A pipe section type intelligent identification method based on deep learning

Abstract

The application discloses a kind of pipe section type intelligent identification method based on deep learning, utilize IMU detection data to identify pipe section type, it is mainly related to pipeline internal detection field.Utilize the method as follows: pipe section type identification of deep learning to IMU detection data is carried out noise reduction, calculates pipeline bending strain;Pipe section division is carried out using sliding window, obtains the data set of pipe section bending strain and angular velocity;Convolutional Neural Network (CNN) is constructed;Bidirectional Long Short-Term Memory Network (BiLSTM) is constructed;CNN-BiLSTM deep learning model is built, parameter is set;Pipe section type data set is divided into training set and test set, model training is carried out in training set, the deep feature of pipe section type data is extracted, and pipe section type is identified on test set.The method of the application can efficiently and accurately identify the pipe section type on pipeline, which is helpful to carry out pipeline integrity management and safety evaluation on bending deformation pipe section.

Description

Technical Field

[0001] This patent relates to the identification of pipe section types such as bends, welds, bending deformations, and dents using IMU detection data, which belongs to the field of internal pipe inspection. Background Technology

[0002] Pipeline transportation has become the primary mode of oil and gas transportation due to its advantages such as low cost and short cycle time. However, long-distance oil and gas pipelines are laid over long distances, making it difficult to manage pipeline integrity through excavation and inspection. Furthermore, the complex terrain surrounding oil and gas pipelines often leads to bending and deformation due to external forces, threatening their safe operation. Pipeline inspection technology equipped with an IMU (Insulated Measurement Unit) can eliminate the need for excavation, allowing for comprehensive pipeline inspection. As the IMU travels through the pipeline, it continuously records data such as the IMU's angle, angular velocity, and acceleration, thereby obtaining information on the bending strain and displacement of the entire pipeline.

[0003] Oil and gas pipelines mainly consist of sections with bends, welds, and bending deformations and dents caused by external forces. IMU (Insulated Measurement Unit) data is used to identify these section types, and combined with ground-based calibration point systems to calibrate pipeline information for pipeline integrity management. IMU-detected angle data is often used as an evaluation criterion for pipeline bending deformation; therefore, calculating the bending strain of the pipeline by angle changes helps identify section types. However, in engineering practice, pipe section intervals are often divided based on strain thresholds, and then section types are screened, which suffers from low efficiency and inconsistent judgment standards. Currently, there is a lack of efficient and intelligent methods for pipeline feature identification using IMU data. Summary of the Invention

[0004] Given the low efficiency and inconsistent identification standards in existing manual pipe segment type identification methods, this application provides a deep learning-based intelligent pipe segment type identification method. A CNN-BiLSTM deep learning model is constructed within the PyTorch framework for pipe segment type identification. The method specifically includes the following steps:

[0005] Step 1: Wavelet transform is used to denoise the angular velocity and angle data acquired by the IMU. The angle data includes the pitch angle P and azimuth angle A, which are the angles between the direction of the internal detector's central axis and the horizontal and true north directions, respectively. The mileage interval of the IMU acquisition is Δs, and the pipe diameter is D. The pitch angle difference ΔP and azimuth angle difference ΔA between each acquisition mileage interval are calculated. The horizontal bending strain ε of the pipe is then calculated using the formula. h Vertical bending strain ε v And total bending strain ε.

[0006] Step 2: Use a sliding window to process the pitch angular velocity ω after step 1. y Angular velocity ω z Horizontal bending strain ε h Vertical bending strain ε v The pipe segments are divided based on the total bending strain ε. The width of the sliding window for capturing signals is set to w, and the step size of the sliding window is l. The pipe segment length is standardized, and the division stops when the pipe segment length L < w, yielding the total number of pipe segments N.

[0007] Step 3 involves constructing a Convolutional Neural Network (CNN), which mainly includes building the input layer, convolutional layers, activation layers, pooling layers, and fully connected layers. The convolutional layers are primarily used to extract features from the input signal, reducing the difficulty of network learning. Their convolution operations are... Where k and b represent the weights and biases of the i-th neuron in the l-th layer, and x is the l-th input of the j-th layer. Pooling layers are used to simplify computation and select and filter extracted features; their expression is: Where q represents l neurons in the i-th channel, and W represents the pooling kernel size.

[0008] Step 4 involves constructing a Bidirectional Long Short-Term Memory (BiLSTM) network. Long Short-Term Memory (LSTM) neural networks introduce memory units into recurrent neural networks, including forget gates, input gates, and output gates. Here, σ is the sigmoid function. As a candidate state, forget layer f t Its function is to determine which data to forget; input gate i t The LSTM network determines which data enters the current state, using the forget gate and input gate to decide which data to discard and which to keep, and the output gate... t The expressions for generating the output at the current moment are as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) o t =σ(W o ·[h t-1 ,x t ]+b o ) Where W and b are the weight matrix and bias to be trained, respectively, and h t-1 It is the output of the previous layer. BiLSTM consists of two LSTM layers, each with an input sequence, but transmitting information in opposite directions, i.e., forward LSTM and backward LSTM. The two hidden state vectors extracted from the forward and backward directions are concatenated to combine the characteristics of the preceding and following signals.

[0009] Step 5: Build the CNN-BiLSTM deep learning model. Convolutional layers are used to extract features from the input data. The extracted features are then fed into the BiLSTM layer to learn the sequential relationship between the input features. The output of the BiLSTM is then used as the input to the fully connected layer. The bidirectional architecture of the BiLSTM allows for the simultaneous extraction of information from both forward and backward time steps using forward and backward LSTM layers.

[0010] Step 6: Divide the data from Step 3 into training and testing sets, and normalize them. Use the normalized signals as input to the model for depth feature extraction. Set the number of neurons and kernel size for the convolutional layers. The convolutional layers traverse the entire training set of pipe segment type data sequences and extract the depth features of the input signals. Based on the extracted depth feature information, predict the pipe segment types of the test set data sequences and evaluate the CNN-BiLSTM model. Beneficial effects

[0011] Existing IMU (Insulated Measurement Unit) technology can calculate the bending strain of pipelines to reflect their bending deformation. However, traditional methods classify pipe segments based on strain thresholds to screen pipe types, which suffers from low efficiency and inconsistent identification standards. This application can intelligently and efficiently identify pipe segment types, including: straight pipe segments, welded pipe segments, hot-bent pipe segments, cold-bent pipe segments, bent and deformed pipe segments, and dented pipe segments. Identifying pipe segment types using this method can help improve pipeline integrity management and conduct safety assessments of bent and deformed pipe segments. Attached Figure Description

[0012] Figure 1 Divide the pipe segment data into the sliding window. Figure 2 This is a diagram of the BiLSTM structure. Figure 3 The structure diagram of the constructed CNN-BiLSTM model is shown. Figure 4 Confusion matrix for identifying each pipe segment type. Figure 5 To identify the precision, recall, and F1 score of various pipe segment types. Specific implementation methods

[0013] To more clearly illustrate the algorithm of this application, the following description is provided in conjunction with the accompanying drawings and specific embodiments. In the embodiment, the pipe inner diameter D is 1016 mm, and the pipe conveying pressure is 8 MPa. This patent application's method for intelligent identification of pipe segment types based on deep learning includes the following steps:

[0014] Step 1: The angle and angular velocity data acquired by the IMU are denoised using a 6th-order wavelet transform ("db4"), as shown in the figure. The pipe inner diameter D is 1016 mm, and the acquisition interval Δs is 0.05 m. Calculate the Δs between two points at interval Δs. i Δs i+1 The difference in pitch angle ΔP and azimuth angle ΔA between the intervals are calculated, and then the vertical strain ε on that interval mileage section is calculated. vi Horizontal strain ε hi and total strain ε i .

[0015] Step 2: Use a sliding window to process the pitch angular velocity ω after step 1. y Angular velocity ω z Vertical bending strain ε v Pipe sections are divided based on the total bending strain ε, such as Figure 1 As shown. The signal interception width of the sliding window is set to 6m, and the step size of the sliding window is 0.1m. A total of 10 types of pipe segments are divided into datasets: straight pipe segment, welded pipe segment, hot-bent pipe segment, cold-bent pipe segment, bent and deformed pipe segment, recessed pipe segment, straight pipe + bent and deformed, straight pipe + elbow, straight pipe + recessed, and bent and deformed + welded.

[0016] Step 3: Construct a CNN network using two convolutional layers. Convolutional layer 1 is configured with 32 channels, a kernel size of 3, and a stride of 1. Convolutional layer 2 is configured with 16 channels, a kernel size of 3, and a stride of 1. The activation function for the convolutional layers is ReLU. The kernel size for both max-pooling layers 1 and 2 is set to 3, and the stride is set to 1. Dropout_1 and Dropout_2 (dropout rate) are set to 0.2.

[0017] Step 4, the BiLSTM network built based on the bidirectional architecture of LSTM, such as... Figure 2 As shown, the input dimension of BiLSTM is set to 16, the output dimension to 128, the activation function to Tanh, and Dropout_3 to 0.5. The input dimension of linear layer 1 is set to 256, the output dimension to 64, and the activation function to ReLU. The input dimension of linear layer 2 is set to 64, the output dimension to 10, and the activation function to Softmax.

[0018] Step 5, construct the CNN-BiLSTM model, as follows: Figure 3 As shown, the model parameters are set as follows: the learning rate is 0.001; the loss function is cross-entropy loss; and the batch size is set to 128. The pipe segment type dataset from step 2 is normalized, and then the data is divided into training and test sets in an 8:2 ratio. The training set is input into the model for training. The model automatically extracts deep features from the training data (data length, standard deviation, minimum value, mean value, kurtosis, skewness, margin factor, etc.). The model tends to stabilize when the number of training epochs exceeds 200.

[0019] Step 6: Input the test set data into the model to predict the pipe segment type. The confusion matrix of the pipe segment type identification results (0—straight pipe segment type, 1—welded pipe segment type, 2—hot-bent pipe segment type, 3—cold-bent pipe segment type, 4—bent and deformed pipe segment type, 5—recessed pipe segment type, 6—straight pipe + bent and deformed pipe segment type, 7—straight pipe + elbow pipe segment type, 8—straight pipe + recessed pipe segment type, 9—bent and deformed + welded pipe segment type) can quickly reflect the identification effect of each pipe segment type, such as... Figure 4 As shown. The model's recognition performance was evaluated using precision, recall, and F1 score. The CNN-BiLSTM model achieved a precision of 0.965, a recall of 0.964, and an F1 score of 0.963 for identifying pipe segment types, demonstrating good recognition ability for various pipe segment types. Figure 5 As shown.

Claims

1. A method for intelligent identification of pipe segment types based on deep learning, characterized in that, The method includes: Acquire detection data from within the IMU, and perform noise reduction processing on the angular velocity and angle data acquired by the IMU using wavelet transform, where the angle data is the pitch angle. With azimuth These are the angles between the direction of the central axis of the internal detector and the horizontal direction and the due north direction, respectively; the mileage interval acquired by the IMU is... The pipe diameter is Calculate the pitch angle difference between each data acquisition interval. and azimuth difference Calculate the horizontal bending strain of the pipe according to the formula. Vertical bending strain and total bending strain ; Pipe segmentation is performed using the internal detection dataset and the bending strain dataset; A convolutional neural network (CNN) is constructed, comprising an input layer, convolutional layers, activation layers, pooling layers, and fully connected layers. The convolutional layers are used to extract features from the input signal, reducing the difficulty of network learning. The convolution operation is as follows: ,in k , b Indicates the first l layer i The weights and biases of each neuron x for j The first layer l One input; pooling layers are used to simplify computation, select and filter extracted features, and their expression is: ,in q For the first i On each channel l One neuron, W The size of the pooling kernel; A bidirectional long short-term memory (BiLSTM) network is constructed. The long short-term memory (LSTM) neural network in this BiLSTM network is based on a recurrent neural network and incorporates memory units, including a forget gate, an input gate, and an output gate. For the sigmoid function, For candidate states, forget layer Its function is to determine which data to forget; the input gate The LSTM network determines which data enters the current state through the forget gate and input gate, and the output gate... The expressions for generating the output at the current moment are as follows: in W and b These are the weight matrix and bias to be trained, respectively; BiLSTM consists of two layers of LSTM network, each with an input sequence, but transmitting information in opposite directions, namely forward LSTM and backward LSTM; the two hidden state vectors extracted from the forward and backward directions are connected to combine the characteristics of the signals before and after; Construct a CNN-BiLSTM deep learning model. Convolutional layers are used to extract features from the input data. The extracted features are then fed into the BiLSTM layer to learn the relationship between the input features in the order of their sequence. The output of the BiLSTM is then used as the input to the fully connected layer. Based on the constructed model, the pipe segment type is predicted, and the pipe segment type results are obtained.

2. The method as described in claim 1, characterized in that, Pipe segmentation is performed using the internal inspection dataset and the bending strain dataset, including: Using a sliding window to measure the pitch angular velocity after noise reduction angular velocity of heading Horizontal bending strain Vertical bending strain and total bending strain Perform pipe segment division; set the width of the sliding window to... The step size of the sliding window is ; Standardize pipe segment length, when the pipe segment length Stop dividing when the total number of pipe segment types is obtained. .

3. The method as described in claim 1, characterized in that, In the CNN-BiLSTM deep learning model, the bidirectional architecture of BiLSTM can extract information from both previous and next time points simultaneously from two directions through forward LSTM layers and backward LSTM layers.

4. The method as described in claim 1, characterized in that, Based on the constructed model, the pipe segment type is predicted, and the pipe segment type results are obtained, including: The pipe segment type data obtained by the aforementioned pipe segment division method is divided into two parts: a training set and a test set. The data is then normalized, and the normalized signal is used as the input to the model for deep feature extraction. The number of neurons and kernel size of the convolutional layer are set, and the convolutional layer traverses the entire training set pipe segment type data sequence, extracting the deep features of the input signal. Based on the extracted deep feature information, the pipe segment type of the test set data sequence is predicted, and the CNN-BiLSTM model is evaluated.

Images