A Fault Diagnosis Method and System for Submersible Electric Pumps Based on CFD and Deep Learning

Abstract

This invention relates to the field of submersible electric pump (FET) fault diagnosis technology, and particularly to a method and system for FET fault diagnosis based on CFD and deep learning. The method includes constructing a full-channel 3D model, simulating FET faults by modifying key geometric feature parameters, and using a hybrid mesh strategy to divide the mesh; extracting time-domain and frequency-domain features from the simulation data, forming feature vectors from these features, and constructing multi-dimensional labeled simulation data using these feature vectors and multi-dimensional label vectors; constructing a multi-channel deep neural network, training the multi-channel deep neural network using the multi-dimensional labeled simulation data; and deploying the trained multi-channel deep neural network to a well site edge computing gateway to diagnose FET sensor data. This invention utilizes CFD to generate high-quality data to drive deep learning model training, achieving rapid and accurate collaborative diagnosis of multiple faults in FETs.

Description

Technical Field

[0001] This invention relates to the field of submersible electric pump fault diagnosis technology, and in particular to a submersible electric pump fault diagnosis method and system based on CFD and deep learning. Background Technology

[0002] Submersible electric pumps (ESPs) operate for extended periods in harsh downhole environments characterized by high temperatures, high pressures, and corrosive media. These environments lead to frequent failures exhibiting significant synergistic effects from multiple faults. Existing data-driven diagnostic methods require vast amounts of fault data, while field fault samples are scarce, with a high volume of single-fault data and very little data on multi-fault coupling. Diagnostic methods based on traditional mechanistic models lack sensitivity to multi-fault coupling and early, subtle faults, rely heavily on expert experience, and are difficult to automate. Existing methods often rely on single signals and lack the synergistic utilization of multi-source physical information, making accurate separation and diagnosis of complex faults challenging.

[0003] Computational fluid dynamics (CFD) simulations can accurately simulate complex flow fields and fault states, but they are computationally expensive and time-consuming, making them unsuitable for direct real-time diagnostics. Deep learning excels at mining complex patterns and learning fault characteristics from massive amounts of data, but it lacks physical mechanism constraints, is prone to overfitting in small sample scenarios, and has poor model interpretability. Summary of the Invention

[0004] This invention aims to at least solve one of the technical problems existing in related technologies. To this end, this invention provides a method and system for fault diagnosis of submersible electric pumps based on CFD and deep learning. It utilizes CFD to generate high-quality data, drives the training of a deep learning model, and ultimately achieves fast and accurate collaborative diagnosis of multiple faults in submersible electric pumps.

[0005] This invention provides a fault diagnosis method for submersible electric pumps based on CFD and deep learning, comprising:

[0006] S1: Construct a full-flow-channel 3D model based on the actual dimensions of the submersible electric pump and define key geometric feature parameters. Simulate submersible electric pump failure by modifying the key geometric feature parameters. S2: Based on the full-channel 3D model, a hybrid mesh strategy is used to divide the impeller rotating domain and the stationary pump casing domain into meshes; S3: Based on the mesh, a transient solver is set up using the three-dimensional transient RANS equations based on the finite volume method. The Euler-Euler Mixture model is used to simulate airlock failure, the DPM coupled Oka erosion model is used to simulate erosion wear, the SSTk-ω model is used to simulate turbulence, and the Coupled algorithm and the second-order implicit transient scheme are used to perform CFD simulation to obtain simulation data. S4: Extract the time-domain and frequency-domain features of the simulation data and define a multi-dimensional parameterized label vector for fault combinations. Combine the time-domain and frequency-domain features into a feature vector, and construct simulation data with multi-dimensional labels using the feature vector and the multi-dimensional label vector. S5: Construct a multi-channel deep neural network and train it using simulation data with multi-dimensional labels; S6: Deploy the trained multi-channel deep neural network to the well site edge computing gateway to diagnose the submersible pump sensor data.

[0007] Furthermore, the key geometric feature parameters include impeller blade outlet width, impeller outlet angle, blade tip-pump casing clearance, and blade thickness; Simulating submersible electric pump failures by modifying key geometric feature parameters includes: Wear failure was simulated by reducing blade thickness and impeller blade exit width; Increased clearance is simulated by increasing the fit clearance between the blade tip and the pump casing; Simulate leaf breakage by directly removing part of the leaf using Boolean operations; Airlock can be simulated by increasing the gas volume fraction at the inlet. Scale deposition was simulated by thickening the pressure surface of the blades.

[0008] Furthermore, the impeller rotation domain is divided using an unstructured tetrahedral core mesh with additional multi-layer prism meshes; A structured mesh is used to partition the static pump casing region.

[0009] Furthermore, step S3 also includes monitoring and sampling the simulation data; Multiple monitoring points are set at the pump inlet, pump outlet, and between each stage of impeller and guide wheel; Define the entire surface of the impeller or guide wheel as the force monitoring surface; Monitor fluid torque fluctuations in the impeller.

[0010] Furthermore, step S4 includes: S41: Define a multi-dimensional label vector based on specific fault combinations and operating conditions; S42: Extract time-domain features from simulation data to obtain time-domain features; S43: Perform a fast Fourier transform on the simulation data to obtain a spectrum, and extract frequency domain features from the spectrum to obtain frequency domain features; S44: Combine time-domain features and frequency-domain features into a feature vector, and construct simulation data with multi-dimensional labels based on the feature vector and the multi-dimensional label vector.

[0011] Furthermore, in step S42, the time-domain features include root mean square, kurtosis, and peak factor.

[0012] Furthermore, in step S43, the frequency domain feature is used to determine the relationship between the shaft rotation frequency, blade frequency and their harmonics and the physical structure of the equipment.

[0013] Furthermore, the multi-channel deep neural network includes parallel feature extraction channels, feature splicing layers, attention fusion layers, and multi-task output layers.

[0014] Furthermore, each channel of the parallel feature extraction channel consists of stacked one-dimensional convolutional layers and a bidirectional long short-term memory network; The multi-channel features extracted by the parallel feature extraction channels are concatenated by a feature concatenation layer to obtain a comprehensive feature vector; The attention fusion layer linearly transforms the comprehensive feature vector into query, key, and value through a learnable weight matrix, calculates the correlation between the query and the key, normalizes it into a weight score using the Softmax function, and generates fused features by weighting and summing the values ​​using the weight scores. The multi-task output layer includes a classification branch and a regression branch. The classification branch uses the Sigmoid activation function and outputs the probability of multiple faults to achieve multi-label classification. The regression branch uses the Linear activation function and outputs the severity estimate of each fault.

[0015] This invention also provides a fault diagnosis system for submersible electric pumps based on CFD and deep learning, for executing the aforementioned fault diagnosis method for submersible electric pumps based on CFD and deep learning, including: The end-side acquisition module includes a submersible electric pump unit, a multi-source sensor, and a data acquisition unit. The multi-source sensor is used to collect the vibration, current, pressure, and temperature of the submersible electric pump, and the data acquisition unit converts the vibration, current, pressure, and temperature of the submersible electric pump into a standardized data stream. The edge-side diagnostic module includes an edge computing gateway, a lightweight diagnostic model, and a local early warning submodule. The edge computing gateway receives and preprocesses standardized data streams. The lightweight diagnostic model is a multi-channel deep neural network pre-trained by CFD simulation and fine-tuned by on-site data, used to complete fault type identification and severity quantification in real time. The local early warning submodule realizes local millisecond-level alarm and control for faults. The cloud-side optimization module includes a cloud gateway and a visualization platform. The cloud gateway enables data interaction between the edge-side diagnostic module and the cloud-side optimization module, uploading fault type identification and severity quantification to the cloud center. Through in-depth analysis and cross-validation using historical data and expert knowledge base, a comprehensive decision report is generated, including fault root causes, evolution trends, maintenance priorities, and maintenance recommendations. This report is then pushed to maintenance personnel through the visualization platform. The actual results verified through on-site inspection are fed back as new labeled data to the edge-side diagnostic module for incremental learning and optimization of the model.

[0016] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: By introducing high-fidelity computational fluid dynamics simulation technology, single and complex fault states, including impeller wear, guide wheel erosion, and airlock effects, are accurately reproduced in virtual space, generating a massive standardized training dataset covering all operating conditions. This overcomes the industry bottleneck of high cost and long cycle time for acquiring field fault data. Based on this, a multi-channel deep neural network is built that can process multi-source heterogeneous sensor signals such as vibration, current, and pressure in parallel. Through adaptive weighted fusion at the feature level, it achieves accurate identification and quantitative assessment of the severity of multiple faults, effectively solving the problems of misjudgment and missed diagnosis in complex fault scenarios using traditional diagnostic methods.

[0017] To further enhance the system's practical engineering value, a highly efficient cloud-edge-device collaborative architecture was constructed. A lightweight diagnostic model, pre-trained with simulation data and fine-tuned with field data, was deployed on edge computing nodes to ensure real-time diagnosis and millisecond-level response capabilities for critical faults. Simultaneously, a cloud platform enables centralized monitoring of all equipment status, continuous optimization of the diagnostic model, and dynamic updates of the digital twin. This not only transforms the operation and maintenance model from "reactive maintenance" to "predictive maintenance," but also fosters a self-evolving and continuously optimizing intelligent diagnostic ecosystem through continuous interaction between field data and simulation models. Ultimately, this provides oilfield production with a reliable, efficient, and continuously learning intelligent operation and maintenance solution, significantly improving equipment reliability, reducing unplanned downtime risks, and generating substantial economic benefits.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a fault diagnosis method for submersible electric pumps based on CFD and deep learning provided by the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of a submersible electric pump fault diagnosis system based on CFD and deep learning provided by the present invention.

[0022] Figure label: 101. End-side acquisition module; 102. Side-side diagnostic module; 103. Cloud-side optimization module. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0024] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0025] The following is combined Figures 1 to 2 This invention describes a fault diagnosis method and system for submersible electric pumps based on CFD and deep learning.

[0026] like Figure 1 As shown, a fault diagnosis method for submersible electric pumps based on CFD and deep learning includes: S1: Construct a full-flow-channel 3D model based on the actual dimensions of the submersible electric pump and define key geometric feature parameters. Simulate submersible electric pump failure by modifying the key geometric feature parameters. Based on the actual dimensions of the submersible electric pump, a complete three-dimensional model of the entire flow channel, including the first stage, intermediate stage, final stage impellers, pump casing, and diffuser, was created using the SpaceClaim tool.

[0027] Key geometric parameters include impeller blade exit width Impeller outlet angle The clearance between the impeller tip and the pump casing and blade thickness .

[0028] Simulate submersible electric pump failure by modifying key geometric feature parameters; By reducing the blade thickness and impeller blade outlet width Simulate wear and tear failure; By increasing the clearance between the blade tip and the pump casing The simulated gap increased (uneven wear); Simulate leaf breakage by directly removing part of the leaf using Boolean operations; Airlock can be simulated by increasing the gas volume fraction at the inlet. Scale deposition is simulated by thickening the pressure surface of the blades.

[0029] In some specific embodiments of the present invention, a single-stage impeller-guide wheel assembly of a submersible electric pump is taken as the research object. Its main design parameters are: flow rate Q is 200 m³ / d, head H is 500 m, and rotational speed n is 2850 rpm.

[0030] Using ANSYS SpaceClaim software, a complete three-dimensional model of the entire flow channel, including the first stage, intermediate stage, final stage impellers, pump casing, and diffuser, was established based on the actual dimensions of the submersible electric pump.

[0031] The healthy size is 5mm; set up Simulate moderate wear, set Simulates mild wear; Inlet gas volume fraction The health status is 0; set up Simulate mild airlock, set Simulates moderate airlock; generate[ mm, ]as well as[ , [Cooperative fault status, etc.]

[0032] S2: Based on the full-channel 3D model, a hybrid mesh strategy is used to divide the impeller rotating domain and the stationary pump casing domain into meshes; For complex impeller rotation domains, an unstructured tetrahedral core mesh is used, supplemented with multiple layers of prism meshes to accurately capture near-wall flow.

[0033] A structured grid is used for the static pump casing domain to improve computational efficiency.

[0034] Three schemes with increasing grid numbers were set up. When the rate of change of key performance parameters was less than 2%, grid independence was considered to be valid. Finally, the scheme that balanced computational accuracy and resource consumption was selected.

[0035] In some specific embodiments of the present invention, ANSYS Meshing is used to generate hybrid meshes.

[0036] Ten boundary layer meshes are generated near the wall, ensuring that the y+ value of the first mesh layer is between 30 and 50 to meet the requirements of the SSTk-ω turbulence model.

[0037] Mesh independence verification was performed, and the final generated mesh totaled approximately 9 million cells. At this number, the rate of change of the calculated head value was less than 1.5%, which was considered mesh independent.

[0038] S3: Based on the mesh, a transient solver is set up using the three-dimensional transient RANS equations based on the finite volume method. The Euler-Euler Mixture model is used to simulate airlock failure, the DPM coupled Oka erosion model is used to simulate erosion wear, the SSTk-ω model is used to simulate turbulence, and the Coupled algorithm and the second-order implicit transient scheme are used to perform CFD simulation to obtain simulation data. The transient solver is configured using the three-dimensional transient RANS equations based on the finite volume method. The continuity equation is:

[0039] in, For density, For time, For gradient operators, It is a velocity vector; Momentum equation:

[0040] in, For pressure, For stress tensor, It is the acceleration due to gravity. It is an external volume force.

[0041] Multiphase flow model: For airlock faults: the Euler-Euler Mixture model is used to simulate gas-liquid two-phase flow, and different degrees of airlock are simulated by setting different gas phase volume fractions at the inlet.

[0042] For erosion wear: the discrete phase model (DPM) is used to simulate the motion trajectory of solid particles, and the Oka erosion model is coupled to calculate the wall wear rate.

[0043] Turbulence model: The SST k-ω model is used to simulate the turbulence model. This model can accurately predict flow separation under adverse pressure gradients and is suitable for simulating complex flows inside impellers.

[0044] Transient settings: Solution method: The Coupled algorithm and the second-order implicit transient scheme are used.

[0045] Time step: Set to the time required for the impeller to rotate 1°. ,Right now seconds, of which The value is the rotational speed.

[0046] Total computation time: At least 10 complete impeller rotation cycles must be calculated to ensure that the flow develops fully and to obtain a stable periodic signal.

[0047] To achieve integration with deep learning diagnostics, the simulation needs to output a time-domain signal with the same data format as the physical sensor data.

[0048] Set up pressure pulsation monitoring points, force monitoring surfaces, and torque monitoring.

[0049] Pressure pulsation monitoring points: Multiple monitoring points are set at the pump inlet, pump outlet, and between each stage of impeller and guide wheel.

[0050] Force monitoring surface: Define the entire impeller or guide wheel surface as the force monitoring surface, used to calculate radial force. and axial force Fluctuations.

[0051] Torque monitoring: Monitors the fluctuation of fluid torque T in the impeller.

[0052] Set data sampling frequency for Hz, synchronized with the time step, generates a high sampling rate time-domain signal.

[0053] Pressure pulsation at monitoring points Radial force of the impeller and torque As the primary output.

[0054] Post-processing can convert radial force fluctuations into vibration acceleration signals. ,in To achieve equivalent mass, torque fluctuations are converted into harmonics of the current signal.

[0055] S4: Extract the time-domain and frequency-domain features of the simulation data and define a multi-dimensional parameterized label vector for fault combinations. Combine the time-domain and frequency-domain features into a feature vector, and construct simulation data with multi-dimensional labels using the feature vector and the multi-dimensional label vector. S41: Define a multi-dimensional label vector based on specific fault combinations and operating conditions; Before the simulation, a unique multidimensional label vector is defined for each calculation. For example, the label [wear level: 0.7, airlock level: 0.2, flow rate: 100%] accurately describes the specific fault combination and operating condition simulated in this simulation.

[0056] Using this set of parameters, a high-fidelity CFD transient simulation was run, and raw high-sampling-rate time-series signals from multiple channels, including pressure, vibration, and torque, were simultaneously acquired. The obtained Fangzheng data is a time-series signal.

[0057] S42: Extract time-domain features from simulation data to obtain time-domain features; The root mean square (RMS), kurtosis, and crease factor are extracted from the time-domain signal. RMS reflects the overall energy level of the signal; kurtosis is highly sensitive to early impulse faults; and the crease factor is used to identify the difference between transient peaks and background noise. These features directly influence the time-series signal. Select The calculation is performed on each data point without any transformation; the calculation expression is as follows:

[0058]

[0059]

[0060] in, It is the root mean square. The number of data points, For the first A timing signal For ravine, The mean of the time series signal. As the peak factor, To find the function with the maximum value; S43: Perform a fast Fourier transform on the simulation data to obtain a spectrum, and extract frequency domain features from the spectrum to obtain frequency domain features; For frequency domain feature extraction, this process requires processing the time domain signal. Perform a Fast Fourier Transform (FFT) to obtain the spectrum. The time-domain signal is converted to the frequency domain, and then the characteristic frequency points determined by the equipment structure (such as shaft rotation frequency, blade throughput frequency, bearing failure frequency) are accurately located, and their corresponding amplitudes are extracted.

[0061] S44: Combine time-domain features and frequency-domain features into a feature vector, and construct simulation data with multi-dimensional labels based on the feature vector and the multi-dimensional label vector.

[0062] For example, [RMS vibration = 4.2, kurtosis vibration = 5.1, peak factor current = 3.8, rotational frequency amplitude = 0.15, blade frequency amplitude = 0.07, ...]. This, together with the defined specific fault combinations and the "multi-dimensional label vector" of operating conditions, constitutes a complete causal sample, thus realizing a specific one-to-one correspondence between physical faults and mathematical conditions. This pairing relationship explicitly tells the model: when the data exhibits "this combination of numerical characteristics," the underlying physical cause is "this type and degree of fault."

[0063] Each data file generated from a fault combination and operating condition simulation case has a unique and precise multidimensional label. All data is stored in the standard .h5 format, constructing a standardized dataset with a one-to-one correspondence between cause and effect. The construction of the labeled simulation dataset is a systematic process that ensures that each data sample precisely corresponds to the physical conditions and fault states that caused it.

[0064] All sample pairs are systematically stored in HDF5 (.h5) files. This format can efficiently store large-scale scientific data and naturally maintains the tree-like structure of data and metadata (tags), ensuring traceability from cause (simulation parameters) to result (output signal).

[0065] S4: Construct a multi-channel deep neural network (MC-CNN-BiLSTM-ATT) and train the multi-channel deep neural network using simulation data with multi-dimensional labels; Multichannel deep neural networks include parallel feature extraction channels, feature splicing layers, attention fusion layers, and multi-task output layers.

[0066] Parallel feature extraction channels: Each channel consists of stacked one-dimensional convolutional layers (1D-CNN) and bidirectional long short-term memory networks (Bi-LSTM). The 1D-CNN is responsible for extracting local features in the signal (such as vibration and shock, current harmonics), while the Bi-LSTM captures long-range temporal contextual dependencies in the signal.

[0067] The multi-channel features extracted by the parallel feature extraction channels are concatenated by a feature concatenation layer to obtain a comprehensive feature vector.

[0068] Attention Feature Fusion Layer: The core of the attention feature fusion layer is to enable the model to automatically learn and weigh features from different sensor channels such as vibration, current, and pressure to determine the importance of a specific fault. Its fusion process is a dynamic and intelligent weighted summation, rather than a simple concatenation or averaging. Specifically, the attention fusion layer linearly transforms the comprehensive feature vector into query, key, and value using a learnable weight matrix. It calculates the correlation between the query and key and normalizes it to weight scores using the Softmax function. These weight scores directly reflect the contribution of each feature to the current diagnostic task. The weighted summation of the values ​​is then used to generate the fused features.

[0069] The multi-task output layer includes a classification branch and a regression branch. The network ends in two branches: the classification branch uses the Sigmoid activation function and outputs the probability of multiple faults to achieve multi-label classification; the regression branch uses the Linear activation function and outputs the severity estimate of each fault.

[0070] The MC-CNN-BiLSTM-ATT model was pre-trained on a large-scale simulation dataset with multidimensional labels. A small amount of normal and fault data from the field was collected. The main structure of the network remained unchanged, and only the last few layers were fine-tuned to enable the model to quickly adapt to the noise characteristics and operating conditions of real sensors, thus completing the model training.

[0071] S6: Deploy the trained multi-channel deep neural network to the well site edge computing gateway to diagnose the submersible pump sensor data.

[0072] The trained multi-channel deep neural network is quantized and accelerated using tools such as TensorRT or ONNX, and then deployed to the well site edge computing gateway to achieve low-latency, high-throughput analysis of real-time data streams.

[0073] Data from the submersible pump sensors in the field is crucial for driving this diagnostic model from ideal simulation to industrial application. It fine-tunes the pre-trained model through transfer learning, enabling it to adapt to noise and individual differences in real-world operating conditions, thus improving the model's robustness and generalization ability. Simultaneously, this real-world data serves as the ultimate benchmark for verifying and iteratively optimizing the CFD simulation model, and can also discover unknown fault modes to expand the training sample library. Ultimately, the continuous influx of field data forms a digital twin closed loop of "perception-diagnosis-optimization," giving the entire diagnostic system the vitality of online learning and continuous self-evolution, ensuring the long-term accuracy and reliability of the technology in practical applications.

[0074] Applying a pre-trained model for intelligent diagnosis and decision-making constitutes an automated closed-loop process from data to action. Its core methodology can be summarized as "real-time edge perception - cloud-based collaborative analysis - human-machine interactive decision-making." Specifically, a lightweight diagnostic model deployed in edge computing equipment at the well site receives and preprocesses multi-source sensor data (vibration, current, pressure, etc.) from the electric pump in real time, instantly completing fault type identification and severity quantification inference calculations, achieving millisecond-level response and local emergency warnings. Simultaneously, diagnostic results and feature data are uploaded to the cloud center for in-depth analysis and cross-validation within a larger historical data and expert knowledge base, generating a comprehensive decision report including fault root causes, evolution trends, maintenance priorities, and maintenance recommendations, which is then pushed to maintenance personnel through a visualization platform. Finally, the real results verified through on-site inspections are fed back to the system as new labeled data for incremental learning and optimization of the model, thus completing a full intelligent closed loop of "monitoring-diagnosis-decision-verification-evolution," driving a fundamental shift in the maintenance model from passive response to proactive prediction.

[0075] This invention utilizes CFD simulation to generate a large amount of precisely labeled multi-fault coupled data, providing a high-quality training dataset for deep learning and fundamentally solving the data problem.

[0076] Multi-channel deep neural networks can decouple interfering fault features, enabling accurate identification and quantitative assessment of coexisting faults. Deep learning models are extremely sensitive to subtle features, allowing them to detect early-stage faults that are invisible to the human eye. Through an end-to-end automated diagnostic process, reliance on expert experience is significantly reduced, providing key technical support for predictive maintenance.

[0077] like Figure 2 As shown, a fault diagnosis system for submersible electric pumps based on CFD and deep learning is used to execute the aforementioned fault diagnosis method for submersible electric pumps based on CFD and deep learning, including: The end-side acquisition module 101 includes a submersible electric pump unit, a multi-source sensor, and a data acquisition unit. The multi-source sensor is used to collect the vibration, current, pressure, and temperature of the submersible electric pump, and the data acquisition unit converts the vibration, current, pressure, and temperature of the submersible electric pump into a standardized data stream. The edge-side diagnostic module 102 includes an edge computing gateway, a lightweight diagnostic model, and a local early warning submodule. The edge computing gateway receives and preprocesses standardized data streams. The lightweight diagnostic model is a multi-channel deep neural network pre-trained by CFD simulation and fine-tuned by field data, used to complete fault type identification and severity quantification in real time. The local early warning submodule realizes local millisecond-level alarm and control for faults. The cloud-side optimization module 103 includes a cloud gateway and a visualization platform. The cloud gateway enables data interaction between the edge-side diagnostic module and the cloud-side optimization module, uploads fault type identification and severity quantification to the cloud center, performs in-depth analysis and cross-validation through historical data and expert knowledge base, and generates a comprehensive decision report containing fault root causes, evolution trends, maintenance priorities and maintenance suggestions. The report is then pushed to maintenance personnel through the visualization platform. The actual results verified by on-site inspection are fed back to the edge-side diagnostic module as new labeled data for incremental learning and optimization of the model.

[0078] Through the collaborative work of the above modules, high-quality data is generated using CFD to drive the training of deep learning models, ultimately achieving fast and accurate collaborative diagnosis of multiple faults in submersible electric pumps.

[0079] Example: S1: Construct a full-flow-channel 3D model based on the actual dimensions of the submersible electric pump and define key geometric feature parameters. Simulate submersible electric pump failure by modifying the key geometric feature parameters. Model object: The single-stage impeller-guide wheel assembly of a certain type of submersible electric pump is taken as the research object. Its main design parameters are: flow rate of 200 m³ / d, head of 500 m, and speed of 2850 rpm.

[0080] Parametric modeling: Using ANSYS SpaceClaim software, a full-channel 3D model was created based on the original drawings.

[0081] Parameterize the key fault characteristics: Blade exit thickness (5 mm in healthy condition) : Gas volume fraction at the inlet (zero in healthy state) By modifying the parameters, a series of fault geometries are generated: Wear and tear fault: settings Simulate moderate wear, set Simulates mild wear.

[0082] Airlock malfunction: settings and Simulates mild and moderate airlock.

[0083] Cooperative faults: Combining the above parameters, generate [ , ], [ , [Cooperative fault status, etc.]

[0084] S2: Based on the full-channel 3D model, a hybrid mesh strategy is used to divide the impeller rotating domain and the stationary pump casing domain into meshes; Use ANSYS Meshing to generate hybrid meshes.

[0085] Ten boundary layer meshes are generated near the wall, ensuring that the y+ value of the first mesh layer is between 30 and 50 to meet the requirements of the SSTk-ω turbulence model.

[0086] Mesh independence verification was performed: the final generated mesh has a total of approximately 9 million cells. At this number, the rate of change of the calculated head is less than 1.5%, which is considered to be mesh independent.

[0087] S3: Based on the mesh, a transient solver is set up using the three-dimensional transient RANS equations based on the finite volume method. The Euler-Euler Mixture model is used to simulate airlock failure, the DPM coupled Oka erosion model is used to simulate erosion wear, the SSTk-ω model is used to simulate turbulence, and the Coupled algorithm and the second-order implicit transient scheme are used to perform CFD simulation to obtain simulation data. Solver settings: The ANSYS Fluent transient solver was used.

[0088] Multiphase flow model: The Eulerian Mixture model was selected to simulate gas-liquid two-phase flow.

[0089] Turbulence model: Select the SST k-ω model.

[0090] Operating conditions: The liquid phase is set to water, the gas phase to air, and the main phase to water.

[0091] Boundary conditions: Inlet: Mass-Flow Inlet, which sets the respective mass flow rates of water and air based on flow rate and air content.

[0092] Outlet: Pressure Outlet, the outlet static pressure is set according to the head.

[0093] Wall: No-slip boundary conditions.

[0094] Rotating Domain: The impeller region is set as a rotating fluid domain with a rotational speed of 2850 rpm, and the dynamic-static interface is handled using MRF (Multiple Reference Frames Model).

[0095] Transient calculation settings: Time step: Set to the time required for the impeller to rotate 1°, i.e. This setting accurately captures the unsteady effects caused by impeller rotation.

[0096] Total steps: 15 complete impeller rotation cycles are calculated, with the first 5 cycles used for flow initialization and the data from the last 10 cycles used for analysis.

[0097] Monitoring points: Pressure monitoring points are set at the guide wheel throat and pump outlet flange; the entire impeller wall is set as a force monitoring surface for calculating radial force.

[0098] Computation execution: All fault combinations are computed in parallel on a high-performance computing cluster (HPC). Each set of faults requires approximately 48-72 hours of computation.

[0099] S4: Extract the time-domain and frequency-domain features of the simulation data, combine the time-domain and frequency-domain features into a feature vector, and construct simulation data with multi-dimensional labels using the feature vector and the multi-dimensional label vector; Signal extraction: Export pressure pulsation signals from monitoring points using Fluent. radial force component of the impeller , .

[0100] radial force It can be used to reflect vibration signals.

[0101] Impeller fluid torque The fluctuations can be obtained through the monitoring panel, and they are proportional to the harmonic components of the motor current.

[0102] Feature engineering: right , , The three sets of signals were preprocessed (mean removal and standardization).

[0103] Extract the following features for each sample (10 cycles of data under one operating condition): Temporal characteristics: Root mean square (RMS), kurtosis, and crease factor.

[0104] Frequency domain characteristics: Perform FFT transformation to extract the 1x transition frequency, leaf pass frequency (BPF) and the amplitudes of their 2x and 3x harmonics.

[0105] Ultimately, each sample constitutes a 9-dimensional feature vector (3 signals × 3 features).

[0106] Building the dataset: This invention calculates a total of 7 states: 5 states (healthy, mild wear, moderate wear, mild airlock, and moderate airlock) and 2 cooperative fault states.

[0107] Each state contains 3 different flow points (80%, 100%, 120% of rated flow).

[0108] The final dataset contains 7 states × 3 flow points = 21 samples.

[0109] Each sample is labeled with a 4-dimensional vector: [whether wear exists, degree of wear, whether airlock exists, degree of airlock]. For example, [1, 0.6, 1, 0.2] represents moderate wear and mild airlock.

[0110] S5: Construct a multi-channel deep neural network and train it using simulation data with multi-dimensional labels; Model building: The MC-CNN-BiLSTM-ATT model was built using Python and the TensorFlow / Keras framework.

[0111] Input layer: Receives (3,300) input data (3 signal channels, 300 time steps per channel).

[0112] CNN layer: Each channel uses 2 layers of 1D-CNN, with 64 and 128 filters respectively, a kernel size of 3, and the activation function is ReLU.

[0113] Bi-LSTM layer: Each channel is connected to a Bi-LSTM layer containing 50 units.

[0114] Attention Fusion Layer: Feature fusion is performed using a 4-head self-attention mechanism layer.

[0115] Multi-task output layer: Employs multi-task learning.

[0116] Task 1 (Classification): Use the Sigmoid activation function to output two nodes, representing the probabilities of "wear and tear" and "airlock" respectively.

[0117] Task 2 (Regression): Use the Linear activation function to output two nodes, representing the severity of wear and airlock (between 0 and 1).

[0118] Model training: The 21 samples were randomly divided into a training set and a test set in a 7:3 ratio.

[0119] Loss function: Weighted loss is used, and the calculation expression is as follows:

[0120] in, For the total loss, The weighting coefficients for the binary classification cross-entropy loss are... The weighting coefficient for mean squared error loss is... For binary classification, cross-entropy loss, For mean square error loss, , .

[0121] Optimizer: Using Adam with an initial learning rate of 0.001.

[0122] The training was performed for 200 cycles, and an early stopping method was used to prevent overfitting.

[0123] Model validation: On the test set, the model achieved a classification accuracy of 98.5% for fault types.

[0124] The mean absolute error (MAE) for predicting the severity of the fault is less than 6%.

[0125] By visualizing the attention weights, it was found that the model pays more attention to vibration signals when diagnosing wear. High-frequency components; when diagnosing airlock, pay more attention to pressure signals. The low-frequency pulsations are consistent with the physical mechanism, enhancing the interpretability of the model.

[0126] S6: Deploy the trained multi-channel deep neural network to the well site edge computing gateway to diagnose the submersible pump sensor data.

[0127] The trained model is converted into a lightweight format using TensorFlow Lite and deployed to the existing edge computing gateway (such as Huawei Atlas500) at the well site.

[0128] Write a data interface program to receive vibration, current, and pressure sensor signals from the PLC in real time.

[0129] The system continuously collects data in the form of a sliding window and performs inference every 5 seconds.

[0130] The diagnostic results (fault type, severity, and confidence level) are displayed in real time on the local HMI screen at the well site and uploaded to the cloud database via a 4G network.

[0131] The system was deployed and operated on an oilfield well. Three months later, the system continued to alarm, displaying "moderate wear (0.65) with slight gas lock (0.18)".

[0132] The maintenance team planned an overhaul based on this information. Upon inspection after starting the pump, it was found that the first-stage impeller did indeed exhibit significant erosion wear, highly consistent with the severity predicted by the model.

[0133] This successful early warning prevented unplanned downtime and verified the effectiveness and engineering practical value of the invention.

[0134] The actual fault data (sensor data and maintenance records) will be transmitted back to the cloud platform.

[0135] The existing CFD simulation model is calibrated using new data (e.g., wear model parameters are corrected), and the diagnostic model is incrementally trained using the expanded dataset.

[0136] The new, performance-optimized model version is then redeployed to the edge gateway, completing a self-evolution of diagnostic capabilities and providing more targeted diagnostic services for the well.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method for submersible electric pumps based on CFD and deep learning, characterized in that, include: S1: Construct a full-flow-channel 3D model based on the actual dimensions of the submersible electric pump and define key geometric feature parameters. Simulate submersible electric pump failure by modifying the key geometric feature parameters. S2: Based on the full-channel 3D model, a hybrid mesh strategy is used to divide the impeller rotating domain and the stationary pump casing domain into meshes; S3: Based on the mesh, a transient solver is set up using the three-dimensional transient RANS equations based on the finite volume method. The Euler-Euler Mixture model is used to simulate airlock failure, the DPM coupled Oka erosion model is used to simulate erosion wear, the SST k-ω model is used to simulate turbulence, and the Coupled algorithm and the second-order implicit transient scheme are used to perform CFD simulation to obtain simulation data. S4: Extract the time-domain and frequency-domain features of the simulation data and define a multi-dimensional parameterized label vector for fault combinations. Combine the time-domain and frequency-domain features into a feature vector, and construct simulation data with multi-dimensional labels using the feature vector and the multi-dimensional label vector. S5: Construct a multi-channel deep neural network and train it using simulation data with multi-dimensional labels; S6: Deploy the trained multi-channel deep neural network to the well site edge computing gateway to diagnose the submersible pump sensor data.

2. The method for fault diagnosis of submersible electric pumps based on CFD and deep learning according to claim 1, characterized in that, The key geometric feature parameters include impeller blade outlet width, impeller outlet angle, blade tip-pump casing clearance, and blade thickness. Simulating submersible electric pump failures by modifying key geometric feature parameters includes: Wear failure was simulated by reducing blade thickness and impeller blade exit width; Increased clearance is simulated by increasing the fit clearance between the blade tip and the pump casing; Simulate leaf breakage by directly removing part of the leaf using Boolean operations; Airlock can be simulated by increasing the gas volume fraction at the inlet. Scale deposition was simulated by thickening the pressure surface of the blades.

3. The method for fault diagnosis of submersible electric pumps based on CFD and deep learning according to claim 1, characterized in that, The impeller rotation domain is divided using an unstructured tetrahedral core mesh with additional multi-layer prismatic meshes. A structured mesh is used to partition the static pump casing region.

4. The method for fault diagnosis of submersible electric pumps based on CFD and deep learning according to claim 1, characterized in that, Step S3 also includes monitoring and sampling the simulation data; Multiple monitoring points are set at the pump inlet, pump outlet, and between each stage of impeller and guide wheel; Define the entire surface of the impeller or guide wheel as the force monitoring surface; Monitor fluid torque fluctuations in the impeller.

5. The method for fault diagnosis of submersible electric pumps based on CFD and deep learning according to claim 1, characterized in that, Step S4 includes: S41: Define a multi-dimensional label vector based on specific fault combinations and operating conditions; S42: Extract time-domain features from simulation data to obtain time-domain features; S43: Perform a fast Fourier transform on the simulation data to obtain a spectrum, and extract frequency domain features from the spectrum to obtain frequency domain features; S44: Combine time-domain features and frequency-domain features into a feature vector, and construct simulation data with multi-dimensional labels based on the feature vector and the multi-dimensional label vector.

6. The method for fault diagnosis of submersible electric pumps based on CFD and deep learning according to claim 5, characterized in that, In step S42, the time-domain features include root mean square, kurtosis, and peak factor.

7. The method for fault diagnosis of submersible electric pumps based on CFD and deep learning according to claim 5, characterized in that, In step S43, the frequency domain feature is used to determine the relationship between the shaft rotation frequency, blade frequency and their harmonics and the physical structure of the equipment.

8. The method for fault diagnosis of submersible electric pumps based on CFD and deep learning according to claim 1, characterized in that, Multichannel deep neural networks include parallel feature extraction channels, feature splicing layers, attention fusion layers, and multi-task output layers.

9. The method for fault diagnosis of submersible electric pumps based on CFD and deep learning according to claim 8, characterized in that, Each channel of the parallel feature extraction channel consists of stacked one-dimensional convolutional layers and a bidirectional long short-term memory network; The multi-channel features extracted by the parallel feature extraction channels are concatenated by a feature concatenation layer to obtain a comprehensive feature vector; The attention fusion layer linearly transforms the comprehensive feature vector into query, key, and value through a learnable weight matrix, calculates the correlation between the query and the key, normalizes it into a weight score using the Softmax function, and generates fused features by weighting and summing the values ​​using the weight scores. The multi-task output layer includes a classification branch and a regression branch. The classification branch uses the Sigmoid activation function and outputs the probability of multiple faults to achieve multi-label classification. The regression branch uses the Linear activation function and outputs the severity estimate of each fault.

10. A fault diagnosis system for submersible electric pumps based on CFD and deep learning, characterized in that, To perform the CFD and deep learning-based submersible electric pump fault diagnosis method as described in any one of claims 1 to 9, comprising: The end-side acquisition module includes a submersible electric pump unit, a multi-source sensor, and a data acquisition unit. The multi-source sensor is used to collect the vibration, current, pressure, and temperature of the submersible electric pump, and the data acquisition unit converts the vibration, current, pressure, and temperature of the submersible electric pump into a standardized data stream. The edge-side diagnostic module includes an edge computing gateway, a lightweight diagnostic model, and a local early warning submodule. The edge computing gateway receives and preprocesses standardized data streams. The lightweight diagnostic model is a multi-channel deep neural network pre-trained by CFD simulation and fine-tuned by on-site data, used to complete fault type identification and severity quantification in real time. The local early warning submodule realizes local millisecond-level alarm and control for faults. The cloud-side optimization module includes a cloud gateway and a visualization platform. The cloud gateway enables data interaction between the edge-side diagnostic module and the cloud-side optimization module, uploading fault type identification and severity quantification to the cloud center. Through in-depth analysis and cross-validation using historical data and expert knowledge base, a comprehensive decision report is generated, including fault root causes, evolution trends, maintenance priorities, and maintenance recommendations. This report is then pushed to maintenance personnel through the visualization platform. The actual results verified through on-site inspection are fed back as new labeled data to the edge-side diagnostic module for incremental learning and optimization of the model.

Images