A real-time torque-based fault diagnosis method for an electrically-driven reciprocating pump

Abstract

The application provides a real-time torque-based electric drive reciprocating pump fault diagnosis method, and belongs to the technical field of oil and gas development. The method comprises collecting a normal sample set and an abnormal sample set, and performing data preprocessing; a real-time torque and a convolutional neural network are used to construct an electric drive reciprocating pump fault diagnosis model; the preprocessed normal sample set and abnormal sample set are input into the electric drive reciprocating pump fault diagnosis model for supervised training, with 360° rotation of a crankshaft as a cycle; a loss function is used to train the electric drive reciprocating pump fault diagnosis model; in the process of training the electric drive reciprocating pump fault diagnosis model, the output feature map is recorded, all feature channels of the feature map are weighted and summed to generate a heat map, and fault positioning is performed through the heat map. The application solves the problem that when torque fluctuation abnormal faults occur in the power end or the hydraulic end of the electric drive reciprocating pump during operation, the conventional torque curve monitoring method cannot timely or accurately determine the fault position.

Description

Technical Field

[0001] This invention belongs to the field of oil and gas development technology, and in particular relates to a fault diagnosis method for an electric reciprocating pump based on real-time torque. Background Technology

[0002] In oil and gas development, fault diagnosis of electrically driven reciprocating pumps plays a crucial role in equipment maintenance. During oil and gas development operations, when problems occur at the power or hydraulic end of an electrically driven reciprocating pump, monitoring is typically performed using torque curves. A downward fluctuation in the torque curve indicates an abnormality in the hydraulic end, while an upward fluctuation indicates an abnormality in the power end. However, current torque curve monitoring methods can only provide a rough assessment of equipment malfunctions and cannot accurately determine the type and location of the fault. Summary of the Invention

[0003] To address the aforementioned shortcomings in the existing technology, this invention provides a fault diagnosis method for electric reciprocating pumps based on real-time torque. This method solves the problem that conventional torque curve monitoring methods cannot determine the fault location in a timely or accurate manner when abnormal torque fluctuations occur at the power or hydraulic end of the electric reciprocating pump during operation.

[0004] To achieve the above objectives, the technical solution adopted by this invention is: a fault diagnosis method for an electric reciprocating pump based on real-time torque, comprising the following steps: S1, Parameter Reading: Read the parameters of the reciprocating pump; S2. Data Acquisition: Collect torque curve data of the reciprocating pump under normal operating conditions as a normal sample set, and collect torque curve data of the reciprocating pump under typical fault conditions as an abnormal sample set. S3. Data Processing: Preprocess the normal and abnormal sample sets; S4. Model Establishment: Based on real-time torque and convolutional neural network, a fault diagnosis model for electric reciprocating pump is constructed. The pre-processed normal sample set and abnormal sample set are input into the electric reciprocating pump fault diagnosis model for supervised training with the crankshaft rotating 360° as the cycle. S5. Model Training: The fault diagnosis model of the electric reciprocating pump is trained using the loss function; S6. Fault Location: During the training of the fault diagnosis model for the electric reciprocating pump, the output feature map is recorded, and all feature channels of the feature map are weighted and summed to generate a heat map. The fault is located through the heat map.

[0005] The beneficial effects of this invention are as follows: This invention relates to a diagnostic method for abnormal main motor torque in an electrically driven reciprocating pump during oil and gas development caused by a fault in the power or hydraulic end. It employs real-time torque curve analysis to locate the fault in the power and hydraulic ends, enabling real-time monitoring of their operating status. When equipment malfunctions, the fault location is immediately determined, providing real-time guidance for operators to adjust equipment operation and accurate maintenance guidance for maintenance personnel. Compared to traditional torque curve monitoring, this invention increases the torque sampling frequency and utilizes a convolutional neural network algorithm to extract features from the torque dataset for each stroke. By comparing the results with a trained model, the fault type of the electrically driven reciprocating pump can be accurately determined.

[0006] Further, S3 includes the following steps: Based on normal and abnormal sample sets, each data point is labeled at the sample level, with each data point corresponding to a signal with a 360° cycle. Based on the annotation results, before the convolution operation, the padding data at the beginning of the sequence is obtained from the end of the sequence, and the bottom padding data at the end of the sequence is obtained from the beginning of the sequence. After basic data filling, outliers were removed using the three-standard-deviation criterion. After removing outliers, the data is standardized to complete the data preprocessing.

[0007] The beneficial effects of the above-mentioned further solutions are: to provide the model with clear data objectives and data foundation, and to build an accurate mapping relationship between "input signal and fault category".

[0008] Furthermore, the standardized expression is as follows: ; in, Indicates the first The standardized values ​​of each feature dimension Indicates the first The mean of each feature dimension, Indicates the first Standard deviation of each feature dimension Indicates the first Initial signal values ​​for each feature dimension.

[0009] The beneficial effects of the above-mentioned further solutions are: eliminating interference from multiple feature dimensions and accelerating model convergence.

[0010] Furthermore, the fault diagnosis model for the electrically driven reciprocating pump includes: The input layer is used to receive the preprocessed sample set data, where the input shape is (3600, 4), where 3600 corresponds to the length of the torque in the angular dimension, and 4 is the number of feature channels; Convolutional layers are used to extract features from the received sample set data to obtain local and global features. The convolutional layers include a first convolutional layer, a second convolutional layer, and a third convolutional layer. Pooling layers are used to output feature vectors based on local and global features. These pooling layers include a first pooling layer located between a first and a second convolutional layer, a second pooling layer located between a second and a third convolutional layer, and a third pooling layer. The first and second pooling layers both employ max pooling, while the third pooling layer employs global average pooling. Specifically, the first and second pooling layers perform max pooling on the local features extracted by the convolutional layers, and the third pooling layer performs global average pooling on the global features extracted by the convolutional layers. The fully connected layer is used to map and transform the feature vectors output by the pooling layer, and to learn and perform nonlinear transformations on the global features extracted by the convolutional layer; the fully connected layer includes a first fully connected layer and a second fully connected layer. The output layer is used for feature vectors, outputting the probability of fault categories in the electric reciprocating pump.

[0011] The beneficial effects of the above-mentioned further scheme are as follows: the input layer receives the multi-dimensional periodic signals of the pump to build an adaptation interface, the convolutional layer extracts fault features from local correlation to global correlation layer by layer, the pooling layer retains key features and reduces the model's computational cost through dimensionality reduction, the fully connected layer realizes feature compaction and cross-dimensional fusion, and the output layer completes the output of four types of fault probabilities through Softmax, forming a closed loop of "signal reception - progressive feature extraction - feature optimization - accurate classification".

[0012] Furthermore, The first convolutional layer is used to extract local and global features of the signal based on the received sample set data using one-dimensional convolutional kernels of different sizes. The first convolutional layer is set with a convolutional kernel of 64, a kernel size of 11, and an activation function of ReLU. The second convolutional layer is used to extract mesoscale features based on the features reduced by the first pooling layer using one-dimensional convolutional kernels, and to fuse the information of the four original feature channels into 64 intermediate channels. The second convolutional layer uses 128 one-dimensional convolutional kernels, the kernel size is 7, and the activation function is ReLU. The third convolutional layer is used to capture global correlation features and fault modes of the data based on the processing of the second pooling layer. The first pooling layer is used to perform dimensionality reduction on the features extracted by the first convolutional layer, and the pooling kernel size of the first pooling layer is set to 2. The second pooling layer is used to perform repeated pooling operations on the mid-scale features extracted by the second convolutional layer. The pooling kernel size of the second pooling layer is set to 2. The third pooling layer is used to calculate the average value in the angular dimension for each feature channel and output the feature vector; The first fully connected layer is used to perform comprehensive learning and nonlinear transformation on global features and to map and transform the feature vector output by the pooling layer. The activation function is ReLU. The second fully connected layer is used to optimize the compactness of feature representation. It removes redundant information and condenses key information from the high-dimensional feature vector of the first fully connected layer, providing accurate input to the output layer. The activation function is ReLU.

[0013] The beneficial effects of the above-mentioned further scheme are: clarifying the specific parameters of each layer to ensure the transmission of data features; realizing progressive fault feature mining of "local-medium-global" based on different convolution kernel sizes; and using two fully connected layers to make the feature vector more focused on the fault core.

[0014] Furthermore, the expression for the probability of the electric reciprocating pump failure category is as follows: ; in, Indicates a fault in the electric reciprocating pump. j The predicted probabilities of each category, The output of the fully connected layer represents the first... j Values ​​for each category, k This represents the category index corresponding to the fault. This indicates that the output of the fully connected layer belongs to k The original score for each category.

[0015] The beneficial effects of the above-mentioned further solutions are: outputting the failure probability of the model under various conditions, supporting the optimization and localization of the model.

[0016] Furthermore, the expression for the loss function is as follows: ; ; ; in, This represents the weighted focus-center mixed loss function. This represents the weighted focus loss function. Represents the central loss function. and All represent weighting coefficients. Indicates category weight, This represents the focusing factor, with a value of 2. Indicates the number of samples. Indicates the first i The sample belongs to the first jReal labels for each category Indicates the first i The sample belongs to the first j The predicted probabilities of each category, Indicates the first i The feature vector output by the second fully connected layer of each sample express The feature center of the category.

[0017] The beneficial effects of the above-mentioned further scheme are: by using a collaborative design of "weighted + focused + feature clustering", the imbalance of pump failure samples can be balanced, the learning of minor faults can be strengthened, the feature space distribution can be optimized, and the model prediction accuracy can be improved.

[0018] Furthermore, step S6 includes the following steps: During the training of the fault diagnosis model for the electric reciprocating pump, the backpropagation calculation of the third convolutional layer and the third pooling layer is used to output the original value without the Softmax layer, record the value and output the feature map. The feature map is weighted and summed to generate a heatmap. The data lengths on the heatmap are mapped one-to-one with the original torque curve data in the angular dimension, and the trough threshold is set through the abnormal feature waveform diagram. Continuous intervals in the heatmap that are above the trough threshold are identified as abnormal intervals. Compare the abnormal intervals with the interval divisions of the waveform diagram of torque and angle of the reciprocating pump; Based on the interval comparison results, the fault location is determined; The trained electric reciprocating pump fault diagnosis model is deployed in the reciprocating pump condition monitoring system to identify features and detect anomalies in real-time torque curve data, output fault types, and realize fault diagnosis of electric reciprocating pumps.

[0019] The beneficial effects of the above-mentioned further solution are: the heat map generated by the weighted feature channel is precisely aligned 1:1 with the angle dimension of the original torque curve, and by combining the abnormal waveform map to set the trough threshold and compare it with the torque-angle waveform division interval, the "angle-component" mapping of the fault location is realized.

[0020] Furthermore, the expression for the heatmap is as follows: ; in, This indicates the heatmap in the angular dimension. The value at each position, The neuron importance weights represent the feature channels. This indicates the feature channel in the angular dimension. Feature values ​​at each position, Indicates the characteristic channel, This represents the number of activation functions.

[0021] The beneficial effect of the above-mentioned further scheme is that it transforms the abstract fault characteristics of the model into an intuitive thermal distribution. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the fault diagnosis process of the present invention.

[0023] Figure 2 This is a diagram of a convolutional neural network architecture.

[0024] Figure 3 This is the theoretical torque curve of a normal five-cylinder fracturing pump under a certain working condition.

[0025] Figure 4 The theoretical torque curve of a five-cylinder fracturing pump with a damaged discharge valve of cylinder #1 under a certain working condition.

[0026] Figure 5 The theoretical torque curve of a five-cylinder fracturing pump with a damaged intake valve in cylinder #1 under a certain operating condition.

[0027] Figure 6 The theoretical torque curve of a five-cylinder fracturing pump with a damaged crosshead on the power end of cylinder #1 under a certain working condition. Detailed Implementation

[0028] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0029] Example like Figure 1 As shown, this invention provides a fault diagnosis method for an electric reciprocating pump based on real-time torque, the implementation method of which is as follows: S1, Parameter Reading: Read the parameters of the reciprocating pump; In this embodiment, the mechanical parameters of the reciprocating pump, the pulse signal of the key phase sensor, the real-time torque, the discharge pump pressure, and the pump surge are read.

[0030] S2. Data Acquisition: Collect torque curve data of the reciprocating pump under normal operating conditions as a normal sample set, and collect torque curve data of the reciprocating pump under typical fault conditions as an abnormal sample set. S3. Data Processing: Preprocessing is performed on the normal and abnormal sample sets. The implementation method is as follows: Based on normal and abnormal sample sets, each collected data point is labeled at the sample level, with each data point corresponding to a signal with a 360° period. Based on the labeling results, before the convolution operation, the padding data at the beginning of the sequence is obtained from the end of the sequence, and the bottom padding data at the end of the sequence is obtained from the beginning of the sequence. After basic padding, outliers are removed using the three-standard-deviation criterion. The data after outlier removal is then standardized to complete the data preprocessing.

[0031] In this embodiment, data preprocessing is performed on the normal and abnormal sample sets, including data labeling, standardization, data cleaning, and feature extraction. First, the sample sets are labeled. Each data point collected by the device (each data point corresponds to a signal with a 360° cycle) is labeled at the sample level. The labels include 0 (normal), 1 (damaged hydraulic discharge valve), 2 (damaged hydraulic suction valve), and 3 (damaged power end). This labeling result serves as the basis for subsequent model training. Second, to fully utilize the periodicity of the device data in the angular domain, a cyclic padding method is used to pad the data before convolution. That is, during convolution, padding data at the beginning of the sequence (to the left of 0°) is obtained from the end of the sequence (near 360°), and padding data at the end of the sequence is obtained from the beginning of the sequence, ensuring that the periodicity of the signal is not lost during convolution. Then, outliers caused by sensor malfunctions are removed using the SG filtering method based on the 3x standard deviation criterion. Finally, the input features are Z-score standardized to ensure that the processed signal has a mean of 0 and a standard deviation of 1. The standardization formula is as follows:

[0032] in, =1,2,3,4 (corresponding to 4 feature dimensions). For the first The standardized values ​​of each feature dimension For the first The mean of each feature dimension, For the first Standard deviation of each feature dimension Indicates the first Initial signal values ​​for each feature dimension.

[0033] S4. Model Establishment: A fault diagnosis model for the electric reciprocating pump is constructed based on real-time torque and convolutional neural networks. Preprocessed normal and abnormal sample sets are input into the electric reciprocating pump fault diagnosis model for supervised training, with a crankshaft rotation period of 360°. The electric reciprocating pump fault diagnosis model includes: The input layer is used to receive the preprocessed sample set data, where the input shape is (3600, 4), where 3600 corresponds to the length of the torque in the angular dimension, and 4 is the number of feature channels; Convolutional layers are used to extract features from the received sample set data to obtain local and global features. The convolutional layers include a first convolutional layer, a second convolutional layer, and a third convolutional layer. Pooling layers are used to output feature vectors based on local and global features. These pooling layers include a first pooling layer located between a first and a second convolutional layer, a second pooling layer located between a second and a third convolutional layer, and a third pooling layer. The first and second pooling layers both employ max pooling, while the third pooling layer employs global average pooling. Specifically, the first and second pooling layers perform max pooling on the local features extracted by the convolutional layers, and the third pooling layer performs global average pooling on the global features extracted by the convolutional layers. The fully connected layer is used to map and transform the feature vectors output by the pooling layer, and to learn and perform nonlinear transformations on the global features extracted by the convolutional layer; the fully connected layer includes a first fully connected layer and a second fully connected layer. The output layer is used for feature vectors, outputting the probability of fault categories in the electric reciprocating pump.

[0034] In this embodiment, the first convolutional layer is used to extract local and global features of the signal based on the received sample set data using one-dimensional convolutional kernels of different sizes. The first convolutional layer is configured with a convolutional kernel of 64, a kernel size of 11, and an activation function of ReLU. The second convolutional layer is used to extract mesoscale features based on the features reduced by the first pooling layer using one-dimensional convolutional kernels, and to fuse the information of the four original feature channels into 64 intermediate channels. The second convolutional layer uses 128 one-dimensional convolutional kernels, the kernel size is 7, and the activation function is ReLU. The third convolutional layer is used to capture global correlation features and fault modes of the data, based on the processing of the second pooling layer. The first pooling layer is used to reduce the dimensionality of the features extracted by the first convolutional layer. The pooling kernel size of the first pooling layer is set to 2. The second pooling layer is used to perform repeated pooling operations on the mesoscale features extracted by the second convolutional layer to further reduce dimensionality and improve the robustness of the model. The pooling kernel size of the second pooling layer is set to 2. The third pooling layer is used to calculate the average value in the angular dimension for each feature channel and output the feature vector; The first fully connected layer is used to perform comprehensive learning and nonlinear transformation on global features and to map and transform the feature vector output by the pooling layer. The activation function is ReLU. The second fully connected layer is used to optimize the compactness of feature representation. It removes redundant information and condenses key information from the high-dimensional feature vector of the first fully connected layer, providing accurate input to the output layer. The activation function is ReLU.

[0035] In this embodiment, a fault diagnosis model for an electric reciprocating pump is constructed based on real-time torque and a convolutional neural network. Preprocessed normal and abnormal sample sets are input into the network model for supervised training, with a crankshaft rotation period of 360°. Figure 2 As shown, the model includes an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer (containing 3 convolutional layers, 3 pooling layers (including 2 max pooling layers and 1 global average pooling layer), 2 fully connected layers, and 1 softmax layer for result classification). The input layer receives the device data after preprocessing in step S3, with an input shape of (3600, 4), where 3600 corresponds to the length of the torque in the angular dimension, and 4 is the number of feature channels. Three convolutional layers are configured. Convolutional layer 1 uses one-dimensional convolutional kernels to extract features from the input. It extracts local and global features of the signal using kernels of different sizes. The kernel size is 11, with 64 kernels, and the activation function is ReLU. This layer extracts fine-grained local features of the signal. Convolutional layer 2 uses 128 one-dimensional convolutional kernels with a kernel size of 7 and the activation function is ReLU. This layer further extracts more complex mesoscale features based on the dimensionality reduction features from pooling layer 1. Information from the four original feature channels is fused into the 64 intermediate channels. Convolutional layer 3 uses 256 one-dimensional convolutional kernels with a kernel size of 5 and the activation function is ReLU. This layer serves as the core feature extraction layer of the model, capturing global correlation features and fault-related patterns in the data, providing highly recognizable feature maps for subsequent fault localization in Grad-CAM. The convolution operation formula is:

[0036] in, The first term represents the output of a one-dimensional convolutional layer. The feature channel in the angular dimension The value at each position, This indicates that the input data is in the angular dimension. + -1 value at position, Indicates the first The first convolutional kernel Each weight, Indicates the first The bias of each convolution kernel This indicates the kernel size.

[0037] like Figure 2As shown, three pooling layers are set. Pooling layer 1 uses max pooling, with a kernel size of 2 between convolutional layers 1 and 2. This reduces the dimensionality of the features extracted by convolutional layer 1, decreases the model's computational parameters, and preserves the key local features of each feature channel. Pooling layer 2 also uses max pooling, with a kernel size of 2 between convolutional layers 2 and 3. Pooling layer 3 uses global average pooling, calculating the average value in the angular dimension for each feature channel and outputting the feature vector.

[0038] like Figure 2 As shown, two fully connected layers are set up, both with ReLU activation function. The feature maps output by the pooling layer are converted into vectors, and the global features are comprehensively learned and nonlinearly transformed. The output layer uses the Softmax activation function. The probability of the output device belonging to the categories of normal (0), abnormal 1 (hydraulic end discharge valve failure), abnormal 2 (hydraulic end suction valve failure), and abnormal 3 (power end failure) is given by the following formula:

[0039] in, Indicates a fault in the electric reciprocating pump. j The predicted probabilities of each category, The output of the fully connected layer represents the first... j Values ​​for each category, k This represents the category index corresponding to the fault. This indicates that the output of the fully connected layer belongs to k The original score for each category.

[0040] S5. Model Training: Training the fault diagnosis model for electric reciprocating pumps using the loss function; In this embodiment, since the conventional classification cross-entropy function is insufficient to meet the core requirements of "class imbalance and high feature discrimination", a weighted focus-center hybrid loss function is selected as the loss function for the electric reciprocating pump fault diagnosis model. Adam is used as the optimizer to train the electric reciprocating pump fault diagnosis model, which can effectively measure the difference between the predicted results of the electric reciprocating pump fault diagnosis model and the true labels, and guide the optimization of model parameters. The formula is as follows: ; ; ; in, This represents the weighted focus-center mixed loss function. This represents the weighted focus loss function. Represents the central loss function. and These represent weighting coefficients, set to 0.7 and 0.3 respectively. Indicates category weight, This represents the focusing factor, with a value of 2. Indicates the number of samples. Indicates the first i The sample belongs to the first j Real labels for each category Indicates the first i The sample belongs to the first j The predicted probabilities of each category, Indicates the first i The feature vector output by each sample through the fully connected layer 2 express The feature center of the category.

[0041] S6. Fault Location: During the training of the fault diagnosis model for the electric reciprocating pump, the output feature map is recorded. All feature channels of the feature map are weighted and summed to generate a heat map. Fault location is then performed using the heat map. The implementation method is as follows: During the training of the fault diagnosis model for the electric reciprocating pump, the backpropagation calculation of the third convolutional layer and the third pooling layer is used to output the original value without the Softmax layer, record the value and output the feature map. The feature map is weighted and summed to generate a heatmap. The data lengths on the heatmap are mapped one-to-one with the original torque curve data in the angular dimension, and the trough threshold is set through the abnormal feature waveform diagram. Continuous intervals in the heatmap that are above the trough threshold are identified as abnormal intervals. Compare the abnormal intervals with the interval divisions of the waveform diagram of torque and angle of the reciprocating pump; Based on the interval comparison results, the fault location is determined; The trained electric reciprocating pump fault diagnosis model is deployed in the reciprocating pump condition monitoring system to identify features and detect anomalies in real-time torque curve data, output fault types, and realize fault diagnosis of electric reciprocating pumps.

[0042] In this embodiment, during model training, the output feature maps are automatically recorded through convolutional layers and global pooling layers. Using Grad-CAM technology, all feature channels of the feature maps are weighted and summed, and a heatmap is generated using the ReLU activation function. The calculation formula is as follows: ; in, This indicates the heatmap in the angular dimension. The value at each position, The neuron importance weights represent the feature channels. This indicates the feature channel in the angular dimension. Feature values ​​at each position, Indicates the characteristic channel, This represents the activation function, used to filter out features that contribute negatively to the predicted category, retaining only features that contribute positively.

[0043] The generated heatmap data lengths are matched one-to-one with the original data in the angular dimension. A threshold for the trough is set using three types of abnormal feature waveforms. Continuous intervals in the heatmap exceeding the threshold are identified as abnormal intervals. Since the angle difference between each cylinder is 72°, the angle fault range for cylinder 1 is set to (65, 72°], for cylinder 2 to (137, 144°], for cylinder 3 to (209, 216°], and so on. The abnormal intervals are compared with the intervals defined in the waveform of the reciprocating pump's torque versus angle to determine which cylinder is experiencing the problem. Finally, the trained feature model is deployed to the reciprocating pump status monitoring system for real-time torque curve data feature recognition and anomaly detection, outputting the fault type and thus achieving fault diagnosis of the electric reciprocating pump.

[0044] In this embodiment, Figure 1 In this process, by collecting pulse signals from the key phase sensor, real-time torque, and discharge pump pressure, and combining them with the equipment's mechanical parameters, the real-time torque curves under normal and abnormal conditions are manually marked. After data preprocessing, the electric reciprocating pump fault diagnosis model is trained. The trained electric reciprocating pump fault diagnosis model is then deployed to the reciprocating pump condition monitoring system, and the signal to be diagnosed is acquired. The signal to be diagnosed is then preprocessed and its features are extracted. Finally, the reciprocating pump condition monitoring system with the deployed electric reciprocating pump fault diagnosis model is used to determine the fault type.

[0045] In this embodiment, a five-cylinder electrically driven fracturing pump is typically used for fracturing. During operation, the fracturing pump may experience torque fluctuations due to problems on either the power or hydraulic end, and conventional torque curve monitoring cannot reflect the specific problem. By combining the mechanical parameters of the five-cylinder fracturing pump and theoretical mechanics formulas, and taking a 360° crankshaft rotation as the cycle, the torque curve per stroke at a fixed pump pressure can be calculated. Under normal conditions, the plunger of the five-cylinder fracturing pump performs work when extended and does not perform work when retracted. When the suction valve of a cylinder on the hydraulic end fails, the corresponding plunger does not perform work during reciprocating motion, affecting the torque curve per stroke by 180°. When the discharge valve of a cylinder on the hydraulic end fails, the cylinder performs work normally when extended, but the plunger experiences a reverse force when retracted, affecting the torque curve per stroke by 180°. Based on the mechanical characteristics of the five-cylinder pump, with a 72° angle difference between each cylinder, a convolutional neural network algorithm is used to extract and model the torque curve for each stroke. Based on these curve characteristics, fault types can be identified and located. Examples include: insufficient water supply to the hydraulic end, a faulty suction valve in cylinder n, or a faulty discharge valve in cylinder n. When a problem occurs on the power end of a cylinder, the corresponding plunger must overcome friction to perform work during reciprocating motion. According to finite element analysis, the normal stress on the bearing is greatest when the crankshaft rotates to 260°, and the torque is also greatest at this point. Based on the mechanical characteristics of the five-cylinder pump, with a 72° angle difference between each cylinder, a convolutional neural network algorithm is used to extract and model the torque curve for each stroke. Based on these curve characteristics, fault types can be identified and located. For example, an abnormal operation of the power end of cylinder n.

[0046] In this embodiment, as Figure 3 As shown, the theoretical torque curve of a normal five-cylinder fracturing pump under a certain working condition was plotted using finite element analysis software, where i represents the crankshaft rotation angle and MXI represents the crankshaft torque value. This curve reflects the change in torque when the crankshaft rotates 360°.

[0047] In this embodiment, as Figure 4 As shown, the theoretical torque curve of a five-cylinder fracturing pump with a damaged discharge valve of cylinder #1 was plotted using finite element analysis software under a certain working condition. Here, i represents the crankshaft rotation angle and MXI represents the crankshaft torque value. The curve reflects the change in torque when the crankshaft rotates 360°.

[0048] In this embodiment, as Figure 5 As shown, the theoretical torque curve of a five-cylinder fracturing pump with a damaged intake valve of cylinder #1 was plotted using finite element analysis software under a certain working condition. Here, i represents the crankshaft rotation angle and MXI represents the crankshaft torque value. The curve reflects the change in torque when the crankshaft rotates 360°.

[0049] In this embodiment, as Figure 6As shown, the theoretical torque curve of a five-cylinder fracturing pump with a damaged crosshead at the power end of cylinder #1 was plotted using finite element analysis software under a certain working condition. Here, i represents the crankshaft rotation angle and MXI represents the crankshaft torque value. The curve reflects the change in torque when the crankshaft rotates 360°.

[0050] In summary, the purpose of this invention is to provide a fault diagnosis method for electric reciprocating pumps based on real-time torque. This method monitors the working status of the power and hydraulic ends in real time, and immediately locates the fault when abnormalities occur. This provides real-time guidance for operators to adjust the equipment's operating status and accurate maintenance guidance for equipment maintenance personnel. First, a key phase sensor is used to mark the crankshaft at 0°, and real-time torque, discharge pump pressure, and pump stroke are collected. Then, based on the mechanical parameters and theoretical mechanics formulas of the reciprocating pump, a theoretical torque curve is constructed using big data under different discharge pump pressures and pump strokes, with a crankshaft rotation of 360° as the cycle. Fault data from different fault types on the power and hydraulic ends are used to construct a fault diagnosis model for the electric reciprocating pump. When a problem occurs on the power or hydraulic end corresponding to a cylinder of the reciprocating pump, the fault location can be accurately determined by comparing the real-time torque curve features with the trained electric reciprocating pump fault diagnosis model. In other words, this invention processes the torque data of each stroke of the electric reciprocating pump using a deep learning algorithm, extracts feature parameters, and compares the results with the trained model to achieve accurate fault type determination for the electric reciprocating pump.

Claims

1. A real-time torque-based fault diagnosis method for an electrically-driven reciprocating pump, characterized in that, Includes the following steps: S1, Parameter Reading: Read the parameters of the reciprocating pump; S2. Data Acquisition: Collect torque curve data of the reciprocating pump under normal operating conditions as a normal sample set, and collect torque curve data of the reciprocating pump under typical fault conditions as an abnormal sample set. S3. Data Processing: Preprocess the normal and abnormal sample sets; S4. Model Establishment: Based on real-time torque and convolutional neural network, a fault diagnosis model for electric reciprocating pump is constructed. The pre-processed normal sample set and abnormal sample set are input into the electric reciprocating pump fault diagnosis model for supervised training with the crankshaft rotating 360° as the cycle. S5. Model Training: The fault diagnosis model of the electric reciprocating pump is trained using the loss function; S6. Fault Location: During the training of the fault diagnosis model for the electric reciprocating pump, the output feature map is recorded, and all feature channels of the feature map are weighted and summed to generate a heat map. The fault is located through the heat map.

2. The real-time torque-based electrically-driven reciprocating pump fault diagnosis method according to claim 1, characterized in that, S3 includes the following steps: Based on normal and abnormal sample sets, each data point is labeled at the sample level, with each data point corresponding to a signal with a 360° cycle. Based on the annotation results, before the convolution operation, the padding data at the beginning of the sequence is obtained from the end of the sequence, and the bottom padding data at the end of the sequence is obtained from the beginning of the sequence. After basic data filling, outliers were removed using the three-standard-deviation criterion. After removing outliers, the data is standardized to complete the data preprocessing.

3. The real-time torque based electric drive reciprocating pump fault diagnosis method according to claim 2, characterized in that, The standardized expression is as follows: ； in, Indicates the first The standardized values ​​of each feature dimension Indicates the first The mean of each feature dimension, Indicates the first Standard deviation of each feature dimension Indicates the first Initial signal values ​​for each feature dimension.

4. The fault diagnosis method for an electrically driven reciprocating pump based on real-time torque according to claim 1, characterized in that, The fault diagnosis model for the electrically driven reciprocating pump includes: The input layer is used to receive the preprocessed sample set data, where the input shape is (3600, 4), where 3600 corresponds to the length of the torque in the angular dimension, and 4 is the number of feature channels; Convolutional layers are used to extract features from the received sample set data to obtain local and global features. The convolutional layers include a first convolutional layer, a second convolutional layer, and a third convolutional layer. Pooling layers are used to output feature vectors based on local and global features. These pooling layers include a first pooling layer located between a first and a second convolutional layer, a second pooling layer located between a second and a third convolutional layer, and a third pooling layer. The first and second pooling layers both employ max pooling, while the third pooling layer employs global average pooling. Specifically, the first and second pooling layers perform max pooling on the local features extracted by the convolutional layers, and the third pooling layer performs global average pooling on the global features extracted by the convolutional layers. The fully connected layer is used to map and transform the feature vectors output by the pooling layer, and to learn and perform nonlinear transformations on the global features extracted by the convolutional layer; the fully connected layer includes a first fully connected layer and a second fully connected layer. The output layer is used for feature vectors, outputting the probability of fault categories in the electric reciprocating pump.

5. The fault diagnosis method for an electrically driven reciprocating pump based on real-time torque according to claim 4, characterized in that, The first convolutional layer is used to extract local and global features of the signal based on the received sample set data using one-dimensional convolutional kernels of different sizes. The first convolutional layer is set with a convolutional kernel of 64, a kernel size of 11, and an activation function of ReLU. The second convolutional layer is used to extract mesoscale features based on the features reduced by the first pooling layer using one-dimensional convolutional kernels, and to fuse the information of the four original feature channels into 64 intermediate channels. The second convolutional layer uses 128 one-dimensional convolutional kernels, the kernel size is 7, and the activation function is ReLU. The third convolutional layer is used to capture global correlation features and fault modes of the data based on the processing of the second pooling layer. The first pooling layer is used to perform dimensionality reduction on the features extracted by the first convolutional layer, and the pooling kernel size of the first pooling layer is set to 2. The second pooling layer is used to perform repeated pooling operations on the mid-scale features extracted by the second convolutional layer. The pooling kernel size of the second pooling layer is set to 2. The third pooling layer is used to calculate the average value in the angular dimension for each feature channel and output the feature vector; The first fully connected layer is used to perform comprehensive learning and nonlinear transformation on global features and to map and transform the feature vector output by the pooling layer. The activation function is ReLU. The second fully connected layer is used to optimize the compactness of feature representation. It removes redundant information and condenses key information from the high-dimensional feature vector of the first fully connected layer, providing accurate input to the output layer. The activation function is ReLU.

6. The fault diagnosis method for an electrically driven reciprocating pump based on real-time torque according to claim 4, characterized in that, The expression for the probability of the fault category of the electric reciprocating pump is as follows: ； in, Indicates a fault in the electric reciprocating pump. j The predicted probabilities of each category, The output of the fully connected layer represents the first... j Values ​​for each category, k This represents the category index corresponding to the fault. This indicates that the output of the fully connected layer belongs to k The original score for each category.

7. The fault diagnosis method for an electric reciprocating pump based on real-time torque according to claim 5, wherein the expression for the loss function is as follows: ； ； ； in, This represents the weighted focus-center mixed loss function. This represents the weighted focus loss function. Represents the central loss function. and All represent weighting coefficients. Indicates category weight, This represents the focusing factor, with a value of 2. Indicates the number of samples. Indicates the first i The sample belongs to the first j Real labels for each category Indicates the first i The sample belongs to the first j The predicted probabilities of each category, Indicates the first i The feature vector output by the second fully connected layer of each sample express The feature center of the category.

8. The fault diagnosis method for an electrically driven reciprocating pump based on real-time torque according to claim 4, characterized in that, S6 includes the following steps: During the training of the fault diagnosis model for the electric reciprocating pump, the backpropagation calculation of the third convolutional layer and the third pooling layer is used to output the original value without the Softmax layer, record the value and output the feature map. The feature map is weighted and summed to generate a heatmap. The data lengths on the heatmap are mapped one-to-one with the original torque curve data in the angular dimension, and the trough threshold is set through the abnormal feature waveform diagram. Continuous intervals in the heatmap that are above the trough threshold are identified as abnormal intervals. Compare the abnormal intervals with the interval divisions of the waveform diagram of torque and angle of the reciprocating pump; Based on the interval comparison results, the fault location is determined; The trained electric reciprocating pump fault diagnosis model is deployed in the reciprocating pump condition monitoring system to identify features and detect anomalies in real-time torque curve data, output fault types, and realize fault diagnosis of electric reciprocating pumps.

9. The fault diagnosis method for an electrically driven reciprocating pump based on real-time torque according to claim 1, characterized in that, The expression for the heatmap is as follows: ； in, This indicates the heatmap in the angular dimension. The value at each position, The neuron importance weights represent the feature channels. This indicates the feature channel in the angular dimension. Feature values ​​at each position, Indicates the characteristic channel, This represents the activation function.

Images