behaviors of vehicle passing, water pump continuous interference, and adapt to various environmental scenes such as highway, wasteland, farmland and mud pit, and cable burying depth. The algorithm has high accuracy and strong universality.

Claims

1. A deep learning-based fiber optic pipeline security early warning algorithm, characterized in that, Includes the following steps: Step 1: Denoise the raw signal collected by the distributed fiber optic vibration sensing system to obtain the vibration ripple signal; Calculate the characteristics of the vibration ripple signal at each spatial location. The signal characteristics of all locations are represented as a spatiotemporal feature image in the time-space dimension, thus obtaining the spatiotemporal feature image. Step 2: Perform image detection on the spatiotemporal feature images to obtain the intrusion location points and spatiotemporal feature images of the intrusion locations; Step 3: Construct a channel 1 conv2D network with the spatiotemporal feature image of the intrusion location as input, and construct a channel 2 conv1D-LSTM network with the signal vibration ripple signal as input. Fuse the dual-channel feature results, add a fully connected layer, a dropout layer and a classification layer to form a dual-channel deep learning network, and make full use of the vibration ripple features of the filtered signal at the intrusion point and the spatiotemporal image features of the intrusion location. Step 4: Use the trained model to classify and identify signals, monitor intrusion behaviors that threaten pipeline safety, and effectively shield against interference; In step 3, the dual-channel deep learning network includes a parallel spatiotemporal image feature extraction layer and a signal vibration ripple feature extraction layer, a feature fusion layer, a fully connected layer, a Dropout layer, and a classification layer. The channel one conv2D network includes three two-dimensional convolutional blocks, each containing a sequentially connected convolutional sub-layer, a batch regularization sub-layer, an activation function sub-layer, another convolutional sub-layer, another batch regularization sub-layer, another activation function sub-layer, and a pooling sub-layer. The channel two conv1D-LSTM network includes two one-dimensional convolutional blocks and a long short-term memory network, each containing a sequentially connected convolutional sub-layer, an activation function sub-layer, another convolutional sub-layer, another activation function sub-layer, and a pooling sub-layer. The dual-channel networks are flattened, feature fusion is performed, and then sequentially connected to a fully connected layer, a Dropout layer, and a classification layer. Step 3 also includes training the dual-channel deep learning model, the specific process of which is as follows: Step 3-1: Initialize the parameters of the dual-channel deep learning model, including the weight parameter w and the bias b; Step 3-2: Input the training sample features into the dual-channel deep learning model for forward propagation to obtain the sample predicted labels; Step 3-3: Calculate the loss values ​​for the predicted label and the true label using the cross-entropy loss function. The calculation formula is as follows: Where x represents a sample, n represents the total number of samples, and a and y are the predicted label and the true label of the sample, respectively; Steps 3-4: Calculate the gradient of each learning parameter using backpropagation based on the loss value, dynamically adjust the learning rate of each parameter using the Adam optimization algorithm, update the model parameters using gradient descent, and minimize the cross-entropy loss function through continuous iteration. Step 3-5: Use the loss value to determine whether the dual-channel deep learning model has converged. If it has converged, end the training process; otherwise, go to step 3-2.

2. The fiber optic pipeline security early warning algorithm based on deep learning according to claim 1, characterized in that, The raw signal acquired by the distributed fiber optic vibration sensing system in step 1 is represented as a two-dimensional matrix containing both temporal and spatial information: D t×l =(dij) t×l ,(i=1 ,2...,t;j=1 ,2...,l), where t represents the time dimension and l represents the number of spatial location points; First, wavelet denoising is performed on the original signal of each location point to filter out the DC component and high-frequency noise, and the vibration ripple signal is obtained; The zero-crossing rate of the vibration ripple signal at each spatial location point is calculated, and the zero-crossing rate features of all location points form a spatiotemporal feature image in the time and space dimension.

3. The fiber optic pipeline security early warning algorithm based on deep learning according to claim 2, characterized in that, In step 3, the activation function sublayer of the dual-channel deep learning network uses the ReLU function; the pooling sublayer uses max pooling; the flattening operation uses global average pooling instead; the Dropout layer is used to prevent overfitting by randomly deactivating neurons with a probability of 0.5 in each iteration of training, thus avoiding the model from relying too much on certain local features and improving the model's generalization performance; the classification layer uses the softmax function.

Images