A method for diagnosing rotating machinery faults in oil production pumping wells based on a parallel diagnosis model

Abstract

The application discloses a kind of based on parallel diagnosis model's oil production machine well on rotating machinery fault diagnosis method, this method is aimed at oil production machine rotating machinery under strong noise condition Fault feature extraction is difficult, the problem that diagnosis model anti-interference ability is insufficient, constructs a kind of double branch fusion diagnosis model based on deformable convolution DCN and Transformer, model combines the local feature self-adapting extraction ability of deformable convolution and the global timing modeling advantage of Transformer, effectively improve the diagnosis robustness under noisy environment by multidimensional feature fusion mechanism, experimental results show that the method has higher average diagnostic accuracy under low signal-to-noise ratio conditions, significantly improve the diagnosis precision of oil production machine well on rotating machinery fault.

Description

Technical Field

[0001] This invention relates to the field of wellhead fault detection in oil well production, and uses deep learning methods to complete fault diagnosis of rotating machinery (such as motors, reducers and support shafts) on the wellhead of the oil well. Background Technology

[0002] As a key piece of equipment in crude oil extraction and production, oil production machines operate continuously under high loads under complex and harsh conditions. Their core components (crank pins, gearboxes, and motor bearings, etc.) are highly susceptible to mechanical failures, directly impacting the safety and continuity of oilfield production. To address this, establishing a fault diagnosis model for key components of oil production machines based on vibration signal analysis is of significant value for real-time equipment status monitoring and fault diagnosis. However, in actual production, the acquisition of vibration signals and fault diagnosis of oil production machines face the following challenges:

[0003] (1) When the pumping unit stroke is adjusted due to fluctuations in the downhole fluid supply, the load at the suspension point of the oil production unit will change non-stationarily. The characteristic frequency of the rotating parts will shift dynamically, and the frequency bands of the wear fault of the crank pin and the gear meshing frequency will overlap. Traditional spectrum analysis will find it difficult to distinguish between the two types of faults, which will affect the diagnostic accuracy.

[0004] (2) In addition, due to the complex and variable influence of external noise under the cluster well platform, the transient impact of early failures of rotating mechanical components (such as bearing microcracks and gear pitting) is submerged by strong background noise, resulting in a high failure detection rate.

[0005] To address the aforementioned issues, existing single-mode diagnostic methods have the following limitations: while one-dimensional vibration signal analysis can capture time-domain dynamic characteristics, it is difficult to characterize the spatial distribution characteristics of frequency band aliasing faults; at the same time, traditional time-frequency conversion methods (such as wavelet transform) have limited ability to characterize transient impacts in noisy environments. Summary of the Invention

[0006] This invention addresses the problems existing in the background technology by proposing a parallel diagnostic model that integrates deformable convolutional neural networks (DCN) and Transformers. The method comprises the following steps:

[0007] 1) Acquire one-dimensional vibration signals x from the rotating machinery on the oil well site. k Normalize it to a standard intensity signal The calculation process is as follows:

[0008]

[0009] In the formula, X represents all data points of the original vibration signal, and N represents the signal length.

[0010] 2) The standard signal is converted into a two-dimensional image using the Gram angular difference field (GADF), and an image sample fault dataset D is constructed. First, the normalized intensity time series is converted to polar coordinates, and the intensity values... and the corresponding timestamp t k After transforming to polar coordinates, the calculation process is as follows:

[0011]

[0012] In the formula φ k and r k Let intensity and time be represented by cosine, and M be the normalization factor. The time correlation at different intervals in polar coordinates is obtained by subtracting sinusoidal angles, and represented by the Gram angle difference field. The calculation process is as follows:

[0013]

[0014] In the formula, A is the Gram angle difference field matrix, θ k Representing time t k The angle of time.

[0015] 3) Construct a deformable convolutional DCN-Transformer dual-branch hybrid diagnostic architecture. DCN introduces a dynamic offset Δp for each point on the basis of traditional convolution. n The calculation process is as follows:

[0016]

[0017] In the formula, p0 represents the position coordinates of the feature map, p n y(p0) is a predefined sampling offset of p0 within the convolution kernel range, y(p0) represents the value of the output feature map at position p0, w is the weight on the corresponding convolution kernel, and the offset pixel value is obtained by using the bilinear interpolation method.

[0018] 4) The self-attention mechanism is a core component of the Transformer. It transforms the input sequence into three feature vector matrices through a linear transformation: query Q, key K, and value V. The attention score is calculated as follows:

[0019]

[0020] In the Transformer, positional information is injected into the input sequence through explicit positional encoding. The calculation process is as follows:

[0021]

[0022] In the formula, pos and i represent the absolute position and dimension index of the position encoding vector in the time series, and d modelThe dimension representing the input vector is used to obtain the positional encoding (PE). By introducing a learnable positional encoding, the model can dynamically adapt to the temporal structure of the input sequence, thereby more accurately capturing the dependencies between positions and improving the performance of sequence modeling.

[0023] 5) A dynamic feature fusion mechanism based on cosine similarity is adopted. Feature fusion is achieved by calculating the cosine similarity between the local features extracted by DCN and the global temporal features extracted by Transformer. The fusion is performed based on the degree of similarity between features. The calculation process of cosine similarity is as follows:

[0024]

[0025] In the formula f dcn and f trans Let w be the feature vectors of deformable convolution and Transformer branch, respectively, σ be the Sigmoid function, and ||·|| denote the L2 norm. Mapping the similarity w to the fusion weight α yields the final weighted fused feature f. fusion The calculation process is as follows:

[0026] f fusion =α·f dcn +(1-α)·f trans

[0027] 6) During the model training phase, the training set is input into the parallel fusion network for training. Parameter configurations include: total training epochs (100 epochs), using the Adam optimizer, and setting the dynamic learning rate to an initial value of 0.001.

[0028] 7) After preprocessing the vibration signal of the rotating machinery of the oil production machine to be diagnosed, input it into the dual branch of the trained model, and output the fault type identification result to realize the fault diagnosis of the rotating machinery of the oil production machine.

[0029] Beneficial effects:

[0030] This invention combines the advantages of deformable convolutional neural networks (DCN) and Transformers: DCN, by dynamically adjusting the sampling position of the convolutional kernel, can accurately capture local features caused by faults; Transformers, utilizing a self-attention mechanism, effectively model the global temporal dependencies of long-sequence vibration signals. By comprehensively analyzing signal changes over long periods, it can better handle temporal interference from noise and distinguish between true fault signals and noise signals. This method overcomes the limitations of traditional CNNs in geometric transformation modeling and solves the problem of insufficient sensitivity of single Transformer models to subtle local features, providing a feasible engineering approach for fault diagnosis of oil production units under low signal-to-noise ratio conditions. Attached Figure Description

[0031] Figure 1 This is a diagram of the overall structure of the DCN-Transformer fusion network in this invention.

[0032] Figure 2 This is a flowchart of the deformable convolution DCN in this invention.

[0033] Figure 3 This is a curve comparing the diagnostic model's loss and accuracy.

[0034] Figure 4 Comparison chart of confusion matrix for model diagnosis.

[0035] Figure 5 Visualization of t-SNE dimensionality reduction for model diagnostic features. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, and the implementation effect of the method in the operation of an oil extraction machine will be illustrated by specific operating procedures. This embodiment is implemented based on the technical solution of the present invention, but the scope of protection of the present invention is not limited to the following embodiments.

[0037] This invention selects processed data from 120 wells of a crude oil extraction company from June 2023 to December 2024. The data sample covers five typical operating conditions: normal operation, motor bearing failure, abnormal gearbox wear, loose crank pin, and oil production machine structural imbalance. The effectiveness of the method presented in this paper for detecting faults occurring during oil production machine operation is analyzed, and a system is constructed as follows: Figure 1 The overall framework shown below has the following core processing flow:

[0038] 1) This invention first constructs a one-dimensional signal dataset S, and then uses the Gram angle difference field (GADF) signal conversion method to convert the acquired one-dimensional vibration signal of rotating machinery, mapping the intensity time series in the Cartesian coordinate system to the polar coordinate system. By calculating the sinusoidal angle difference between adjacent time points, a time correlation matrix in the polar coordinate system is constructed, ultimately forming a two-dimensional image characterizing the temporal features of the signal. This method effectively preserves the temporal correlation characteristics of the original signal through polar coordinate transformation, and uses the Gram angle difference field representation. The calculation process is as follows:

[0039]

[0040] In the formula, A is the Gram angle difference field matrix, θ k Representing time t k The angle of time.

[0041] 2) Deformable convolution introduces a learnable spatial offset, enabling the convolution kernel to automatically adjust its receptive field. Compared to the fixed sampling mode of traditional convolution, deformable convolution introduces a dynamic offset Δp. n This improves the flexibility of feature extraction. The specific calculation process is as follows:

[0042]

[0043] In the formula, p0 represents the position coordinates of the feature map, p n y(p0) is a predefined sampling offset of p0 within the convolution kernel range, y(p0) represents the value of the output feature map at position p0, and w is the weight on the corresponding convolution kernel. The offset pixel value is obtained using the bilinear interpolation method. The deformable convolution process is as follows: Figure 2 As shown, the offset is obtained by convolution on the original feature map. The convolution kernel has the same resolution as the current convolutional layer, and the generated offset field has the same resolution as the input feature map. During the training process, the convolution kernel and offset are continuously learned and optimized.

[0044] 3) The first branch converts the one-dimensional vibration signal of the oil production machine into a two-dimensional image using the Gram angular difference field (GADF) and enhances the extraction capability of local features using DCN; the second branch inputs the standardized vibration signal of the oil production machine into the Transformer encoder. The two branches extract features from the two dimensions of spatial local deformation and temporal global correlation, respectively, forming a complementary feature representation system.

[0045] The core component of the Transformer architecture, the self-attention mechanism, achieves global modeling of sequence data through dynamic weight allocation. This mechanism first generates three feature spaces—query Q, key K, and value V—from the input sequence through three independent linear transformation layers. It then calculates the similarity between Q and K and divides it by a scaling factor. After applying softmax to the acquired attention scores, normalized attention weights are obtained. Finally, the values ​​V are weighted and summed to obtain the final output. The calculation process is as follows:

[0046]

[0047] The Transformer employs a multi-head attention mechanism, which significantly enhances feature extraction capabilities through parallelization. This mechanism expands a single-head attention mechanism into h independent attention heads, each learning specific association patterns in a different feature subspace. The core idea is to split Q, K, and V into h heads through parallelization, with each head learning features in a different subspace. The model dimension is d. model Then the dimension of each head is d. k =d model / h, the independent calculation process is as follows:

[0048] head i =Attention(QW i Q ,KW i K VW i V )

[0049] In the formula, head i W represents the output value of the i-th attention head. i Q W i K and W i V Let be the transformation matrix of the i-th attention head. After concatenating the outputs of all heads, perform a linear transformation to obtain the final output of the multi-head attention mechanism. The calculation process is as follows:

[0050] MultiHead(Q,K,V)=Concat(head1,...,head h W O

[0051] In the formula, Concat represents the concatenation operation, W O This represents the output linear transformation matrix.

[0052] To effectively enhance the nonlinear modeling and expressive power of the model, after capturing global dependencies through self-attention, a feedforward neural network is used to nonlinearly encode the extracted information. In the Transformer, the feedforward neural network consists of two fully connected layers and a nonlinear activation function ReLU. The network FFN(x) is calculated as follows:

[0053] FFN(x) = max(0, xW1+b1)W2+b2

[0054] In the formula, x represents the input vector, W1 and b1 represent the parameter matrix and bias of the first layer, max represents the nonlinear activation ReLU function, and W2 and b2 represent the parameter matrix and bias of the second layer.

[0055] The feature fusion method based on dynamic adaptive weights achieves optimal fusion of local features and Transformer global features through cosine similarity measurement. This method first calculates the cosine similarity between the two feature vectors as the fusion weight, and then employs a dynamic weighting strategy for feature fusion, assigning a larger fusion weight to features with higher similarity.

[0056] 4) Use an accelerometer to collect vibration signals from key parts of the oil production machine (drive motor, gearbox crank, support shaft, etc.) during operation. Specific faults are shown in Table 1.

[0057] Table 1 Oil Production Machine Fault Information Data

[0058]

[0059]

[0060] To verify the model's advantages in low signal-to-noise ratio scenarios, comparative experiments were conducted using different network models (1DCNN, AlexNet, ResNet50, and DCN). The accuracy and loss function curves are shown below. Figure 3 As shown in the figure, 1DCNN suffers from limited ability to model fault vibration signals and extract features due to the low signal-to-noise ratio under harsh conditions, with an accuracy of only 58.5% and a relatively high training loss. To provide a more comprehensive analysis of model performance, the diagnostic test performance of each comparative model under low signal-to-noise ratio conditions is presented. The performance indicators include accuracy, precision, recall, and F1 score, and the results are shown in Table 2.

[0061] Table 2 Comparison of Model Diagnostic Performance under Low Signal-to-Noise Ratio Conditions

[0062]

[0063] As shown in Table 2, this invention verifies the superiority of the proposed DCN-Transformer fusion model by systematically comparing the performance of different models under low signal-to-noise ratio conditions. Experimental data shows that the model has a high average diagnostic accuracy of 98.75%, far superior to the performance of DCN (87.25%). This is due to the model combining the local feature extraction capability of DCN with the temporal attention mechanism of Transformer, effectively solving the limitations of traditional methods in feature extraction under strong noise environments. 1DCNN exhibits significant performance vulnerability due to its fixed receptive field limitation. Although DCN improves its spatial feature extraction capability through dynamic convolution kernel shifting, it still struggles to completely overcome noise interference when used alone. The experimental results fully demonstrate that the DCN-Transformer model can effectively cope with the complex noise environment of industrial sites, providing a reliable solution for fault diagnosis of rotating machinery such as oil pumping units.

[0064] This invention evaluates the classification performance of different models using a confusion matrix system. Under low signal-to-noise ratio conditions, such as... Figure 4As shown, the performance of the 1DCNN model drops sharply, with an accuracy of less than 50% for the first two types of faults. AlexNet and ResNet50 perform poorly for the fourth type of fault, while the DCN model remains relatively stable, maintaining an accuracy of over 70% for all types of faults. These results further validate that the DCN-Transformer model has optimal and stable fault diagnosis capabilities under low signal-to-noise ratio conditions.

[0065] To deeply analyze the feature learning capability of diagnostic models under low signal-to-noise ratio conditions, the t-SNE dimensionality reduction algorithm is used for visualization analysis of the feature space. For example... Figure 5 As shown, the feature distribution of the DCN-Transformer model exhibits significant clustering advantages, maintaining excellent intra-cluster aggregation and inter-cluster separation. However, the feature distributions of other comparative models show obvious inter-class overlap, with severe aliasing of fault features in the 1DCNN and AlexNet models. These visualization results, from a feature space perspective, confirm the good noise robustness and feature discrimination ability of the DCN-Transformer model.

[0066] In summary, the proposed parallel diagnostic model integrating DCN-Transformer solves the problems of difficult feature extraction and weak noise resistance in the fault diagnosis of rotating machinery on oil wells. Experiments have verified that this method can significantly improve the diagnostic accuracy of rotating machinery faults on oil wells.

Claims

1. A method for fault diagnosis of rotating machinery on oil production rigs based on a parallel diagnostic model, characterized in that... A deformable convolutional neural network (DCN) and a parallel dual-branch architecture using Transformer are constructed. A dynamic feature fusion mechanism is employed to diagnose faults in rotating machinery on oil wells. Specific steps include: 1) Collect raw one-dimensional vibration signals of the rotating machinery on the oil well site under operating conditions. Divide the signals into subsequences using the sliding window method, and perform standardized preprocessing on the collected vibration signals to construct a one-dimensional signal dataset S, which is then divided into a training set S. train Validation set S val and test set S test ; 2) The standardized vibration signal is converted into a two-dimensional image using Gram angle difference field GADF. The temporal correlation of the signal is preserved by polar coordinate mapping. The oil production machine fault image sample dataset D is constructed, and its division ratio is consistent with that of the one-dimensional signal dataset. 3) Construct a two-branch fusion network model of deformable convolutional network (DCN) and Transformer. The DCN network extracts local image features by optimizing and learning the convolutional kernel and offset. In the constructed two-branch fusion network, DCN introduces a dynamic offset for each point on the basis of convolution. The calculation process is as follows: In the formula These are the position coordinates of the feature map. yes Predefined sampling offsets within the convolution kernel range, Indicates the output feature map at position The value of is given, where w is the weight on the corresponding convolution kernel. The offset pixel value is obtained using the bilinear interpolation method. This represents the set of locations sampled by the convolutional kernel on the feature map; 4) The global temporal dependency of a one-dimensional signal is modeled using the Transformer self-attention mechanism. The Transformer encoder contains a multi-head self-attention mechanism, residual connections and layer normalization networks, and a feedforward network. 5) A feature fusion mechanism based on cosine similarity is adopted to adaptively fuse features from deformable convolutional DCN branches and Transformer branches; 6) Input the training set into the dual-branch fusion network model for training until the training set loss and validation set loss tend to stabilize, and obtain the oil production machine rotating machinery fault diagnosis model; 7) Input the test signal into the rotating machinery fault diagnosis model and output the fault type identification result of the oil production machine rotating machinery.

2. The method for fault diagnosis of rotating machinery on oil wells based on a parallel diagnostic model according to claim 1, characterized in that... Step 1) Rotating machinery faults include motor bearing faults, reducer gear wear faults, crank-connecting rod mechanism loosening faults, and support shaft misalignment and imbalance faults. These faults involve one-dimensional vibration signals. The signal is converted to a standard intensity signal through normalization. The calculation process is as follows: In the formula, X represents all data points of the original vibration signal, and N represents the signal length.

3. The method for fault diagnosis of rotating machinery on oil wells based on a parallel diagnostic model according to claim 1, characterized in that... In step 2), the Gram difference field GADF first transforms the normalized intensity time series to the polar coordinate system, and the standard intensity signal... The value and the corresponding timestamp After transforming to polar coordinates, the calculation process is as follows: In the formula and Let intensity and time be represented by cosine, and M be the normalization factor. The time correlation at different intervals in polar coordinates is obtained by subtracting sinusoidal angles, and represented by the Gram angle difference field. The calculation process is as follows: In the formula, A is the Gram angle difference field matrix. Representing time t k The angle of time.

4. The method for fault diagnosis of rotating machinery on oil wells based on a parallel diagnostic model according to claim 1, characterized in that... In step 4), the self-attention mechanism in the Transformer branch allows the model to dynamically focus on information at different locations when processing sequential data, capturing long-distance dependencies. Through input embedding linear transformation, three vector sequences are obtained: query Q, key K, and value V. The attention score is calculated as follows: In the formula The scaling factor is used to apply softmax to the acquired attention scores to obtain normalized attention weights. The value vector V is then weighted and summed to obtain the final output.

5. A method for fault diagnosis of rotating machinery on oil wells based on a parallel diagnostic model, as described in claim 1, characterized in that... Step 5) uses a cosine similarity-based fusion method. By measuring the directional consistency between different feature vectors, it evaluates the similarity and difference between their features and fuses them based on the distance metric results to obtain the final weighted fused features. The calculation process for cosine similarity is as follows: In the formula and These are the feature vectors of deformable convolution and Transformer branches, respectively. For the Sigmoid function, Representing the L2 norm, the similarity is... Mapped to fusion weights The final weighted fusion features are obtained. The calculation process is as follows: 。

Images