A Deep Learning-Based Method and System for Pump Fault Identification and Diagnosis

By using multi-source time-series data cascade coupling correction, CNN-BiLSTM dual-branch parallel feature extraction, and adaptive weight fusion, the problems of single feature extraction, weak anti-interference ability, and poor generalization in existing technologies for pump fault diagnosis are solved, and high-precision, fully automated pump fault identification and operation and maintenance decision-making are realized.

CN123069083APending Publication Date: 2026-11-10SHANDONG HUAFANGYUN ENERGY SAVING INTEGRATION CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611498641.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-24
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing deep learning-based pump fault diagnosis technologies suffer from problems such as one-sided feature extraction, weak anti-interference ability, insufficient generalization ability, lack of systematic solutions, and poor adaptability to small samples, making it difficult to meet the fault identification needs of complex working conditions and various pump models.

Method used

A multi-source time-series data tiered coupling correction, CNN and BiLSTM dual-branch parallel feature extraction, adaptive weight feature fusion, working condition adaptive fine-tuning migration, and full-process closed-loop management system are constructed. Through multi-task learning and adaptive correction mechanisms, high-precision identification and diagnosis of water pump faults are achieved.

Benefits of technology

It improves the model's diagnostic accuracy and anti-interference ability under complex working conditions, enhances its generalization ability across working conditions, realizes fully automated fault identification and operation and maintenance decision-making, and adapts to the fault identification needs of different types of water pumps.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069083A_ABST
    Figure CN123069083A_ABST
Patent Text Reader

Abstract

This invention relates to a deep learning-based method and system for identifying and diagnosing pump faults, belonging to the field of pump condition monitoring and intelligent fault diagnosis. The method includes: acquiring multi-source time-series data of pump operation; performing cascade coupling correction on vibration and temperature signals based on pressure and flow signals; inputting the corrected data in parallel into a CNN spatial feature extraction branch and a BiLSTM temporal feature extraction branch to mine fault spatial distortion and temporal degradation features; identifying faults after adaptive weight fusion of the extracted dual-branch features; and performing fault location and source tracing, trend prediction, and maintenance decisions based on the diagnostic results. The advantages of this invention are: improved input data quality through coupling correction; synchronous mining of spatiotemporal features through a dual-branch parallel structure; and enhanced robustness to non-stationary operating conditions through adaptive fusion. This effectively solves the problems of existing technologies, such as single feature extraction, weak anti-interference, poor generalization, and lack of systematic solutions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a deep learning-based method for identifying and diagnosing pump faults, belonging to the field of pump condition monitoring and intelligent fault diagnosis. Background Technology

[0002] As core fluid transport equipment in industrial production, municipal water supply, water conservancy pumping stations, and HVAC systems, water pumps operate continuously at high speeds for extended periods. They are susceptible to various faults caused by factors such as media erosion, mechanical wear, load fluctuations, and sudden changes in operating conditions, including bearing damage, impeller blockage, pump body cavitation, seal leakage, and motor overload. Water pump faults are characterized by their insidious nature, suddenness, and strong interconnectedness. Even minor faults, if not identified and addressed promptly, can easily lead to equipment downtime, system paralysis, and pipeline leaks, causing serious economic losses and safety hazards. Therefore, achieving high-precision real-time monitoring of water pump operating status and accurate fault diagnosis is crucial for ensuring safe equipment operation and reducing maintenance costs.

[0003] Currently, various technical systems have been developed both domestically and internationally for pump fault identification and diagnosis. Traditional fault diagnosis methods mainly rely on signal processing and manual feature engineering, such as Fourier transform, wavelet packet decomposition, and empirical mode decomposition, to extract statistical features of vibration, pressure, and temperature signals. These features are then combined with shallow machine learning models such as support vector machines and backpropagation neural networks for classification and identification. However, traditional methods heavily depend on expert experience for manual feature design, resulting in limited feature generalization capabilities and making it difficult to address the non-stationarity, nonlinearity, and multi-source coupling characteristics of pump fault signals under complex operating conditions.

[0004] In recent years, deep learning technology, with its advantages of automatic feature extraction, strong nonlinear fitting, and high-dimensional data processing, has been gradually applied to the field of pump fault diagnosis. Existing research mostly employs basic deep learning models such as single convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and autoencoders to achieve automatic identification and classification of pump faults, thus to some extent eliminating the reliance on manual feature design in traditional methods. Among these, CNNs excel at extracting local spatial features of signals (e.g., application number 202110542679.7, invention titled "Deep Learning Classification Model and Fault Diagnosis Method for Fault Diagnosis," which uses CNNs to extract one-dimensional amplitude features); LSTMs can capture long-term dependency features of time-series data (e.g., the aforementioned literature simultaneously uses LSTMs to extract two-dimensional time-series features in both the time and frequency domains). Basic autoencoders are used for data dimensionality reduction and feature denoising (e.g., patent application number 202110529358.3, invention title: A method for graded diagnosis of emulsion pump faults), which uses stacked autoencoders to build a deep autoencoder network (DAE) for adaptive feature learning and fault identification of multi-source monitoring parameters of emulsion pumps; another example is patent application number 202210072680.2, invention title: A method and device for intelligent diagnosis of important plant water pumps, which uses stacked autoencoder neural networks to reduce the dimensionality and extract features from the vibration signals of important plant water pumps). The application of the above-mentioned basic deep learning models has improved the fault diagnosis performance under complex working conditions to a certain extent.

[0005] However, existing deep learning-based pump fault diagnosis technologies still have many unresolved technical challenges and cannot meet the actual needs of industrial applications: First, single-model feature extraction is one-sided. Existing methods often employ a single network structure. For example, application number 202110542679.7, entitled "Deep Learning Classification Model and Fault Diagnosis Method for Fault Diagnosis," while using a multi-dimensional feature extraction approach—CNN for one-dimensional features, LSTM for two-dimensional features, and MFCC for three-dimensional features—still employs a serial structure for its one-dimensional, two-dimensional, and three-dimensional feature extraction modules. Features are extracted sequentially and then fused for training, failing to achieve simultaneous parallel mining of spatial and temporal domain features. Similarly, patent application number 202110529358.3, entitled "A Fault Grading Diagnosis Method for Emulsion Pumps," uses a deep autoencoder network (DAE) that primarily focuses on dimensionality reduction and global feature learning, also lacking the ability to jointly capture local spatial distortion features and temporal degradation features. These methods struggle to simultaneously consider the spatiotemporal correlation features of pump operation data, failing to comprehensively characterize the equipment's operating status and exhibiting insufficient accuracy in identifying early, subtle, and complex faults.

[0006] Second, the models have weak anti-interference capabilities. Existing basic deep learning models have poor adaptability to raw data containing noise and abnormal disturbances in industrial settings. Although patent application number 202210072680.2, entitled "A Method and Device for Intelligent Diagnosis of Important Plant Water Pumps," employs wavelet threshold denoising preprocessing, it only performs one-time denoising at the data input end and does not design embedded denoising and feature enhancement mechanisms within the model; patent application number 202110529358.3, entitled "A Method for Fault Grading Diagnosis of Emulsion Pumps," uses a deep autoencoder network, which has a certain denoising and reconstruction capability, but does not design dedicated feature enhancement and noise suppression strategies for the complex operating scenarios of water pumps. The above methods have prominent noise sensitivity issues, which can easily lead to misdiagnosis and missed diagnosis.

[0007] Third, the generalization ability is insufficient. Existing models are mostly trained on datasets from single operating conditions and single devices, resulting in poor model transferability. For example, patent application number 202110529358.3, entitled "A Method for Fault Grading Diagnosis of Emulsion Pumps," uses training data from a single model of emulsion pump (BRW500 / 31.5 type). While its deep autoencoder network achieves high diagnostic accuracy on a single device, it lacks mechanisms for model transfer and adaptive adjustment across different models and operating conditions. Similarly, patent application number 202210072680.2, entitled "A Method and Device for Intelligent Diagnosis of Important Plant Water Pumps," also fails to optimize model generalization under different operating conditions. Faced with different pump models, different load conditions, and different operating environments, the diagnostic accuracy of these models drops significantly, exhibiting extremely poor versatility.

[0008] Fourth, there is a lack of systematic solutions. Existing technologies mostly focus on algorithm model optimization, only achieving the single function of fault identification and classification. While patent application number 202210072680.2, entitled "A Method and Device for Intelligent Diagnosis of Important Plant Water Pumps," sets up a hierarchical diagnostic architecture of "first-level online fault identification" and "second-level offline fault location," its expert system is offline-triggered, requiring manual input of necessary fault information before fault location can be performed. It cannot achieve a fully automated closed-loop process from data acquisition, preprocessing, and model inference to fault location, trend prediction, and operation and maintenance decisions. Similarly, it stops at fault type identification and output, failing to form a complete solution for industrial application. The above technologies are difficult to directly apply to actual scenarios such as industrial pumping stations and industrial production.

[0009] Fifth, poor sample fit. Industrial pump failure samples are scarce, positive and negative samples are imbalanced, and annotation costs are high. While the deep autoencoder network in patent application 202110529358.3 (invention title: A Method for Grading and Diagnosing Faults of Emulsion Pumps) reduces reliance on labeled data by combining unsupervised pre-training with supervised fine-tuning, a large number of labeled samples are still required during the supervised fine-tuning stage. Patent applications 202110542679.7 (invention title: Deep Learning Classification Model and Fault Diagnosis Method for Fault Diagnosis) and 202210072680.2 (invention title: A Method and Device for Intelligent Diagnosis of Important Plant Water Pumps) both employ supervised learning for model training, relying on a large amount of high-quality labeled fault data. Existing supervised and semi-supervised deep learning models perform extremely poorly in small-sample scenarios, making it difficult to meet practical engineering needs.

[0010] To address the shortcomings of existing technologies, this invention provides a deep learning-based method and system for identifying and diagnosing water pump faults. By constructing a multi-source time-series data cascade coupling correction, parallel feature extraction using CNN and BiLSTM dual branches, adaptive weighted feature fusion, adaptive fine-tuning and migration based on operating conditions, and a closed-loop management system throughout the entire process, this method achieves high-precision identification and diagnosis of water pump faults. It effectively improves the model's anti-interference capability, cross-operating condition generalization capability, and applicability to industrial scenarios, overcoming the problems of single feature extraction, weak anti-interference capability, poor generalization, lack of systematic solutions, and poor small sample adaptability in existing technologies. Summary of the Invention

[0011] To overcome the shortcomings of existing technologies, this invention provides a water pump fault identification and diagnosis method based on deep learning. The technical solution of this invention is as follows: A deep learning-based method for water pump fault identification and diagnosis includes the following steps: Step S1: Acquire multi-source time-series data during the operation of the water pump. The multi-source time-series data includes at least vibration signal, pressure signal, flow signal, and temperature signal. Using the pressure signal and flow signal as steady-state reference signals, perform stepwise coupling correction on the vibration signal and temperature signal to achieve physical alignment and coupling enhancement of multi-source heterogeneous data, and output coupled preprocessed data. Step S2: Input the coupled preprocessed data into the parallel spatial feature extraction branch and temporal feature extraction branch respectively, and extract the fault spatial distortion features and fault temporal degradation features simultaneously; The spatial feature extraction branch uses a convolutional neural network to extract fine-grained fault spatial features from shallow, middle and deep layers through multi-level convolutional pooling operations, and outputs a spatial feature set. The temporal feature extraction branch adopts a bidirectional long short-term memory network, which traverses the temporal fluctuation pattern of the coupled preprocessed data in both forward and reverse directions, captures the temporal dependency features of long-term fault degradation, and outputs a temporal feature set after splicing forward and reverse features. Step S3: Fuse the spatial feature set with the temporal feature set, and perform fault identification and classification on the pump operating status based on the fused coupled features, and output the fault diagnosis results; Step S4: Based on the fault diagnosis results of step S3, perform fault location and source tracing, operation trend prediction and intelligent operation and maintenance decision-making to form a closed-loop system for the entire process of water pump fault identification and diagnosis.

[0012] The step-by-step coupling correction in step S1 specifically includes: The raw time-series signal acquired by the i-th type of sensor Where i = 1, 2, 3, 4 correspond to vibration signal, pressure signal, flow signal, and temperature signal, respectively. The coupled preprocessed data is calculated according to the following formula. : ; Among them, the coupling weights of each branch signal The signal from this branch and the pressure reference signal Flow reference signal The Pearson correlation coefficient is adaptively determined to achieve deep coupling correction of multi-source heterogeneous data.

[0013] The spatial feature extraction branch employs a three-level convolutional pooling tiered structure to extract fine-grained features of the fault space step by step, and the coupled preprocessed data output from step S1 is denoted as the input of layer 0. Specifically, the features of the k-th layer are calculated step by step according to the following formula: in, The weights of the k-th layer convolutional kernel are... The feature map output from layer (k-1) is downsampled after each convolutional layer by pooling. The three-level convolutional-pooling structure extracts shallow edge texture fault features, mid-level local distortion fault features, and deep global semantic fault features, respectively. The outputs of each level are concatenated to form the spatial feature set. .

[0014] The temporal feature extraction branch uses a bidirectional long short-term memory network to extract temporal features, and the coupled preprocessed data output from step S1 is denoted as... The forward and backward time series features are calculated using the following formulas: ; ; forward temporal features With reverse time series characteristics The data is spliced ​​together to form a temporal feature set. : ; in, This represents the forward LSTM computation process. This represents the reverse LSTM computation process. This represents a vector concatenation operation, resulting in a temporal feature set. It fully covers the timing fluctuations and degradation patterns of water pumps during operation.

[0015] In step S3, the specific steps of feature fusion include: For the spatial feature set and the time series feature set Adaptive correction of feature weights is performed separately, and the spatial feature weights are dynamically adjusted based on the signal quality assessment results of the current input signal. and time series feature weights The fusion features are calculated according to the following formula: ;in, = 1, and The value of is dynamically determined by the signal quality assessment results under the current operating conditions.

[0016] Before performing step S2, the process also includes model pre-training and condition-adaptive fine-tuning transfer steps: The convolutional neural network and the bidirectional long short-term memory network are pre-trained using operational datasets of multiple scenarios and various types of water pumps to obtain globally universal model parameters. ; When applied to a target scenario, the global general model is adaptively adjusted using small sample data from the target scenario. This involves freezing the parameters of the first few network layers and adjusting only the parameters of the last few layers to obtain scenario-fine-tuning parameters. This allows the model to adapt to the fault diagnosis requirements of the current target operating conditions. The model pre-training employs a multi-task learning strategy, simultaneously optimizing the main task of pump fault classification and the auxiliary task of operating condition type identification to enhance the model's feature transfer capability across different operating conditions.

[0017] The closed-loop system in step S4 includes: S4-1. Deploy various sensors to collect multi-source time-series data of water pump operation in real time; S4-2. Perform missing value imputation, outlier detection and removal, normalization, and multi-source data cascade coupling correction on the collected raw data. S4-3. Deploy the parallel two-branch feature extraction coupled model of the trained convolutional neural network and bidirectional long short-term memory network, and output the fault identification classification results and fault probability confidence. S4-4. Based on the spatial feature activation map and temporal feature attention distribution output by the coupling model, locate the physical component location and start time of the fault, and output a fault tracing report. S4-5. Based on the current diagnostic results and historical degradation trajectory, predict the evolution trend of the pump's operating status and output the predicted value of its remaining service life; S4-6. Based on the fault type, fault severity, and remaining service life prediction results, match the preset operation and maintenance strategy knowledge base, and output an operation and maintenance decision scheme including maintenance priority, maintenance measure suggestions, and spare parts replacement plan.

[0018] A data preprocessing step is set before step S1: The original multi-source time-series data is timestamped and resampled to ensure that the signals from various sensors have a uniform sampling frequency and time window length. Missing values ​​are filled using linear interpolation. Outliers exceeding a preset threshold are detected and removed. The vacant positions after removal are filled with valid values ​​from adjacent time points. The data of each branch that has been filled in are standardized so that the signals of each physical dimension are uniformly mapped to the same numerical range.

[0019] The training process of the convolutional neural network and the bidirectional long short-term memory network includes the following steps: Step T1: Collect multi-source time-series data samples of multiple scenarios and various models of water pumps under normal operating conditions and various fault conditions, and label each sample with fault type label and operating condition type label to form a training dataset; Step T2: Input the training dataset into the coupled model composed of a convolutional neural network and a bidirectional long short-term memory network in parallel, set the loss function of the main task of fault classification and the loss function of the auxiliary task of working condition identification, pre-train using a multi-task joint optimization strategy, iteratively update the network parameters through the backpropagation algorithm until the loss functions of the main task and the auxiliary task converge, and save the global general model parameters θg. Step T3, adaptive fine-tuning under working conditions: Obtain small sample labeled data of the target scene, load the global general model parameters θg, freeze the parameters of the first few layers of the network, fine-tune the parameters of the last few layers of the network with the small sample data of the target scene, obtain the scene fine-tuning parameters θs, and complete the adaptation and deployment of the model under the target working conditions.

[0020] A deep learning-based water pump fault identification and diagnosis system includes: A multi-source data acquisition unit is used to acquire multi-source time-series data output by various types of sensors during the operation of the water pump in real time. The multi-source time-series data includes at least vibration signals, pressure signals, flow signals, and temperature signals. The data preprocessing unit is connected to the multi-source data acquisition unit. It is used to perform missing value interpolation, outlier detection and removal, timestamp alignment and normalization on the acquired multi-source time series data. Then, it uses the pressure signal and flow signal as steady-state reference signals to perform step-coupling correction on the vibration signal and temperature signal, and outputs coupled preprocessed data. A coupled feature extraction unit, connected to the data preprocessing unit, includes a parallel spatial feature extraction subunit and a temporal feature extraction subunit. The spatial feature extraction subunit uses a convolutional neural network to extract fault spatial distortion features step by step through multi-level convolutional pooling operations and outputs a spatial feature set. The temporal feature extraction subunit uses a bidirectional long short-term memory network to traverse the temporal fluctuation patterns of the coupled preprocessed data in both forward and backward directions and outputs a temporal feature set. The feature fusion and fault classification unit is connected to the coupled feature extraction unit and is used to fuse the spatial feature set and the temporal feature set, perform fault identification and classification on the pump operating status based on the fused coupled features, and output fault diagnosis results. The full-process closed-loop management unit is connected to the feature fusion and fault classification unit, and is used to perform fault location and source tracing, operation trend prediction and / or intelligent operation and maintenance decision-making based on the fault diagnosis results, forming a full-process closed-loop system for water pump fault identification and diagnosis.

[0021] The advantages of this invention are: 1. Using pressure and flow signals as steady-state benchmarks, the coupling weights of each branch signal are adaptively determined using the Pearson correlation coefficient, and tiered coupling correction is performed on vibration and temperature signals. Since pressure and flow parameters directly reflect the steady-state operating point of the pump during operation, while vibration and temperature signals are often superimposed with operating condition fluctuations and environmental noise interference, this invention calibrates unsteady signals using steady-state benchmark signals through physical coupling relationships. This effectively eliminates abnormal data components introduced by load fluctuations and environmental disturbances, resulting in cleaner data input to the deep learning model with stronger physical consistency. Compared to existing technologies that only perform one-time wavelet threshold denoising at the input end, this invention's tiered coupling correction utilizes the inherent physical coupling relationship of the pump for targeted correction at the data source. This more effectively preserves fault-related features and suppresses operating condition disturbances unrelated to faults, providing high-quality input data for subsequent deep learning models, thereby improving the model's diagnostic accuracy and anti-interference capability in complex industrial scenarios.

[0022] 2. The coupled and corrected data is input in parallel into the CNN spatial feature extraction branch and the BiLSTM temporal feature extraction branch. The CNN branch uses a three-level convolutional pooling ladder structure to extract shallow edge texture fault features, mid-level local distortion fault features, and deep global semantic fault features, which can keenly capture the local distortion morphology of pump faults in the vibration waveform. The BiLSTM branch traverses the temporal data in both forward and backward directions to fully capture the long-term degradation law and temporal dependency of the pump's operating state. The two branches work synchronously and in parallel without interfering with each other, and finally fuse the spatial and temporal features.

[0023] 3. When fusing spatial and temporal feature sets, the spatial feature weight α and temporal feature weight β are dynamically determined based on the signal quality assessment results of the current input signal, with α + β = 1. Under stable operating conditions with high signal-to-noise ratios, both spatial and temporal features provide effective information, and their weights are evenly distributed. However, in scenarios with large fluctuations in operating conditions and degraded signal quality, the system automatically adjusts the weight distribution, allowing the feature branch with better quality to dominate the fusion result. This adaptive correction mechanism ensures that the feature fusion result is always based on the most reliable features under the current operating conditions, avoiding the problem of a sharp drop in overall diagnostic accuracy caused by the failure of a certain dimension feature when operating conditions change, which is a problem with the fixed-weight fusion method in existing technologies. This significantly improves the model's adaptability and diagnostic stability under complex and non-stationary industrial operating conditions.

[0024] 4. First, the coupled model is pre-trained using multi-task datasets of various pump models across multiple scenarios, while simultaneously optimizing the main fault classification task and the auxiliary task of operating condition identification. By introducing the auxiliary task of operating condition identification, the model is forced to learn to distinguish data distribution differences under different operating conditions during the pre-training stage, thereby extracting common fault features that are insensitive to changes in operating conditions. When deployed to the target scenario, only a small amount of sample data from the target scenario is needed to achieve rapid operating condition adaptation by freezing the front-layer network and fine-tuning the back-layer network. The general fault representation capability obtained through multi-task pre-training, combined with a lightweight fine-tuning strategy, maintains a high diagnostic accuracy even in industrial scenarios where fault samples are scarce, fundamentally overcoming the problem of extremely poor diagnostic performance in small-sample scenarios in existing technologies.

[0025] 5. A closed-loop system covering six key stages has been constructed, encompassing data acquisition, intelligent preprocessing, coupled model reasoning, fault location and tracing, operational trend prediction, and intelligent operation and maintenance decision-making. In particular, fault location and tracing, based on spatial feature activation maps and temporal feature attention distribution, can accurately pinpoint the physical component location and start time of the fault; operational trend prediction outputs a predicted remaining service life value based on current diagnostic results and historical degradation trajectories; and intelligent operation and maintenance decision-making directly outputs an executable maintenance plan by matching a pre-set operation and maintenance strategy knowledge base with the fault type, severity, and remaining service life.

[0026] In summary, this invention, through the systematic synergy of five technical approaches—multi-source data hierarchical coupling correction, CNN-BiLSTM dual-branch parallel feature extraction, adaptive weight feature fusion, multi-task pre-training and conditional fine-tuning transfer, and an end-to-end closed-loop system—comprehensively addresses the technical shortcomings of existing technologies, such as single feature extraction, weak anti-interference ability, poor generalization, poor sample adaptability, and lack of systematic solutions, from five dimensions: data quality improvement, enhanced feature extraction capability, optimized fusion strategy, improved small-sample generalization ability, and construction of a systematic solution. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the method of the present invention.

[0028] Figure 2 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0029] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of the present invention. Those skilled in the art should understand that modifications or substitutions can be made to the details and form of the technical solutions of the present invention without departing from the spirit and scope of the present invention, but all such modifications and substitutions fall within the protection scope of the present invention.

[0030] See Figure 1 and Figure 2 This invention relates to a deep learning-based method for identifying and diagnosing water pump faults, comprising the following steps: Step S1: Acquire multi-source time-series data during the operation of the water pump. The multi-source time-series data includes at least vibration signal, pressure signal, flow signal, and temperature signal. Using the pressure signal and flow signal as steady-state reference signals, perform stepwise coupling correction on the vibration signal and temperature signal to achieve physical alignment and coupling enhancement of multi-source heterogeneous data, and output coupled preprocessed data. Step S2: Input the coupled preprocessed data into the parallel spatial feature extraction branch and temporal feature extraction branch respectively, and extract the fault spatial distortion features and fault temporal degradation features simultaneously; The spatial feature extraction branch uses a convolutional neural network to extract fine-grained fault spatial features from shallow, middle and deep layers through multi-level convolutional pooling operations, and outputs a spatial feature set. The temporal feature extraction branch adopts a bidirectional long short-term memory network, which traverses the temporal fluctuation pattern of the coupled preprocessed data in both forward and reverse directions, captures the temporal dependency features of long-term fault degradation, and outputs a temporal feature set after splicing forward and reverse features. Step S3: Fuse the spatial feature set with the temporal feature set, and perform fault identification and classification on the pump operating status based on the fused coupled features, and output the fault diagnosis results; Step S4: Based on the fault diagnosis results of step S3, perform fault location and source tracing, operation trend prediction and intelligent operation and maintenance decision-making to form a closed-loop system for the entire process of water pump fault identification and diagnosis.

[0031] In step S1, this invention uses pressure and flow signals as steady-state benchmarks and leverages the inherent fluid dynamics and physical coupling relationship of the pump to perform tiered coupling correction on vibration and temperature signals. The weights of each branch are adaptively determined using the Pearson correlation coefficient, giving higher weights to effective signals with strong physical correlation. This effectively suppresses operational disturbances and noise interference unrelated to faults, improving input quality from the data source and providing coupled preprocessed data with stronger physical consistency and more significant fault characteristics for subsequent deep learning models.

[0032] In step S2, this invention employs a parallel dual-branch structure of CNN and BiLSTM to simultaneously extract spatial and temporal features, overcoming the limitations of existing technologies that rely on cascaded extraction or single-network structures. The CNN branch captures shallow edge textures, mid-level local distortions, and deep global semantic fault features through a three-level convolutional pooling cascaded structure; the BiLSTM branch traverses the temporal data bidirectionally along both the forward and backward paths, fully covering the long-term degradation patterns of the pump's operating state. The two branches work in parallel without interference, avoiding the cumulative propagation of errors from earlier stages to later stages in a cascaded structure, thus achieving comprehensive and accurate capture of early weak faults and complex fault features.

[0033] In step S3, this invention introduces an adaptive weight correction mechanism in the feature fusion stage, dynamically adjusting the spatial feature weight α and the temporal feature weight β based on the current signal quality. When the quality of a certain dimension feature deteriorates due to fluctuations in operating conditions, the system automatically reduces the weight of that dimension and increases the contribution of another dimension, ensuring that the fused features are always based on the most reliable features under the current operating conditions. Compared to the fixed-weight fusion method in existing technologies, the adaptive fusion strategy of this invention significantly improves the diagnostic robustness and accuracy of the model under non-stationary industrial operating conditions.

[0034] In step S4, this invention constructs an end-to-end closed-loop system encompassing data acquisition, preprocessing, model inference, fault location and tracing, operational trend prediction, and intelligent operation and maintenance decision-making. Fault location accurately determines the location and start time of the faulty component based on spatial feature activation maps and temporal attention distribution; trend prediction outputs the remaining service life; and operation and maintenance decision-making directly matches the strategy knowledge base to provide an executable maintenance plan. The entire process requires no manual intervention, meeting the practical needs of intelligent fault diagnosis throughout the entire life cycle of water pumps in industrial pumping stations, municipal water supply, and other scenarios.

[0035] The step-by-step coupling correction in step S1 specifically includes: The raw time-series signal acquired by the i-th type of sensor Where i = 1, 2, 3, 4 correspond to vibration signal, pressure signal, flow signal, and temperature signal, respectively. The coupled preprocessed data is calculated according to the following formula. : ; Among them, the coupling weights of each branch signal The signal from this branch and the pressure reference signal Flow reference signal The Pearson correlation coefficient is adaptively determined to achieve deep coupling correction of multi-source heterogeneous data.

[0036] The following example briefly illustrates the calculation process of the tiered coupling correction in step S1: Suppose that at a certain sampling time, the normalized raw signals collected by each sensor of a water pump are as follows: Vibration signal Pressure signal ; Flow signal Temperature signal ; Among them, pressure signal Pick = 0.92, flow signal Pick = 0.88.

[0037] Step 1: Calculate the coupling weights of each branch signal. Based on historical window data (such as the past 100 sampling points), calculate the Pearson correlation coefficients between each branch signal and the pressure reference signal and the flow reference signal: The correlation coefficient between the vibration signal and pressure is 0.75, and the correlation coefficient with flow rate is 0.65. Therefore: ; The pressure signal has a correlation coefficient of 1.00 with itself as the reference signal, and a correlation coefficient of 0.80 with the flow rate. Therefore: ; The correlation coefficient between the flow rate signal and the pressure is 0.80, and the correlation coefficient with itself is 1.00. Therefore: ; The correlation coefficient between temperature and pressure is 0.40, and the correlation coefficient with flow rate is 0.35. Therefore: ; Step 2: Calculate the coupling preprocessed data : ; 0; Pressure signal ( ) and flow signals ( As a steady-state benchmark, it has a high weight due to its high correlation with itself or with each other, and occupies a dominant position in the coupling result, playing the role of stabilizing and correcting the benchmark.

[0038] Vibration signal ( The correlation coefficient with pressure and flow benchmarks is moderate, and the weight is appropriate. While the effective fault components are retained, the abnormal disturbance components with large deviations from the benchmarks are suppressed.

[0039] Temperature signal ( The low correlation coefficient between the current temperature data and the pressure and flow rate benchmarks indicates that the current temperature data is greatly affected by environmental interference or has a weak correlation with the current steady-state operating point of the water pump. Therefore, the weight is automatically reduced, and its contribution to the coupled output is weakened.

[0040] The spatial feature extraction branch employs a three-level convolutional pooling tiered structure to extract fine-grained features of the fault space step by step, and the coupled preprocessed data output from step S1 is denoted as the input of layer 0. Specifically, the features of the k-th layer are calculated step by step according to the following formula: in, The weights of the k-th layer convolutional kernel are... The feature map output from layer (k-1) is downsampled after each convolutional layer by pooling. The three-level convolutional-pooling structure extracts shallow edge texture fault features, mid-level local distortion fault features, and deep global semantic fault features, respectively. The outputs of each level are concatenated to form the spatial feature set. .

[0041] Assuming the coupled preprocessed data output in step S1 It is a one-dimensional time-series signal with a length of L = 1024, i.e. This data integrates comprehensive information from four types of sensors: vibration, pressure, flow, and temperature, after physical coupling correction.

[0042] Step 1: Level 1 (Shallow Feature Extraction, k=1) Configuration: Layer 1 convolutional kernel Size is The number is 16, the step size is 1; the pooling size is... The step size is 2.

[0043] ; Calculation process: After convolution, the feature length is approximately 1024 - 16 + 1 = 1009. After pooling downsampling, the length is approximately... Output feature map .

[0044] Feature type extracted: shallow edge texture fault features. At this stage, the receptive field is small, primarily capturing localized subtle changes in the signal, such as spikes and glitches in vibration waveforms, minute fluctuations in pressure pulsations, and instantaneous jumps in temperature signals. These edge texture features correspond to early, minor anomalies in pump operation, such as the weak impact response caused by initial pitting of bearing balls.

[0045] The first-level convolutional kernel focuses on changes in signal morphology within a short time window, and can keenly capture local waveform distortions caused by faults, but it does not yet have the ability to distinguish the overall fault mode.

[0046] Step 2: Level 2 (Mid-level feature extraction, k=2) Configuration: Layer 2 convolutional kernel Size is The number is 32, the step size is 1; the pooling size is... The step size is 2. ; Calculation process: The feature length after convolution is approximately 505 - 8 + 1 = 498, and the length after pooling is approximately... =249, output feature map .

[0047] Feature type extracted: Mid-level local distortion fault features. The receptive field expands to dozens of time points, beginning to combine shallow textures into local structural patterns. At this stage, it is able to identify medium-scale signal distortion patterns, such as periodic abnormal increases in vibration amplitude, deformation of flow waveforms, and abnormal recovery characteristics of pressure signals.

[0048] The second-level convolutional kernel begins to abstract local structural combinations, corresponding to medium-scale symptoms of pump failure. For example, periodic pulsation distortion of flow caused by slight impeller blockage, and changes in the vibration envelope morphology caused by increased bearing wear. At this point, the features begin to have a certain pattern discrimination capability.

[0049] Step 3: Level 3 (Deep Feature Extraction, k=3) Configuration: 3rd layer convolutional kernel Size is The number is 64, the step size is 1; the pooling size is... The step size is 2.

[0050] ; Calculation process: The feature length after convolution is approximately 249 - 4 + 1 = 246, and the length after pooling is approximately... = 123, output feature map .

[0051] Feature extraction type: Deep global semantic fault features. The receptive field covers the entire signal range, enabling the synthesis of global information to form a high-level semantic representation of fault modes. This stage no longer focuses on specific waveform details, but outputs abstract semantic features of fault categories.

[0052] The third-level convolutional kernel obtains the global receptive field, which can integrate various abnormal signs throughout the entire operation cycle of the water pump to form a high-level semantic judgment of the overall failure mode. For example, it can distinguish global category features such as "bearing failure", "cavitation failure" and "seal leakage".

[0053] Step 4: Multi-level feature splicing to form a spatial feature set The three-level outputs are concatenated according to the channel dimension: ; Output result: Includes multi-scale spatial feature information ranging from local details to global semantics, with a feature dimension of [missing information]. = 8080 + 7968 + 7872 = 23920 dimensions (assuming it is flattened into a one-dimensional vector).

[0054] Through a three-level convolutional pooling hierarchical structure, the CNN branches abstract spatial features progressively from coarse to fine and from local to global, preserving detailed information about early, weak faults (Level 1), capturing structural patterns in fault development (Level 2), and ultimately forming high-level semantic discrimination criteria (Level 3). The three levels of features are then concatenated to form a spatial feature set. This provides a comprehensive, multi-scale spatial fault representation for subsequent fusion with BiLSTM temporal features.

[0055] The temporal feature extraction branch uses a bidirectional long short-term memory network to extract temporal features, and the coupled preprocessed data output from step S1 is denoted as... The forward and backward time series features are calculated using the following formulas: ; ; forward temporal features With reverse time series characteristics The data is spliced ​​together to form a temporal feature set. : ; in, This represents the forward LSTM computation process. This represents the reverse LSTM computation process. This represents a vector concatenation operation, resulting in a temporal feature set. It fully covers the timing fluctuations and degradation patterns of water pumps during operation.

[0056] Assume the coupled preprocessed data output from step S1 is a time-series signal comprising T=200 time steps, denoted as: ; Each time step This is the fused characteristic value after coupling correction in step S1 (a comprehensive representation of the four types of sensor signals after physical coupling correction). This data records the entire process of the water pump gradually developing from normal operation to bearing wear failure.

[0057] Step 1: Forward LSTM computation The forward LSTM starts at time step t=1 and iterates forward to t=200, reading data and updating the hidden state at each time step. At each time step t, the output of the forward LSTM is: ; in, This is the hidden state vector output by the feedforward LSTM at time step t (assuming the hidden layer dimension is 128).

[0058] At t=1, the network only knows the normal operating status of the water pump at the initial moment.

[0059] As the time step progresses to t=50, the network has accumulated information from the previous 50 time steps and begins to detect a slow increase in vibration amplitude and a slight rise in temperature. At this point, the forward hidden state... It includes information about the occurrence of the fault from the beginning to the current moment.

[0060] By t=150, the network had observed most of the running cycle and was in the forward hidden state. It includes complete evolution information from normal operation to mid-bearing wear, and captures degradation patterns such as periodic increase in vibration amplitude and continuous rise in temperature.

[0061] The final hidden state of the forward LSTM at t=200 It fully compiles all historical information from the past 200 time steps, and grasps the evolution path of when the fault started and how it developed.

[0062] For example, suppose a bearing begins to show early pitting at t=20. Before t=20, the vibration amplitude is approximately 0.2 mm / s, gradually increasing until it reaches 0.6 mm / s at t=150. The forward LSTM can accurately capture the evolution trajectory of "amplitude gradually increasing from 0.2 to 0.6" and learn the rule that "continuous increase in amplitude corresponds to accelerated bearing wear".

[0063] Step 2: Reverse LSTM computation The inverse LSTM starts at time step t=200 and iterates backwards to t=1, reading data and updating the hidden state at each time step. At each time step t, the output of the inverse LSTM is: ; in, This is the hidden state vector (also 128-dimensional) output by the inverse LSTM at time step t.

[0064] At t=200, the network starts from the endpoint state with the most severe fault and observes the entire sequence in reverse.

[0065] When the process progresses to t=150, the network has already traversed the time interval from 200 to 150 in reverse, knowing that the fault has reached a critical state at t=200, and thus hiding the state in reverse. It includes information on the degradation stage, "tracing back from the end of the failure to the present".

[0066] When t=1 is reached, the final hidden state of the inverse LSTM fully gathers the information of the entire sequence traversed in reverse, and can infer from the perspective of "the fault has occurred" how "the fault gradually developed to the current state".

[0067] If the forward LSTM can only see "amplitude gradually increasing from 0.2 to 0.6", but cannot accurately determine the specific moment when the amplitude exceeds the safety threshold (such as 0.5 mm / s), the backward LSTM works backward from the end point: it is known that the amplitude is 0.6 (severe fault) at t=200, and by going back, it is found that the amplitude first reaches 0.5 (fault alarm threshold) at t=120, thus accurately locating the key time point when the fault worsens.

[0068] Step 3: Concatenate forward and backward temporal features The hidden state of the forward LSTM at time step t is concatenated with the hidden state of the backward LSTM at time step t: ; For each time step t, the concatenated feature vector has a dimension of 128 + 128 = 256. The complete temporal feature set is: ; For time step t=100 (mid-fault stage) It includes two aspects of information: It includes the fault evolution history (past information) obtained from forward traversal, from normal operation to step 100. This includes the fault development trend (future information) obtained from the reverse traversal, "from step 100 to the critical fault endpoint".

[0069] Step 4: Output the temporal feature set The final output time series feature set of the BiLSTM branch It also covers: By employing bidirectional temporal traversal, the BiLSTM branch can simultaneously acquire historical cumulative information and future trend information at any time step t. The concatenated temporal feature set... It comprehensively covers the temporal fluctuations and degradation patterns of water pump operation, enabling it to trace the initiation and evolution of faults (forward) and predict their development direction (backward), providing a comprehensive and bidirectional temporal fault representation for subsequent fusion with CNN spatial features. This bidirectional modeling capability makes the model particularly sensitive to slowly degrading faults (such as bearing wear and seal aging), allowing for earlier detection of fault signs and more accurate prediction of fault trends.

[0070] In step S3, the specific steps of feature fusion include: For the spatial feature set and the time series feature set Adaptive correction of feature weights is performed separately, and the spatial feature weights are dynamically adjusted based on the signal quality assessment results of the current input signal. and time series feature weights The fusion features are calculated according to the following formula: ;in, = 1, and The value of is dynamically determined by the signal quality assessment results under the current operating conditions.

[0071] Assuming that at a certain pump operating moment, step S2 outputs the following characteristics: Spatial feature set Extracted by a CNN branch, it mainly reflects the local distortion morphology in the vibration waveform, including the impact pulse characteristics caused by bearing wear; Temporal feature set Extracted by the BiLSTM branch, it mainly reflects the long-term evolution trend of vibration amplitude, including the degradation law of amplitude gradually increasing from 0.2 mm / s to 0.5 mm / s.

[0072] Operating Condition 1: The water pump is operating in a steady state (pressure and flow are stable). Signal quality assessment results: Pressure fluctuation ≤ ±2%, flow fluctuation ≤ ±3%, stable operating conditions; The sensor signal has a high signal-to-noise ratio, the spatial features extracted by CNN are clear and reliable, and the temporal features extracted by BiLSTM are equally effective; Signal quality assessment system determines: spatial characteristics Quality score = 0.85, time series characteristics Quality score = 0.82, both are at a relatively high level.

[0073] Weights are determined dynamically: ; Fusion computing: ; Results Analysis: Under steady-state conditions, both spatial and temporal characteristics can provide high-quality fault information, and their weights are evenly distributed. , The fusion feature fully utilizes the local distortion capture capability of CNN and the global degradation tracking capability of BiLSTM to achieve optimal feature complementarity.

[0074] Operating Condition 2: The water pump is in a transient operating condition (severe load fluctuations and large pressure jumps). Signal quality assessment results: Pressure fluctuation ≥ ±15%, flow fluctuation ≥ ±12%, drastic changes in operating conditions; Severe pressure / flow fluctuations introduce a large number of non-fault impact components into the vibration signal, which seriously interferes with the CNN's extraction of real fault spatial features, resulting in a large amount of operating condition noise mixed in the spatial feature F_s, and a significant decrease in quality. In contrast, BiLSTM branches, due to their long-term memory capabilities, can smooth out the impact of short-term operational shocks through historical information, thus enhancing their temporal characteristics. The quality remains relatively stable; Signal quality assessment system determines: quality score of spatial features = 0.42, quality score of time series features = 0.78.

[0075] Weights are determined dynamically: ; ; Fusion computing: ; Under transient conditions of drastic load fluctuations, the system automatically adjusts the spatial feature weights. Reduce the weight of time series features to 0.35. The value was increased to 0.65, which makes the fusion features rely more on the temporal features with stronger anti-interference capabilities, effectively suppressing the spatial feature noise introduced by the operating condition disturbance and avoiding misdiagnosis and missed diagnosis.

[0076] Operating Condition 3: Local anomalies occur in the sensor signal (occasional drift of the vibration sensor). Signal quality assessment results: The vibration signal exhibits amplitude jumps (sensor drift) within a certain time period, resulting in spatial characteristics. A false impact pulse appeared in the middle; However, the BiLSTM branch, through bidirectional time series modeling, can identify that the transition does not conform to the normal degradation pattern (both forward and backward predictions show large deviations), and judge it as an abnormal disturbance rather than a real fault evolution; Signal quality assessment system determines: quality score of spatial features = 0.38, quality score of time series features = 0.80.

[0077] Weights are determined dynamically: ; ; Fusion computing: ; Results analysis: The spatial feature weight was further reduced to 0.32, which greatly reduced the interference of false spatial features introduced by sensor drift on the diagnostic results and ensured the accuracy of fault identification and classification.

[0078] Feature fusion is achieved through signal quality-driven adaptive weight correction. Under all operating conditions, it maintains the optimal combination state of "primarily based on the most reliable features at present," significantly improving the diagnostic model's adaptability and robustness against interference in complex and non-stationary operating conditions. Compared with existing technologies (such as CN202110529358.3) that use a fixed-weight fusion method, the adaptive mechanism of this invention can adjust the weight allocation in real time when changes in operating conditions cause a certain dimension feature to fail, so as to avoid the adverse effects of that dimension feature on the diagnostic results.

[0079] Before performing step S2, the process also includes model pre-training and condition-adaptive fine-tuning transfer steps: The convolutional neural network and the bidirectional long short-term memory network are pre-trained using operational datasets of multiple scenarios and various types of water pumps to obtain globally universal model parameters. ; When applied to a target scenario, the global general model is adaptively adjusted using small sample data from the target scenario. This involves freezing the parameters of the first few network layers and adjusting only the parameters of the last few layers to obtain scenario-fine-tuning parameters. This allows the model to adapt to the fault diagnosis requirements of the current target operating conditions. The model pre-training employs a multi-task learning strategy, simultaneously optimizing the main task of pump fault classification and the auxiliary task of operating condition type identification to enhance the model's feature transfer capability across different operating conditions.

[0080] Suppose a company developing an intelligent pump diagnostic system plans to deploy its model to multiple customer sites. The company has previously accumulated rich historical data from two different scenarios: Scenario A is a municipal water supply pumping station equipped with ISG single-stage centrifugal pumps, with a large amount of labeled data covering fault types such as normal operation, bearing wear, seal leakage, cavitation, and motor overload; Scenario B is an industrial production circulating water pump system equipped with S-type double-suction pumps, similarly accumulating labeled data covering fault types such as normal operation, bearing wear, seal leakage, cavitation, and impeller blockage, totaling over 100,000 sets. Now, the model needs to be deployed to Scenario C, a newly built large-scale water conservancy pumping station equipped with vertical diagonal-flow pumps, which can only provide a small number of labeled samples (approximately 50 sets), covering two operating conditions: normal operation and bearing wear. If traditional supervised learning methods are used to directly train the model on the 50 small samples in Scenario C, the model will severely overfit, resulting in extremely low diagnostic accuracy. Therefore, this invention adopts a strategy of multi-scenario pre-training combined with adaptive fine-tuning based on operating conditions to solve this problem.

[0081] Step 1: Multi-scenario pre-training The system aggregates historical operating data from scenario A (municipal water supply pumping station, ISG centrifugal pump) and scenario B (industrial production circulating water pump, S-type double-suction pump) to form a massive training dataset, totaling over 100,000 sets. Each sample is labeled not only with a specific fault type but also with an additional operating condition type label, covering information such as equipment model, flow range, and pressure range. Subsequently, this training dataset is input into a coupled model composed of CNN and BiLSTM in parallel, and pre-trained using a multi-task joint optimization strategy.

[0082] Within the multi-task learning framework, the model simultaneously optimizes two tasks. The primary task, fault classification, identifies fault types such as bearing wear, seal leakage, cavitation, impeller blockage, motor overload, and normal operation, using cross-entropy as the loss function. The secondary task, operating condition identification, determines operating condition information such as the equipment model and flow range to which the current data belongs, also employing cross-entropy loss. The total loss function is a weighted sum of the losses from the primary and secondary tasks. The weight coefficient for the secondary task is set to 0.3, ensuring that the primary task dominates the optimization process while the secondary task effectively constrains model training.

[0083] By introducing the working condition identification auxiliary task, the model is forced to learn to distinguish the differences in data distribution under different working conditions, thereby forming the ability to extract common fault features in the underlying network parameters that are insensitive to changes in working conditions. For example, the model jointly learns the bearing wear features of the ISG centrifugal pump in scenario A and the bearing wear features of the S-type double-suction pump in scenario B. The learned bearing wear features are consistent across different models and flow ranges, rather than being applicable only to a specific working condition. After pre-training, all network weights and biases of the CNN and BiLSTM are saved, which are the global universal model parameters θg.

[0084] Step 2: Adaptive Fine-tuning of Operating Conditions When the model needs to be deployed to scenario C (large-scale water conservancy pumping station, vertical mixed-flow pump), since this scenario only has 50 sets of labeled samples, it is impossible to train an effective deep learning model from scratch. The system first loads the pre-trained and saved global general model parameters θg, and then uses a freeze-fine-tuning strategy to adaptively adjust the parameters according to the working conditions.

[0085] In the freeze-fine-tuning strategy, the parameters of the first few network layers are frozen, and their weights remain unchanged during subsequent training. Specifically, the first two convolutional layers of the CNN branch and the first LSTM layer of the BiLSTM branch are kept fixed because these shallow networks have learned general waveform feature extraction capabilities (such as edge texture of vibration signals and short-term dependency patterns of time-series signals). These feature extraction capabilities can be directly reused without adjustment when transferring between different scenes. The parameters of the last few network layers are then unfrozen and fine-tuned using 50 sets of small sample data from the target scene. This includes the last convolutional and fully connected layers of the CNN branch, the last LSTM layer of the BiLSTM branch, the feature fusion layer, and the classification layer. These high-level networks are responsible for extracting semantic features related to specific fault categories and need to be adapted to the specific characteristics of bearing wear under the new model of vertical diagonal flow pump. The learning rate during fine-tuning is set to one-tenth of that in the pre-training stage to avoid destroying the learned general features during fine-tuning. After a few iterations of training, the scene fine-tuning parameters θs are obtained, completing the model's adaptation and deployment under the target working condition.

[0086] The shallow network (the first few layers) extracts general features, such as the edge texture of vibration waveforms and short-term dependency patterns of time-series signals. These features are universal across different pump models and operating conditions, eliminating the need for relearning for new scenarios. The higher-level network (the later few layers) extracts semantic features, such as abstract representations of fault categories. These features need to be adapted to the fault mode characteristics of different pump models in the target scenario. Freezing the parameters of the shallow network effectively prevents damage to the general feature extraction capability when fine-tuning on a limited number of samples. Adjusting only the parameters of the higher-level network allows the model to quickly adapt to the target scenario, achieving the transfer learning goal of solving large problems with small samples.

[0087] The closed-loop system in step S4 includes: S4-1. Deploy various sensors to collect multi-source time-series data of water pump operation in real time; S4-2. Perform missing value imputation, outlier detection and removal, normalization, and multi-source data cascade coupling correction on the collected raw data. S4-3. Deploy the parallel two-branch feature extraction coupled model of the trained convolutional neural network and bidirectional long short-term memory network, and output the fault identification classification results and fault probability confidence. S4-4. Based on the spatial feature activation map and temporal feature attention distribution output by the coupling model, locate the physical component location and start time of the fault, and output a fault tracing report. S4-5. Based on the current diagnostic results and historical degradation trajectory, predict the evolution trend of the pump's operating status and output the predicted value of its remaining service life; S4-6. Based on the fault type, fault severity, and remaining service life prediction results, match the preset operation and maintenance strategy knowledge base, and output an operation and maintenance decision scheme including maintenance priority, maintenance measure suggestions, and spare parts replacement plan.

[0088] Suppose that in an industrial pumping station, an S-type double-suction centrifugal pump has been running continuously for more than 6 months. The system, through a closed-loop process, starts by collecting raw data from sensors, and gradually completes data cleaning, model reasoning, fault location, and trend prediction, ultimately outputting an executable operation and maintenance decision plan. The entire process requires no manual intervention.

[0089] S4-1: Deploy various sensors to collect multi-source time-series data on water pump operation in real time. Accelerometers (installed in the drive-end bearing housing and non-drive-end bearing housings, collecting vibration acceleration signals at a sampling frequency of 12.8kHz), pressure sensors (installed in the pump outlet flange and inlet flange, collecting inlet and outlet pressures at a sampling frequency of 1kHz), flow sensors (installed in the outlet pipe, collecting instantaneous flow rates at a sampling frequency of 1kHz), and temperature sensors (installed in the bearing housing and pump casing, collecting bearing and pump casing temperatures at a sampling frequency of 0.1Hz) are deployed at key locations on the pump body. All sensors synchronously collect data in real time, which is then aggregated by a data acquisition card and continuously transmitted to the diagnostic system, organized into 10-second analysis windows.

[0090] S4-2: Perform missing value imputation, outlier detection and removal, normalization, and multi-source data cascade coupling correction on the collected raw data.

[0091] Within this analysis window, the temperature sensor experienced a momentary data loss due to a communication interruption. The system employed linear interpolation, averaging the measured values ​​from the moments before and after the missing moment according to a time ratio to fill the gap and restore the continuity of the temperature sequence. A single-point amplitude abrupt change occurred in the vibration signal, with the peak value exceeding the normal value by several times. Based on the 3σ criterion, this was determined to be occasional spike noise caused by electromagnetic interference, which was removed. The signal was then filled with weighted averages of the effective values ​​from adjacent moments to restore a smooth waveform. Subsequently, the four signals with different physical dimensions—vibration amplitude, pressure, flow rate, and temperature—were Z-score standardized to eliminate dimensional differences.

[0092] After standardization, tiered coupling correction is performed. Using pressure and flow signals as steady-state references, the Pearson correlation coefficients between the vibration signal and the pressure and flow references, as well as the correlation coefficients between the temperature signal and both references, are calculated. The correlation coefficient between the vibration signal and outlet pressure is 0.72, and with flow rate is 0.68, with an average coupling weight of approximately 0.70. The correlation coefficient between the temperature signal and pressure is 0.35, and with flow rate is 0.30, with an average weight of approximately 0.325. After correction, the effective vibration components that closely match the pressure and flow rate references receive higher weights, while the weight of the temperature signal, which is strongly correlated with environmental factors, is adaptively reduced, resulting in pre-processed coupled data with stronger physical consistency.

[0093] S4-3: Deploy a parallel dual-branch feature extraction coupled model of a trained convolutional neural network and a bidirectional long short-term memory network to output fault identification classification results and fault probability confidence.

[0094] The preprocessed data was input into the coupled model. After three levels of convolutional pooling extraction, the CNN branch detected periodic impact pulses in the vibration signal, with an impact interval of approximately 0.02 seconds, matching the frequency of the bearing rolling elements. The impact morphology evolved from sharp to blunt, indicating that the bearing was experiencing damage progressing from pitting to spalling. The BiLSTM branch, traversing forward, found that the bearing temperature gradually increased from 50.2℃ to 52.3℃, and the rate of increase accelerated from 0.1℃ / day in the past 7 days to 0.5℃ / day in the last 3 days. The reverse traversal confirmed that the time when the temperature first exceeded the 50℃ warning threshold coincided with the time when the vibration impact pulse first appeared, pointing to the same fault source. The fault identification classification result and fault probability confidence were output.

[0095] S4-4: Based on the spatial feature activation map and temporal feature attention distribution output by the coupled model, locate the physical component location and start time of the fault, and output a fault tracing report.

[0096] Inverse activation mapping was performed on the feature map output of level 3 of the CNN branch. High activation regions were found to be concentrated in the frequency band corresponding to the drive-end bearing housing, indicating that the fault source was located in the rolling elements of the drive-end bearing, excluding non-drive-end bearings and other components. Time-series analysis of the attention weights of the BiLSTM branch revealed that the model assigned the highest attention weight to data segments from the past 3 days. Combined with the operation logs, it was found that the pump experienced an overload operation on day 7, suggesting that this overload event was likely the inducing factor for the initial microcracks in the bearing. Based on this, the fault source report output: the faulty components were the rolling elements and inner raceway of the drive-end bearing; the fault started approximately 3 days prior; the inducing factor was overload operation leading to bearing overload; and the initial microcracks gradually expanded into spalling.

[0097] S4-5: Based on the current diagnostic results and historical degradation trajectory, predict the evolution trend of the pump's operating status and output the predicted value of its remaining service life.

[0098] The system retrieves the pump's operating history data from the past 6 months to construct a degradation trajectory model. The current vibration amplitude is approximately 0.45 mm / s, the system alarm threshold is set at 0.5 mm / s, and the shutdown threshold is set at 0.7 mm / s. Based on the actual degradation rate over the past 3 days—with the vibration amplitude increasing from 0.32 mm / s to 0.45 mm / s and the temperature rising from 50.2℃ to 52.3℃—an exponential growth curve is fitted, predicting that the vibration amplitude will reach the alarm threshold in approximately 2 days, and the shutdown threshold will be reached in approximately 7 to 8 days. The remaining service life is predicted to be approximately 7 days, with a confidence interval of ±1.5 days.

[0099] S4-6: Based on the fault type, fault severity, and remaining service life prediction results, match the preset operation and maintenance strategy knowledge base, and output an operation and maintenance decision scheme including maintenance priority, maintenance measure suggestions, and spare parts replacement plan.

[0100] The system inputs the diagnostic results (bearing rolling element wear, mid-term) and the predicted results (remaining service life approximately 7 days) into the preset maintenance strategy knowledge base. For severe bearing wear failures with less than 10 days of remaining service life, the preset strategy in the knowledge base is "high" priority for maintenance, recommending a shutdown for repair within 3 days to avoid unplanned downtime due to sudden failures. Based on this, the system outputs maintenance recommendations: replace the drive-end bearing, model SKF 6310, simultaneously check the lubricating oil condition and replace it if necessary, and check the bearing housing clearance. The spare parts replacement plan confirms that the current inventory of SKF 6310 bearings is sufficient for replacement, and also recommends simultaneously procuring 2 sets as backups, with a procurement cycle of approximately 5 to 7 days. In addition, the system outputs an alternative plan: if a shutdown cannot be arranged within 3 days, it is recommended to reduce the pump load to below 70% of the rated flow to slow the rate of failure deterioration.

[0101] The entire process achieves full automation, from raw sensor data to executable maintenance plans. When the system diagnoses a bearing failure on day 10 and predicts a remaining lifespan of approximately 7 days (i.e., reaching the downtime threshold around day 17), maintenance personnel can arrange for downtime repairs and bearing replacement before day 13, thus avoiding unplanned downtime and production interruptions caused by sudden bearing seizure. Compared to existing technologies that stop at fault type identification or require manual intervention to trigger offline expert systems (a semi-closed-loop architecture), this invention achieves a complete automated closed loop from "fault discovery" to "cause localization," then to "trend prediction," and finally "solution provision," meeting the implementation requirements of intelligent fault diagnosis throughout the entire pump lifecycle for industrial pumping stations.

[0102] A data preprocessing step is set before step S1: The original multi-source time-series data is timestamped and resampled to ensure that the signals from various sensors have a uniform sampling frequency and time window length. Missing values ​​are filled using linear interpolation. Outliers exceeding a preset threshold are detected and removed. The vacant positions after removal are filled with valid values ​​from adjacent time points. The data of each branch that has been filled in are standardized so that the signals of each physical dimension are uniformly mapped to the same numerical range.

[0103] Timestamp alignment and resampling eliminate sensor heterogeneity, providing a unified time reference for subsequent physical coupling correction. Due to differences in hardware configuration, various sensors exhibit significant variations in sampling frequencies (e.g., vibration signals reach 12.8kHz, while temperature signals are only 0.1Hz), making it impossible to correlate raw data point-to-point on the timeline. This step, through timestamp alignment and unified resampling, ensures that all signals have the same sampling frequency and time window length, guaranteeing that the four types of signals at the same moment physically correspond to the same operating state. This alignment is a prerequisite for subsequent tiered coupling correction; if the timeline is misaligned, correcting a vibration signal lagging by several seconds using the current pressure and flow reference signals will lose physical meaning. A unified time base ensures that correlation coefficient calculations in coupling correction accurately reflect the true physical relationships between the signals.

[0104] Missing value imputation and outlier detection and removal ensure data integrity and reliability, preventing non-faulty data from contaminating model judgments. In industrial settings, occasional factors such as intermittent sensor communication interruptions and electromagnetic interference are unavoidable, often resulting in missing values ​​and spike noise in the raw data. Directly inputting data containing missing values ​​into the model would prevent the network from performing convolution and time-series calculations correctly; misjudging spike noise as fault features would lead to misdiagnosis. This step uses linear imputation to appropriately fill in missing values, maintaining the continuous and gradual characteristics of the signal waveform; it employs the 3σ criterion to detect and remove outliers whose amplitudes significantly deviate from the mean, using weighted imputation based on effective values ​​from adjacent time points to eliminate the damage to the signal morphology caused by occasional interference. After these processes, the data sequence is restored to a complete and smooth state, providing a clean and reliable input for subsequent feature extraction.

[0105] Standardization eliminates differences in multiple physical dimensions, ensuring fair treatment of signals from all channels during model training. Vibration amplitude, pressure, flow rate, and temperature signals have completely different physical dimensions and numerical ranges. If directly input into the network, signals with larger values ​​will dominate the gradient update of the loss function, suppressing the effective feature representation of signals with smaller values, causing the model to favor learning one type of signal while ignoring others. This step standardizes the data from each branch using Z-scores, mapping them uniformly to a standard normal distribution space with a mean of 0 and a standard deviation of 1, ensuring all signals are on the same numerical order of magnitude. After standardization, the model can treat each type of sensor information equally, avoiding the negative impact of dimensional differences on model training and diagnostic accuracy, and ensuring the deep fusion effect of multi-source heterogeneous data.

[0106] The training process of the convolutional neural network and the bidirectional long short-term memory network includes the following steps: Step T1: Collect multi-source time-series data samples of various scenarios and pump models under normal and various fault conditions. Label each sample with fault type and operating condition type labels to form a training dataset. By collecting multi-source time-series data of various scenarios and pump models under normal and various fault conditions, and simultaneously labeling them with fault type and operating condition type labels, a diverse training dataset covering a wide range of operating condition and fault spaces is constructed. The construction of this multi-source heterogeneous dataset allows the model to be exposed to the differences in data distribution under different equipment models and operating conditions during the pre-training stage, laying a data foundation for subsequent learning of common fault features independent of operating conditions. At the same time, the introduction of operating condition type labels provides supervision signals for multi-task learning, which is not available in traditional single-task pre-training datasets, fundamentally supporting the effective implementation of subsequent operating condition identification auxiliary tasks.

[0107] Step T2: Input the training dataset into a coupled model composed of a convolutional neural network and a bidirectional long short-term memory network in parallel. Set the loss function for the main task of fault classification and the loss function for the auxiliary task of working condition identification. Pre-train using a multi-task joint optimization strategy, iteratively update the network parameters through backpropagation until the loss functions of the main and auxiliary tasks converge, and save the global universal model parameters θg. Pre-training is performed using a multi-task joint optimization strategy for the main task of fault classification and the auxiliary task of working condition identification, with the introduction of the auxiliary task of working condition identification being a core technical approach. During the optimization process, the model is forced to learn both "fault type discrimination" and "working condition type discrimination" simultaneously, thereby enabling the underlying network parameters to form a common fault feature extraction capability that is insensitive to changes in working conditions. That is, the model learns to distinguish between "differences in the manifestation of the same fault under different working conditions" and "essential feature differences between different faults." Compared with single-task pre-training, multi-task learning effectively decouples working condition differences from fault differences, giving the globally universal model parameters θg obtained through pre-training stronger cross-scene transfer capabilities. Meanwhile, the convergence of the two loss functions serves as the criterion for determining the completion of pre-training, ensuring that the model reaches a stable state on both the main and auxiliary tasks, thus avoiding overfitting of the working condition features caused by optimizing only the main task.

[0108] In this embodiment, the loss function for both the fault classification main task and the operating condition identification auxiliary task is the cross-entropy loss function.

[0109] The input to the main task of fault classification is the fault category prediction probability vector output by the model and the true fault label of the sample. Its loss function is expressed as: ; Where M is the total number of fault categories, The true probability that a sample belongs to the i-th type of fault (when the true category of the sample is the i-th type). = 1, otherwise = 0), This is the predicted probability of the i-th type of fault output by the model.

[0110] The input to the working condition identification auxiliary task is the working condition category prediction probability vector output by the model and the true working condition label of the sample. Its loss function is expressed as: Where N is the total number of operating condition categories. The true probability that a sample belongs to the j-th type of work condition (when the true category of the sample is the j-th type). = 1, otherwise = 0), This is the predicted probability of the j-th type of working condition output by the model.

[0111] The total loss function is the weighted sum of the loss from the primary task of fault classification and the loss from the auxiliary task of condition identification, expressed as: in, This is the auxiliary task weight coefficient, used to balance the relative importance of the main task and auxiliary tasks in joint optimization. The value ranges from 0.1 to 0.5, with a preferred value of 0.3. By setting an auxiliary task for identifying operating conditions, the coupled model is forced to learn to distinguish the differences in data distribution under different operating conditions during the pre-training process, thereby enabling the underlying network parameters to form the ability to extract common fault features that are insensitive to changes in operating conditions.

[0112] Step T3, Adaptive Fine-tuning under Working Conditions: Obtain small sample labeled data of the target scene, load the global general model parameters θg, freeze the parameters of the first few layers of the network, and fine-tune the parameters of the last few layers of the network using the small sample data of the target scene to obtain the scene fine-tuning parameters θs, thus completing the adaptation and deployment of the model under the target working conditions. A lightweight adaptive working condition strategy is adopted, which freezes the parameters of the first few layers of the network and only fine-tunes the parameters of the last few layers of the network. The frozen shallow network retains the general waveform feature extraction and temporal dependency capture capabilities learned in the pre-training stage, avoiding damage to these general capabilities when fine-tuning on a limited number of small samples; the fine-tuned high-level network quickly adapts to the fault semantic features of specific equipment models through a small number of target scene samples. This strategy enables the model to complete working condition adaptation with only a few iterations when facing new scenes with only a few labeled samples, without the need for retraining from scratch or large-scale retraining. This fundamentally solves the pain point of poor diagnostic performance in small sample scenarios caused by the scarcity of fault samples and high labeling costs in industrial sites, while significantly reducing the computational and time costs of model deployment.

[0113] This invention also relates to a deep learning-based water pump fault identification and diagnosis system, comprising: The multi-source data acquisition unit 1 is used to acquire multi-source time-series data output by various types of sensors during the operation of the water pump in real time. The multi-source time-series data includes at least vibration signals, pressure signals, flow signals, and temperature signals. Data preprocessing unit 2 is connected to the multi-source data acquisition unit. It is used to perform missing value interpolation, outlier detection and removal, timestamp alignment and normalization on the acquired multi-source time series data. Then, it uses pressure signal and flow signal as steady-state reference signals to perform stepwise coupling correction on vibration signal and temperature signal, and outputs coupled preprocessed data. The coupled feature extraction unit 3 is connected to the data preprocessing unit and includes a parallel spatial feature extraction subunit and a temporal feature extraction subunit. The spatial feature extraction subunit uses a convolutional neural network to extract fault spatial distortion features step by step through multi-level convolutional pooling operations and outputs a spatial feature set. The temporal feature extraction subunit uses a bidirectional long short-term memory network to traverse the temporal fluctuation pattern of the coupled preprocessed data in both forward and backward directions and outputs a temporal feature set. Feature fusion and fault classification unit 4 is connected to the coupled feature extraction unit, and is used to fuse the spatial feature set and the temporal feature set, perform fault identification and classification on the pump operating status based on the fused coupled features, and output fault diagnosis results. The full-process closed-loop management unit 5 is connected to the feature fusion and fault classification unit, and is used to perform fault location and source tracing, operation trend prediction and / or intelligent operation and maintenance decision-making based on the fault diagnosis results, forming a full-process closed-loop system for water pump fault identification and diagnosis.

[0114] The multi-source data acquisition unit, through acceleration sensors, pressure sensors, flow sensors, and temperature sensors deployed at key locations on the pump body, simultaneously and in real-time acquires four types of signals: vibration, pressure, flow, and temperature. The collaborative configuration of these multi-source heterogeneous sensors enables the system to comprehensively perceive the pump's operating status from different physical dimensions, providing a rich data foundation for subsequent diagnostics. The different sensors form a redundant and complementary relationship; even if one sensor fails, the others can still maintain basic monitoring functions, improving the system's fault tolerance.

[0115] The data preprocessing unit sequentially performs timestamp alignment and resampling, missing value imputation, outlier detection and removal, standardization, and cascade coupling correction. Timestamp alignment and resampling eliminate inconsistencies in sampling frequencies caused by hardware differences among sensors, ensuring point-to-point correspondence of the four signal types in the time dimension, providing a unified time reference for subsequent correction and feature extraction. Missing value imputation uses a linear imputation method to reasonably fill in missing values ​​caused by communication interruptions, avoiding breakpoint problems due to missing data. Outlier detection and removal uses the 3σ criterion to identify spike noise caused by occasional factors such as electromagnetic interference. After removal, it is filled using a sliding weighted average of effective values ​​from adjacent time points, eliminating interference from non-faulty anomalies to subsequent feature extraction. Standardization maps all physical dimensions to the same numerical range, avoiding the problem of one dimension's value being too large during training and suppressing the effective information of other dimensions. The cascade coupling correction uses pressure and flow rate as steady-state benchmarks and adaptively determines the weight of each branch through the Pearson correlation coefficient. This results in higher weights for effective signals with strong physical correlations and suppression of abnormal disturbance signals that deviate from the physical coupling relationship, thus improving the input quality from the data source.

[0116] The coupled feature extraction unit comprises a parallel spatial feature extraction subunit and a temporal feature extraction subunit. The spatial feature extraction subunit employs a three-level convolutional pooling hierarchical structure, progressively extracting shallow edge textures, mid-level local distortions, and deep global semantic fault features, comprehensively characterizing the spatial domain representation of pump faults from coarse to fine and from local to global perspectives. The temporal feature extraction subunit uses a bidirectional long short-term memory network (LSTM) to traverse temporal data in both forward and backward directions. The forward LSTM traces the fault's initiation and evolution from history to the present, while the backward LSTM predicts the fault's development trend from the endpoint. The two subunits operate in parallel and synchronously without interference, avoiding the cumulative propagation of errors from earlier stages to later stages in a cascaded structure, thus achieving true synchronous joint representation of spatial and temporal domain features.

[0117] The feature fusion and fault classification unit dynamically determines the spatial feature weight α and temporal feature weight β based on the signal quality assessment results of the current input signal, ensuring that the fused features are always dominated by the most reliable features under the current operating conditions. Under steady-state conditions, the weights are evenly distributed, fully utilizing the local distortion capture capability of CNN and the global degradation tracking capability of BiLSTM. Under transient conditions or sensor anomalies, the system automatically reduces the weight of the affected branch and increases the contribution of the other branch, effectively suppressing the negative impact of operating condition fluctuations and sensor drift on the diagnostic results. This adaptive weighting mechanism ensures the reliability of the fused features under various operating conditions.

[0118] The described closed-loop management unit locates faulty components and their starting time points based on spatial feature activation maps and temporal feature attention distributions. It predicts remaining service life based on current diagnostic results and historical degradation trajectories, and outputs an executable maintenance plan based on a knowledge base matching fault type, severity, and remaining service life. The entire process requires no manual intervention, filling the gap in existing technologies that are limited to fault identification and lack systematic solutions.

[0119] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for identifying and diagnosing pump faults, characterized in that, Includes the following steps: Step S1: Acquire multi-source time-series data during the operation of the water pump. The multi-source time-series data includes at least vibration signal, pressure signal, flow signal, and temperature signal. Using the pressure signal and flow signal as steady-state reference signals, perform stepwise coupling correction on the vibration signal and temperature signal to achieve physical alignment and coupling enhancement of multi-source heterogeneous data, and output coupled preprocessed data. Step S2: Input the coupled preprocessed data into the parallel spatial feature extraction branch and temporal feature extraction branch respectively, and extract the fault spatial distortion features and fault temporal degradation features simultaneously; The spatial feature extraction branch uses a convolutional neural network to extract fine-grained fault spatial features from shallow, middle and deep layers through multi-level convolutional pooling operations, and outputs a spatial feature set. The temporal feature extraction branch adopts a bidirectional long short-term memory network, which traverses the temporal fluctuation pattern of the coupled preprocessed data in both forward and reverse directions, captures the temporal dependency features of long-term fault degradation, and outputs a temporal feature set after splicing forward and reverse features. Step S3: Fuse the spatial feature set with the temporal feature set, and perform fault identification and classification on the pump operating status based on the fused coupled features, and output the fault diagnosis results; Step S4: Based on the fault diagnosis results of step S3, perform fault location and source tracing, operation trend prediction and intelligent operation and maintenance decision-making to form a closed-loop system for the entire process of water pump fault identification and diagnosis.

2. The deep learning-based water pump fault identification and diagnosis method according to claim 1, characterized in that, The step-by-step coupling correction in step S1 specifically includes: The raw time-series signal acquired by the i-th type of sensor Where i = 1, 2, 3, 4 correspond to vibration signal, pressure signal, flow signal, and temperature signal, respectively. The coupled preprocessed data is calculated according to the following formula. : ; Among them, the coupling weights of each branch signal The signal from this branch and the pressure reference signal Flow reference signal The Pearson correlation coefficient is adaptively determined to achieve deep coupling correction of multi-source heterogeneous data.

3. The deep learning-based water pump fault identification and diagnosis method according to claim 1, characterized in that, The spatial feature extraction branch employs a three-level convolutional pooling tiered structure to extract fine-grained features of the fault space step by step, and the coupled preprocessed data output from step S1 is denoted as the input of layer 0. Specifically, the features of the k-th layer are calculated step by step according to the following formula: in, The weights of the k-th layer convolutional kernel are... The feature map output from layer (k-1) is downsampled after each convolutional layer by pooling. The three-level convolutional-pooling structure extracts shallow edge texture fault features, mid-level local distortion fault features, and deep global semantic fault features, respectively. The outputs of each level are concatenated to form the spatial feature set. .

4. The deep learning-based water pump fault identification and diagnosis method according to claim 1, characterized in that, The temporal feature extraction branch uses a bidirectional long short-term memory network to extract temporal features, and the coupled preprocessed data output from step S1 is denoted as... Calculate the forward and backward timing characteristics using the following formulas. feature: ; ; forward temporal features With reverse time series characteristics The data is spliced ​​together to form a temporal feature set. : ; in, This represents the forward LSTM computation process. This represents the reverse LSTM computation process. This represents a vector concatenation operation, resulting in a temporal feature set. It fully covers the timing fluctuations and degradation patterns of water pumps during operation.

5. The deep learning-based water pump fault identification and diagnosis method according to claim 1, characterized in that, In step S3, the specific steps of feature fusion include: For the spatial feature set and the time series feature set Adaptive correction of feature weights is performed separately, and the spatial feature weights are dynamically adjusted based on the signal quality assessment results of the current input signal. and time series feature weights The fusion features are calculated according to the following formula: ;in, = 1, and The value of is dynamically determined by the signal quality assessment results under the current operating conditions.

6. The deep learning-based water pump fault identification and diagnosis method according to claim 1, characterized in that, Before performing step S2, the process also includes model pre-training and condition-adaptive fine-tuning transfer steps: The convolutional neural network and the bidirectional long short-term memory network are pre-trained using operational datasets of multiple scenarios and various types of water pumps to obtain globally universal model parameters. ; When applied to a target scenario, the global general model is adaptively adjusted using small sample data from the target scenario. This involves freezing the parameters of the first few network layers and adjusting only the parameters of the last few layers to obtain scenario-fine-tuning parameters. This allows the model to adapt to the fault diagnosis requirements of the current target operating conditions. The model pre-training employs a multi-task learning strategy, simultaneously optimizing the main task of pump fault classification and the auxiliary task of operating condition type identification to enhance the model's feature transfer capability across different operating conditions.

7. The deep learning-based water pump fault identification and diagnosis method according to claim 1, characterized in that, The closed-loop system in step S4 includes: S4-1. Deploy various sensors to collect multi-source time-series data of water pump operation in real time; S4-2. Perform missing value imputation, outlier detection and removal, normalization, and multi-source data cascade coupling correction on the collected raw data. S4-3. Deploy the parallel dual-branch feature extraction coupled model of the trained convolutional neural network and bidirectional long short-term memory network, and output the fault identification classification results and fault probability confidence. S4-4. Based on the spatial feature activation map and temporal feature attention distribution output by the coupling model, locate the physical component location and start time of the fault, and output a fault tracing report. S4-5. Based on the current diagnostic results and historical degradation trajectory, predict the evolution trend of the pump's operating status and output the predicted value of its remaining service life; S4-6. Based on the fault type, fault severity, and remaining service life prediction results, match the preset operation and maintenance strategy knowledge base, and output an operation and maintenance decision scheme including maintenance priority, maintenance measure suggestions, and spare parts replacement plan.

8. The deep learning-based water pump fault identification and diagnosis method according to claim 1, characterized in that, A data preprocessing step is set before step S1: The original multi-source time-series data is timestamped and resampled to ensure that the signals from various sensors have a uniform sampling frequency and time window length. Missing values ​​are filled using linear interpolation. Outliers exceeding a preset threshold are detected and removed. The vacant positions after removal are filled with valid values ​​from adjacent time points. The data of each branch that has been filled in are standardized so that the signals of each physical dimension are uniformly mapped to the same numerical range.

9. The deep learning-based water pump fault identification and diagnosis method according to any one of claims 1 to 8, characterized in that, The training process of the convolutional neural network and the bidirectional long short-term memory network includes the following steps: Step T1: Collect multi-source time-series data samples of multiple scenarios and various models of water pumps under normal operating conditions and various fault conditions, and label each sample with fault type label and operating condition type label to form a training dataset; Step T2: Input the training dataset into the coupled model composed of a convolutional neural network and a bidirectional long short-term memory network in parallel, set the loss function of the main task of fault classification and the loss function of the auxiliary task of working condition identification, pre-train using a multi-task joint optimization strategy, iteratively update the network parameters through the backpropagation algorithm until the loss functions of the main task and the auxiliary task converge, and save the global general model parameters θg. Step T3, adaptive fine-tuning under working conditions: Obtain small sample labeled data of the target scene, load the global general model parameters θg, freeze the parameters of the first few layers of the network, fine-tune the parameters of the last few layers of the network with the small sample data of the target scene, obtain the scene fine-tuning parameters θs, and complete the adaptation and deployment of the model under the target working conditions.

10. A water pump fault identification and diagnosis system based on deep learning, characterized in that, include: A multi-source data acquisition unit is used to acquire multi-source time-series data output by various types of sensors during the operation of the water pump in real time. The multi-source time-series data includes at least vibration signals, pressure signals, flow signals, and temperature signals. The data preprocessing unit is connected to the multi-source data acquisition unit. It is used to perform missing value interpolation, outlier detection and removal, timestamp alignment and normalization on the acquired multi-source time series data. Then, it uses the pressure signal and flow signal as steady-state reference signals to perform stepwise coupling correction on the vibration signal and temperature signal, and outputs coupled preprocessed data. A coupled feature extraction unit, connected to the data preprocessing unit, includes a parallel spatial feature extraction subunit and a temporal feature extraction subunit. The spatial feature extraction subunit uses a convolutional neural network to extract fault spatial distortion features step by step through multi-level convolutional pooling operations and outputs a spatial feature set. The temporal feature extraction subunit uses a bidirectional long short-term memory network to traverse the temporal fluctuation patterns of the coupled preprocessed data in both forward and backward directions and outputs a temporal feature set. The feature fusion and fault classification unit is connected to the coupled feature extraction unit and is used to fuse the spatial feature set and the temporal feature set, perform fault identification and classification on the pump operating status based on the fused coupled features, and output fault diagnosis results. The full-process closed-loop management unit is connected to the feature fusion and fault classification unit, and is used to perform fault location and source tracing, operation trend prediction and / or intelligent operation and maintenance decision-making based on the fault diagnosis results, forming a full-process closed-loop system for water pump fault identification and diagnosis.

Citation Information

Patent Citations

  • Fault classification diagnosis method for emulsion pump

    CN113378887A

  • Deep learning classification model for fault diagnosis and fault diagnosis method

    CN113869339A

  • Method and device for intelligent diagnosis of important service water pump

    CN114548154A