Vertical centrifugal pump motor non-driving end bearing noise reduction method based on deep learning

Abstract

The invention relates to the technical field of bearing noise, in particular to a vertical centrifugal pump motor non-driving end bearing noise reduction method based on deep learning, and the method comprises the steps: collecting multi-source mixed noise through an electrical sensor disposed on a bearing seat, and constructing a bearing coherence graph representing a signal relation; on an edge computing device, a blind source decoupling algorithm integrating learnable edge pruning and self-supervised graph contrast learning is adopted, adaptive sparsification and characterization learning are carried out on the coherence graph, and pure bearing vibration signals are separated out; inputting the bearing vibration signal into a deep learning model for online analysis, predicting the remaining effective life of the bearing, and identifying key harmonic components; and a dual noise reduction strategy is executed, the resonance area is avoided by adjusting the rotating speed of the motor, compensation harmonic waves are actively injected into the frequency converter to generate reverse-phase torque ripples, and targeted suppression of noise is achieved. Noise reduction is carried out on the non-driving end bearing of the vertical centrifugal pump motor through the deep learning method.

Description

Technical Field

[0001] This invention relates to the field of bearing noise technology, specifically a method for noise reduction of the non-drive end bearing of a vertical centrifugal pump motor based on deep learning. Background Technology

[0002] Existing vibration and acoustic testing methods for bearings at the non-drive end of vertical centrifugal pump motors have limitations. Traditional signal processing techniques, such as wavelet transform and empirical mode decomposition, rely on manually set thresholds and prior knowledge when extracting early fault features from complex mixed signals. When faced with non-stationary operating conditions involving the coupling of multiple sources of signals, such as mechanical vibration, electromagnetic effects, and environmental noise, these methods struggle to effectively decouple and extract weak bearing fault feature signals, easily leading to distortion of diagnostic information or loss of key features. Therefore, accurately separating and identifying bearing fault signals from a strong noise background is a current technical challenge in the field of equipment condition testing.

[0003] With the application of deep learning technology in the field of dynamic system testing and diagnosis, a new approach has been provided to solve the above problems. Deep learning models can automatically learn and extract deep features related to faults directly from raw vibration or acoustic data in an end-to-end manner, without the need for manually designing complex feature extractors. It can effectively model and distinguish between effective signals representing the health status of bearings and background noise, thereby significantly improving the accuracy and robustness of fault diagnosis and making it possible to achieve accurate equipment condition assessment and predictive maintenance.

[0004] To address this, a noise reduction method for the non-drive end bearing of a vertical centrifugal pump motor based on deep learning is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a noise reduction method for the non-drive end bearing of a vertical centrifugal pump motor based on deep learning, which uses deep learning methods to reduce noise in the non-drive end bearing of the vertical centrifugal pump motor.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A deep learning-based method for noise reduction of the non-drive end bearings of a vertical centrifugal pump motor includes:

[0008] An electrical sensor is deployed in the bearing housing at the non-drive end of a vertical centrifugal pump motor to collect bearing mixed noise and perform time-domain alignment to generate a bearing mixed noise signal. A bearing coherence diagram is constructed based on the characterization relationship of the bearing mixed noise signal.

[0009] The edge computing device constructs a bearing coherence graph and performs a self-supervised blind source decoupling algorithm on the bearing coherence graph, the self-supervised blind source decoupling algorithm includes an integrated learnable edge pruning module to perform online adaptive sparsification on the bearing coherence graph to generate a bearing strong coherence graph, and a self-supervised graph contrast learning model is used to separate noise from the bearing strong coherence graph by using a constructed positive sample pair and a negative sample pair to generate a bearing vibration signal.

[0010] The bearing vibration signal is input into a deep learning model for online learning to predict the remaining useful life, identify key harmonic components, and perform a double noise reduction strategy based on the key harmonic components, including adjusting the motor speed to avoid a high noise resonance region and injecting a compensation harmonic into the motor frequency converter to reduce noise through reverse torque pulsation.

[0011] The bearing mixed noise collection process is as follows:

[0012] An electrical sensor is arranged on the non-driven end bearing seat of the vertical centrifugal pump motor, the electrical sensor includes a voltage sensor, a current sensor, and an electromagnetic induction sensor, and the motor stator current signal, the stator voltage signal, and the electromagnetic induction signal of the non-driven end bearing of the vertical centrifugal pump motor are collected respectively and sampled synchronously and aligned in time domain to generate a bearing mixed noise signal containing mechanical noise, electromagnetic noise, and fluid noise.

[0013] The synchronous sampling and time domain alignment eliminate the phase difference and sampling delay of signals collected by different sensors through a time offset correction method based on a cross-correlation function.

[0014] The specific process of constructing a bearing coherence graph based on the bearing noise signal is as follows:

[0015] The amplitude spectrum, the phase spectrum, and the envelope feature are extracted based on the bearing mixed noise signal; the amplitude spectrum and the phase spectrum are extracted through fast Fourier transform, and the envelope feature is extracted through Hilbert transform based on the amplitude spectrum and the phase spectrum.

[0016] The cross-correlation coefficient and the coherence function between the electrical sensors are calculated based on the amplitude spectrum, the phase spectrum, and the envelope feature, the amplitude similarity, the phase consistency, and the envelope correlation are generated and fused into edge weights; the mechanical noise, the electromagnetic noise, and the fluid noise are taken as nodes, and the fused edge weights are taken as weighted edges to construct a bearing coherence graph.

[0017] The edge computing device constructs a bearing coherence graph and performs a self-supervised blind source decoupling algorithm on the bearing coherence graph, the self-supervised blind source decoupling algorithm includes an integrated learnable edge pruning module to perform online adaptive sparsification on the bearing coherence graph to generate a bearing strong coherence graph, and a self-supervised graph contrast learning model is used to separate noise from the bearing strong coherence graph by using a constructed positive sample pair and a negative sample pair to generate a bearing vibration signal.

[0018] A self-supervised blind source decoupling algorithm is deployed in an edge computing device close to the vertical centrifugal pump motor, the self-supervised blind source decoupling algorithm integrates a learnable edge pruning module and a self-supervised graph contrast learning model to separate mechanical noise, electromagnetic noise, and fluid noise.

[0019] The separation process separates the corresponding noise components by class by identifying the different edge weight distribution of mechanical vibration features, electromagnetic induction features and fluid disturbance features in the graph structure, and finally obtains the bearing vibration signal related to the bearing operation state.

[0020] The self-supervised blind source decoupling is specifically:

[0021] The learnable edge pruning module is used to assign adjustable weights to each edge of the bearing coherence graph, and dynamically reduce the low correlation edge weight through an online gradient update algorithm, to adaptively sparsify the bearing coherence graph and generate a bearing strong coherence graph;

[0022] The self-supervised graph contrast learning model learns the node representation of the bearing strong coherence graph by constructing positive sample pairs and negative sample pairs, so that the embedding vectors of the nodes in the positive sample pairs are close in the feature space, and the embedding vectors of the nodes in the negative sample pairs are far away, to separate the bearing mixed noise signal and generate a bearing vibration signal.

[0023] The process of inputting the bearing vibration signal into a deep learning model for online learning to predict the remaining useful life and identifying the key harmonic components is specifically:

[0024] The deep learning model includes a convolutional neural network and a long short-term memory network, which are used to extract time domain features, frequency domain features and time-dependent features of the bearing vibration signal at the same time;

[0025] Through an online learning mechanism, the deep learning model uses historical vibration data and real-time collected bearing vibration signals for incremental training, and outputs a bearing remaining useful life prediction value;

[0026] Based on the prediction result and vibration signal spectrum analysis, the key harmonic components related to the bearing degradation state are identified.

[0027] The double noise reduction strategy is:

[0028] The double noise reduction strategy includes a first noise reduction strategy and a second noise reduction strategy;

[0029] The first noise reduction strategy is to adjust the motor speed to dynamically avoid the high noise resonance region by real-time monitoring of the bearing vibration signal and the key harmonic component, and to reduce the coupling interference of mechanical noise and fluid noise;

[0030] The second noise reduction strategy is to inject a compensation harmonic signal into the motor frequency converter, to generate an anti-phase torque ripple opposite in phase to the key harmonic component, to actively cancel the electromagnetic noise and residual mechanical vibration, and to reduce the bearing mixed noise.

[0031] The injection compensation harmonic signal specifically processes is:

[0032] The edge computing device calculates the corresponding anti-phase harmonic amplitude and phase according to the amplitude and phase information of the key harmonic component identified by the deep learning model in real time.

[0033] The anti-phase harmonic amplitude and phase are injected into the motor stator through the motor frequency converter, so that the generated anti-phase torque pulsation and the original key harmonic component are mutually canceled in phase.

[0034] Compared with the prior art, the beneficial effects of the present application are:

[0035] 1. The present application converts complex timing problems into structured graph problems by deploying multiple electrical sensors and constructing a signal coherence graph; using a self-supervised blind source decoupling algorithm deployed on an edge computing device, through learnable edge pruning and graph contrast learning, the method can automatically and without human intervention separate the vibration signal strongly related to the bearing state from the mixed noise of mechanical, electromagnetic and fluid sources, which overcomes the limitations of traditional signal processing methods that rely on manual feature design and are difficult to process non-stationary signals, significantly improving the separation accuracy of the target signal and the adaptive ability of the algorithm.

[0036] 2. The present application not only focuses on noise reduction, but also uses the separated pure bearing vibration signal for deep analysis, through the constructed convolutional neural network and long short-term memory network deep learning model, the time, frequency domain and time-dependent features of the signal can be extracted at the same time, and combined with the online incremental training mechanism, not only the key harmonic component directly related to the bearing degradation state can be accurately identified, but also the remaining useful life of the bearing can be predicted in real time, so that the method is upgraded to a device health management method with diagnosis and prediction functions.

[0037] 3. Based on the key harmonics identified by the deep learning model, the present application implements an innovative double noise reduction strategy, on the one hand, by adjusting the motor speed to avoid the resonance region, to reduce the coupled noise from a macroscopic point of view; on the other hand, by generating anti-phase harmonics in real time through the edge computing device, and injecting them into the motor through the frequency converter to generate anti-phase torque pulsation, to achieve accurate active cancellation of key electromagnetic and mechanical noise. This closed-loop control method combining avoidance and cancellation is more targeted, more effective and more adaptable than single passive vibration isolation or active noise reduction methods. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flowchart of the deep learning-based vertical centrifugal pump motor non-driven end bearing noise reduction method;

[0039] Figure 2 A self-supervised blind source decoupling algorithm structure diagram for a deep learning-based vertical centrifugal pump motor non-drive end bearing noise reduction method;

[0040] Figure 3 A flowchart for bearing vibration signal noise reduction of the deep learning-based vertical centrifugal pump motor non-drive end bearing noise reduction method. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0042] Embodiment one:

[0043] Please refer to Figure 1 The present application provides a deep learning-based vertical centrifugal pump motor non-drive end bearing noise reduction method, and the technical solutions are as follows:

[0044] The deep learning-based vertical centrifugal pump motor non-drive end bearing noise reduction method comprises:

[0045] An electrical sensor is deployed on the non-drive end bearing seat of the vertical centrifugal pump motor to collect bearing mixed noise and perform time domain alignment to generate a bearing mixed noise signal, and a bearing coherence map is constructed based on the representation relationship of the bearing mixed noise signal;

[0046] An edge computing device is constructed to execute a self-supervised blind source decoupling algorithm on the bearing coherence map, the self-supervised blind source decoupling algorithm comprising an integrated learnable edge pruning module for online adaptive sparsification of the data coherence map to generate a bearing strong coherence map, and through a self-supervised graph comparison learning model, using the constructed positive sample pair and negative sample pair, the bearing strong coherence map is subjected to noise separation to generate a bearing vibration signal;

[0047] The bearing vibration signal is input into a deep learning model for online learning to predict the remaining useful life, identify the key harmonic components, and execute a double noise reduction strategy based on the key harmonic components, including adjusting the motor speed to avoid the high noise resonance region and injecting compensation harmonics into the motor frequency converter, and reducing noise through anti-phase torque pulsation.

[0048] In this embodiment, the data acquisition object is a vertical centrifugal pump driven by a frequency converter, and a set of electrical sensors is deployed near the bearing seat of the non-driven end of the motor. The sensor set includes a Hall effect clamp current sensor for non-invasively acquiring the stator current signal of any phase of the motor; a Hall effect voltage sensor for monitoring the input voltage waveform of the stator winding; and a high-sensitivity electromagnetic induction sensor in close contact with the bearing seat shell for picking up the local magnetic field fluctuations caused by the bearing in operation;

[0049] During data acquisition, the sampling mode is set to synchronous sampling to ensure that the signals of the three channels of voltage, current and electromagnetic induction are collected at the same time at the hardware level, thereby maximizing the elimination of scanning delays between channels. To capture the high-frequency features that may be generated by bearing failure, the sampling frequency is set to 50 kHz, and the sampling accuracy is 24 bits, to ensure sufficient dynamic range and signal resolution;

[0050] Based on the motor stator current signal, the stator voltage signal and the electromagnetic induction signal of the non-driven end bearing of the motor of the vertical centrifugal pump, synchronous sampling and time domain alignment are performed. The time offset correction method based on the cross-correlation function eliminates the phase difference and sampling delay of signals collected by different sensors. With the stator current signal having the highest signal-to-noise ratio as the reference, the cross-correlation function between the voltage signal, the electromagnetic induction signal and the current signal is calculated to determine the time delay of each signal.

[0051] According to the sample point offset corresponding to the peak position of the cross-correlation function, the voltage and electromagnetic induction signal are translated and compensated. After this alignment step, a multi-dimensional data matrix is finally obtained, in which each column represents the stator current, stator voltage and electromagnetic induction signal that have been accurately aligned on the time reference, i.e. the bearing mixed noise signal, which contains mechanical noise, electromagnetic noise and fluid noise.

[0052] By jointly deploying current, voltage and electromagnetic induction sensors, the coupling information related to the bearing operating state is successfully captured from multiple physical dimensions, which is much more comprehensive and rich than single sensor data. At the same time, the combination of hardware synchronous sampling and software fine alignment based on the cross-correlation function ensures high time synchronization accuracy and data fidelity between multiple signals, providing a high-quality, phaseless data basis for subsequent precise separation of mechanical, electromagnetic and fluid mixed noise components and implementation of high-reliability fault diagnosis.

[0053] Based on the internal representation relationship of the bearing mixed noise signal;

[0054] Each time sequence of the acquired stator current, stator voltage and electromagnetic induction signal is segmented, for example, the signal is divided into multiple data frames containing 4096 sampling points and having an overlap rate of 50%. A Hanning window is applied to each data frame to suppress spectral leakage; feature extraction is performed on each windowed data frame, each frame of signal is converted from time domain to frequency domain by applying a fast Fourier transform to obtain a corresponding complex spectrum; an amplitude spectrum, i.e. a modulus of a complex number, and a phase spectrum, i.e. an argument of a complex number, are extracted based on the complex spectrum; at the same time, a Hilbert transform is applied to the original each time domain data frame to construct an analytic signal thereof, and a modulus of the analytic signal is taken to obtain an instantaneous envelope feature of the signal;

[0055] The amplitude similarity between any two sensor signal pairs is quantified by calculating a cross-correlation coefficient between the amplitude spectra corresponding to the sensor signal pairs, and the phase consistency between any two sensor signal pairs is quantified by calculating a mean value of an amplitude square coherence function in a main frequency band; the envelope correlation between any two sensor signal pairs is quantified by calculating a Pearson correlation coefficient between the envelope signals corresponding to the sensor signal pairs, and these calculations are performed on all data frames and averaged; by quantifying the relationship between signals in three dimensions of amplitude, phase and envelope, a comprehensive and in-depth analysis of dynamic correlation is achieved, and statistical robustness of the quantification indicators is ensured by frame averaging and other processing techniques, thereby providing a highly reliable data basis for constructing an accurate coherence graph model;

[0056] Three nodes of a graph are defined, representing mechanical noise, electromagnetic noise and fluid noise, and the amplitude similarity, phase consistency and envelope correlation calculated in the previous step are fused by weighting, and the embodiment is set to 0.3, 0.4 and 0.3, forming a single fused edge weight; this weight is assigned between nodes representing the correlation of corresponding physical processes, for example, the weight calculated from the current and voltage signals defines an edge between the electromagnetic noise node and the fluid noise node, and a bearing coherence graph containing three nodes and three weighted edges is constructed to intuitively represent the interaction strength between different noise sources in a structured manner;

[0057] The bearing coherence graph constructed in this embodiment abstracts and converts the high-dimensional and unstructured time sequence signal relationship into a low-dimensional and structured graph model, the nodes represent physical noise sources, the weighted edges fused with multi-dimensional features quantify the interaction strength between the sources, and a direct and system-level perspective is provided to understand the complex noise coupling mechanism; this structured data expression makes the complex dynamic relationship machine-readable, providing an ideal input and a solid foundation for subsequent application of graph algorithms for precise noise source decoupling.

[0058] A self-supervised blind source decoupling algorithm is executed on the edge computing device based on the bearing coherence graph, referring to Figure 2, the self-supervised blind source decoupling algorithm is deployed on an edge computing unit (for example, NVIDIA Jetson AGXXavier) close to the vertical centrifugal pump, the edge computing unit has GPU acceleration capability and can meet the real-time operation requirements of the graph neural network model, and a core of the self-supervised blind source decoupling algorithm is an encoder based on a graph attention network; the encoder is integrated with a learnable edge pruning module and a self-supervised graph contrastive learning framework;

[0059] The input bearing coherent graph is dynamically optimized by the learnable edge pruning module, the learnable edge pruning module is essentially a small neural network, an importance weight (between 0 and 1) is calculated for each edge according to the initial features of the nodes in the graph, for an edge connecting any two nodes, the initial feature vectors of the two nodes are spliced to form a combined feature vector with doubled dimensions, which completely contains all the initial information of the two nodes connected by the edge; the combined feature vector is input into a multilayer perceptron, the multilayer perceptron includes a hidden layer with a ReLU activation function and an output layer, the output layer has only one neuron, and uses a Sigmoid activation function to map its input to a weight between 0 and 1;

[0060] The weight is multiplied by the original fusion edge weight to dynamically adjust the connection strength of the edges in the graph, and in the training process, the model learns to prune, that is, to weaken or even ignore edges irrelevant to the target task, so that the graph attention network can focus more on the graph structure representing the key noise source coupling relationship, and generate a sparse bearing strong coherent graph;

[0061] The graph attention network encoder is trained in a self-supervised graph contrastive learning manner to learn effective feature representations of each noise source node; two different random data augmentations (for example, random edge dropping and node feature masking) are performed on the input strong coherent graph to construct two associated but different views; for the same noise source node, its representations in the two views constitute a positive sample pair; and the representations of different nodes constitute a negative sample pair; the goal of the self-supervised graph contrastive learning is to optimize the contrastive loss function to make the vector representations of the positive sample pair in the feature space as similar as possible, and the vector representations of the negative sample pair as far away as possible;

[0062] After self-supervised pre-training, the graph attention network encoder generates highly discriminative feature vectors for the three nodes of mechanical noise, electromagnetic noise and fluid noise, performs noise separation and reconstruction of the target signal; using the trained model to identify the edge weight distribution and feature pattern dominated by electromagnetic effect and fluid disturbance in the graph, and using a decoder (for example, a multi-layer perception) to reconstruct the respective noise components from the original mixed noise signal according to the feature vectors of the electromagnetic noise and fluid noise nodes, subtracting these identified and reconstructed noise components from the original mixed noise signal can decouple the interference, and finally obtain the pure bearing vibration signal most directly related to the bearing running state;

[0063] Through the self-supervised graph contrast learning framework, the multi-source highly coupled noise signal is accurately separated without any manual labeled samples, and the learnable edge pruning module is combined to adaptively identify and focus on the key noise coupling path, and the essential distinguishing features of each noise source are learned through the graph attention network, and the entire decoupling model is efficiently deployed in the edge computing unit to realize low-latency and real-time processing at the data source end.

[0064] Specifically, the function of the learnable edge pruning module is to adaptively sparsify the input bearing coherence graph. For each edge in the graph, the module generates a gating weight in the range of (0, 1) through a small multi-layer perception network; during training, end-to-end learning is performed through an online gradient update algorithm based on the Adam optimizer, so that it learns to assign lower weights to edges with low correlation or small contribution to the separation task, dynamically reducing the influence of these edges in the graph attention network, and finally generating a bearing strong coherence graph that only retains the most critical coupling relationships;

[0065] The goal of the self-supervised graph contrast learning model is to efficiently learn node representations for the bearing strong coherence graph without human labels; by data augmentation to construct positive and negative sample pairs, the input strong coherence graph is subjected to two independent and random perturbation operations, such as randomly discarding edges with a probability of 20%, and randomly masking part of the initial feature dimensions of the nodes with a probability of 15%, thereby generating two views with similar content but slight differences; in each iteration of model training, the two enhanced views are simultaneously input into the graph attention network encoder with shared weights to calculate the embedding vectors of each node under the two views; for the same noise source node (such as the mechanical noise node), the two embedding vectors obtained in view A and view B form a positive sample pair, and the embedding vectors of any different nodes form a negative sample pair;

[0066] The self-supervised graph contrastive learning model adopts a normalized temperature scaling cross-entropy loss function for optimization, and the normalized temperature scaling cross-entropy loss function aims to pull the distance between embedding vectors of the same node in different views closer in the feature space through gradient descent, while pushing the distance between embedding vectors of different nodes farther apart; through contrastive learning on a large amount of unlabeled data, the encoder can finally learn to generate embedding vectors that are linearly separable or highly separable in the feature space for the three essentially different nodes of mechanical noise, electromagnetic noise and fluid noise, thereby completing the deep representation of the internal separability of the mixed noise signal and laying a foundation for the subsequent decoder to reconstruct and separate the pure bearing vibration signal;

[0067] Through the synergistic effect of learnable edge pruning and self-supervised contrastive learning, high-quality and unsupervised representation learning of noise source features is achieved; on the one hand, the adaptive edge pruning module can automatically filter out redundant or misleading associated information in the graph, allowing the model to focus on the most critical system dynamics; on the other hand, the contrastive learning based on data augmentation forces the model to discover and learn the most essential and most distinctive invariant features of each noise source, enabling the generation of highly separable node embedding vectors in the feature space without any human labels;

[0068] Referring to Figure 3 The bearing vibration signal is input into a deep learning model for online condition assessment, life prediction and key harmonic recognition, and the deep learning model adopts a hybrid architecture combining convolutional neural networks and long short-term memory networks, aiming to fully utilize the multi-dimensional information in the signal;

[0069] The continuous bearing vibration signal is segmented into input sequences of fixed length (e.g., each segment contains 2048 sampling points), and a one-dimensional convolutional neural network part in the front end of the model is used, which is stacked by convolutional layers and pooling layers. The convolutional layer serves as a learnable feature extractor, automatically capturing local, high-frequency pulse features (time domain features) related to bearing faults and hidden frequency domain patterns (frequency domain features) from the original time domain waveform. The pooling layer reduces the dimensionality of the extracted features and enhances the shift invariance of the features. The output of the one-dimensional convolutional neural network part is a sequence of feature maps, representing the core state information of the original signal in different time segments.

[0070] The feature map sequence extracted by the one-dimensional convolutional neural network is input to the backend of the model, a long short-term memory network stacked by two layers, which learns and memorizes the temporal dependencies between features; By modeling the step-by-step evolution of the feature sequence, the LSTM network can capture the degradation trend of the bearing throughout its life cycle from health to failure, and the output of the LSTM is connected to an independent fully connected layer head for regression prediction, outputting the specific value of the remaining useful life of the bearing;

[0071] The fully connected layer head is used for remaining useful life regression prediction task, the network is a multi-layer perceptron, its structure includes a hidden layer that receives 256-dimensional input and outputs 128 neurons, which uses ReLU as the activation function, and is connected to an output layer that contains only one neuron and uses a linear activation function, ensuring that the final output of the model is a bounded, continuous real number, which can directly correspond to the predicted remaining useful life (e.g., in hours);

[0072] During the training phase of the model, the mean squared error loss function is used to calculate the error between the predicted remaining useful life and the true label for the output of the regression head; By backpropagating this loss, all parameters of the shared backbone network (CNN and LSTM) and the fully connected layer head are updated simultaneously using the Adam optimizer;

[0073] The training and application of the deep learning model includes an online learning mechanism, which uses a historical data set covering multiple operating conditions and complete life cycle data for sufficient offline pre-training before deployment; After deployment to edge computing devices, the model continues to collect new, confirmed vibration signals while making real-time predictions, and periodically monitors significant changes in state, using newly collected data to incrementally train or fine-tune the deployed model; This online learning mechanism ensures that the model can adapt to the unique degradation patterns of the current device, continuously improving the accuracy of the prediction, and when the model's prediction value drops significantly, it will automatically perform high-resolution spectral analysis (such as envelope spectrum analysis) on the vibration signal segment that triggered this judgment, identifying the frequency components with the maximum energy and matching the theoretical frequency of the bearing fault, which are the key harmonic components for subsequent active noise control;

[0074] Through the deep integration of multi-task learning and online fine-tuning mechanisms, we achieve comprehensive intelligent diagnosis from state assessment, life prediction to fault feature positioning, this method outputs the future remaining useful life prediction of the device, and also adaptively optimizes the model to match the degradation patterns of the specific device through the online learning mechanism, continuously improving the diagnostic accuracy, and closely combining the diagnostic results with control requirements: When a fault is predicted, it can automatically and accurately identify the key harmonic components as noise sources, providing direct and executable control basis for subsequent targeted active noise reduction strategies.

[0075] When the deep learning model predicts that the bearing has an early failure and successfully identifies the key harmonic components, a dual noise reduction strategy deployed on the edge computing device will be activated; the dual noise reduction strategy relies on real-time communication between the edge computing device and the motor inverter of the vertical centrifugal pump, the input of the noise reduction process is the decoupled pure bearing vibration signal, and the key harmonic components identified by the deep learning model;

[0076] The dual noise reduction strategy includes a first noise reduction strategy and a second noise reduction strategy, the first noise reduction strategy avoids resonance by dynamically adjusting the motor speed; by slowly increasing and decreasing the speed of the motor, the mechanical or fluid resonance frequency points of the pump system are calibrated and stored in the edge device as a resonance zone; when the real-time identified key harmonic frequency is close to the system resonance zone at the current speed, the edge computing device will start the avoidance algorithm, which calculates the optimal target speed within the allowable speed fluctuation range of the process flow, which can make the key harmonic frequency deviate from the resonance zone, and the edge device sends a new speed instruction to the motor inverter, thereby macroscopically reducing the mechanical and fluid noise coupling amplification effect caused by resonance;

[0077] After the motor is stably running at the new target speed, the second noise reduction strategy actively injects compensation harmonics; the edge computing device calculates a compensation signal for each harmonic that needs to be offset in real time according to the key harmonic components, the compensation signal has the same frequency as the key harmonic, its phase is accurately set to be opposite, and its amplitude is calculated according to a pre-set model of the motor torque response to ensure that a sufficient size of reverse torque pulsation can be generated;

[0078] The edge device digitally synthesizes all the calculated compensation harmonic signals to form a composite compensation waveform, which is sent to the motor inverter of the vertical centrifugal pump through a high-speed communication interface, and is modulated and injected into the stator voltage or current command of the motor by the direct torque control of the motor inverter of the vertical centrifugal pump, thereby generating torque pulsation in the motor air gap that is opposite in phase to the noise source, actively and targetedly canceling specific harmonic noise generated by electromagnetic effects or residual mechanical vibration, and achieving fine suppression of mixed noise; The entire dual strategy continuously runs in a closed-loop control framework, and the speed and compensation signal are continuously fine-tuned by real-time monitoring of the noise reduction effect to achieve the best noise reduction performance.

[0079] The high-efficiency and robust targeted noise reduction is realized by the synergistic control of macro-avoidance and micro-offset; the first noise reduction strategy avoids the resonance region by intelligently adjusting the speed, macroscopically avoids the resonance region, creates a stable operating environment for active noise reduction, and avoids vicious amplification of noise; on this basis, the second strategy realizes the source-level and targeted active offset of key noise harmonics by injecting compensation harmonics into the frequency converter, and the whole strategy runs in a real-time closed loop framework, which can continuously and adaptively optimize the control parameters according to the actual noise reduction effect, and ensure the best noise reduction performance under complex and variable working conditions.

[0080] The embodiment proposes an intelligent noise reduction method with full-process closed loop, realizing seamless connection from signal perception, intelligent separation to active control; by creatively converting multi-source electrical signals into structured coherence graphs, and using self-supervised graph neural networks to realize accurate separation of target vibration signals without human labels, the problem of coupled noise that traditional methods cannot handle is solved; not only can the method based on the separated signals perform deep fault diagnosis and life prediction, but also can execute the double synergistic noise reduction strategy combining speed avoidance and harmonic injection according to the diagnosis results, realize adaptive and targeted suppression of abnormal noise, and achieve excellent results of treating both symptoms and root causes.

[0081] Embodiment two:

[0082] In this embodiment, the application provides a deep learning-based vertical centrifugal pump motor non-driven end bearing noise reduction method, which is applied to a vertical centrifugal pump for a petrochemical device driven by a frequency converter, and the motor non-driven end bearing shows early signs of wear. First, a set of electrical sensors are deployed outside the bearing seat, including a clamp-type current sensor, a parallel voltage sensor, and an electromagnetic induction sensor close to the shell; three signals are collected at a sampling rate of 50 kHz by synchronous sampling, and accurate time alignment is completed by a correction algorithm based on cross-correlation function, generating a mixed noise signal containing mechanical, electromagnetic and fluid noise;

[0083] After obtaining the mixed noise signal, the edge computing device deployed beside the pump machine processes the signal in real time. The device first extracts amplitude spectrum, phase spectrum and envelope features from the signal through fast Fourier transform and Hilbert transform, and calculates the correlation and coherence between signals based on these features, fuses them into weights, and constructs a dynamic bearing coherence graph with mechanical noise, electromagnetic noise and fluid noise as nodes. The self-supervised blind source separation algorithm integrated in the device starts to execute, and the learnable edge pruning module inside the algorithm adaptively sparsifies the coherence graph to generate a strong coherence graph; the self-supervised graph contrast learning model learns the representation of the strong coherence graph by constructing positive and negative sample pairs, and finally successfully separates the pure bearing vibration signal from the complex mixed signal;

[0084] The separated bearing vibration signal is input into a pre-trained CNN-LSTM hybrid deep learning network in real time for online evaluation and prediction. The network model analyzes the time sequence characteristics of the vibration signal to determine that the bearing has entered the early failure stage, and the predicted remaining effective life shows an accelerating downward trend. At the same time, by performing frequency spectrum analysis on the signal segment triggering the fault judgment, the key harmonic component matching the bearing outer ring fault characteristic frequency is identified.

[0085] Once the key harmonic component is identified, the edge computing device immediately starts a dual noise reduction strategy. First, the first strategy is executed, and the device queries the pre-stored system resonance zone map to find that the key harmonic frequency at the current speed is close to a resonance point. Then, the device calculates and sends instructions to the motor frequency converter to reduce the motor speed by 4%, successfully avoiding the risk of resonance amplification. After the new speed stabilizes, the second strategy is immediately executed. The edge computing device calculates a compensation harmonic signal with completely opposite parameters according to the amplitude and phase of the identified key harmonic, and injects the digital signal into the frequency converter through the high-speed bus. The frequency converter generates an anti-phase torque ripple accordingly to actively cancel the specific harmonic noise generated by the bearing fault. The entire process runs continuously in a closed loop, and the noise reduction effect is continuously monitored and fine-tuned, significantly reducing the abnormal noise of the bearing.

[0086] Although embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based vertical centrifugal pump motor non-driving end bearing noise reduction method, characterized in that, The application relates to a method for predicting the remaining useful life of a bearing of a vertical centrifugal pump motor. The method comprises the following steps: An electrical sensor is arranged on a bearing seat at a non-driving end of a motor of a vertical centrifugal pump, the electrical sensor comprises a voltage sensor, a current sensor and an electromagnetic induction sensor, motor stator current signals, stator voltage signals and electromagnetic induction signals of the bearing at the non-driving end of the motor of the vertical centrifugal pump are collected respectively, and synchronous sampling and time domain alignment are performed to generate bearing mixed noise signals containing mechanical noise, electromagnetic noise and fluid noise; The synchronous sampling and time domain alignment eliminate the phase difference and sampling delay of signals collected by different sensors through a time offset correction method based on a cross-correlation function.

2. The deep learning-based vertical centrifugal pump motor non-driving end bearing noise reduction method of claim 1, wherein, The bearing coherence graph is constructed based on a characteristic relationship of the bearing mixed noise signals, and the specific process is as follows: Amplitude spectrum, phase spectrum and envelope features are extracted based on the bearing mixed noise signals; the amplitude spectrum and the phase spectrum are extracted through fast Fourier transform, and the envelope features are extracted through Hilbert transform based on the amplitude spectrum and the phase spectrum; The cross-correlation coefficient and the coherence function between the electrical sensors are calculated based on the amplitude spectrum, the phase spectrum and the envelope features, amplitude similarity, phase consistency and envelope correlation are generated and fused into edge weights; the bearing coherence graph is constructed by taking the mechanical noise, the electromagnetic noise and the fluid noise as nodes and taking the fused edge weights as weighted edges.

3. The deep learning based vertical centrifugal pump motor non-driving end bearing noise reduction method of claim 1, wherein, The self-supervised blind source decoupling algorithm is executed on the bearing coherence graph by an edge computing device, the self-supervised blind source decoupling algorithm comprises a learnable edge pruning module and a self-supervised graph contrast learning model, the bearing coherence graph is adaptively sparsified online, a bearing strong coherence graph is generated, the bearing strong coherence graph is separated from noise through the self-supervised graph contrast learning model and by using constructed positive sample pairs and negative sample pairs, and a bearing vibration signal is generated. The bearing vibration signal is input into a deep learning model for online learning to predict the remaining useful life, identify key harmonic components and execute a double noise reduction strategy based on the key harmonic components, the double noise reduction strategy comprises adjusting the motor speed to avoid a high noise resonance region and injecting compensation harmonics into a motor frequency converter, and noise reduction is performed through reverse torque pulsation. The bearing mixed noise collection process is as follows:

4. The deep learning based vertical centrifugal pump motor non-driving end bearing noise reduction method of claim 1, wherein, An electrical sensor is arranged on a bearing seat at a non-driving end of a motor of a vertical centrifugal pump, the electrical sensor comprises a voltage sensor, a current sensor and an electromagnetic induction sensor, motor stator current signals, stator voltage signals and electromagnetic induction signals of the bearing at the non-driving end of the motor of the vertical centrifugal pump are collected respectively, and synchronous sampling and time domain alignment are performed to generate bearing mixed noise signals containing mechanical noise, electromagnetic noise and fluid noise; The synchronous sampling and time domain alignment eliminate the phase difference and sampling delay of signals collected by different sensors through a time offset correction method based on a cross-correlation function. The bearing coherence graph is constructed based on a characteristic relationship of the bearing mixed noise signals, and the specific process is as follows:

5. The deep learning based vertical centrifugal pump motor non-driving end bearing noise reduction method of claim 4, wherein, Amplitude spectrum, phase spectrum and envelope features are extracted based on the bearing mixed noise signals; the amplitude spectrum and the phase spectrum are extracted through fast Fourier transform, and the envelope features are extracted through Hilbert transform based on the amplitude spectrum and the phase spectrum; The cross-correlation coefficient and the coherence function between the electrical sensors are calculated based on the amplitude spectrum, the phase spectrum and the envelope features, amplitude similarity, phase consistency and envelope correlation are generated and fused into edge weights; the bearing coherence graph is constructed by taking the mechanical noise, the electromagnetic noise and the fluid noise as nodes and taking the fused edge weights as weighted edges. The self-supervised blind source decoupling algorithm is executed on the bearing coherence graph by an edge computing device, the self-supervised blind source decoupling algorithm comprises a learnable edge pruning module and a self-supervised graph contrast learning model, the bearing coherence graph is adaptively sparsified online, a bearing strong coherence graph is generated, the bearing strong coherence graph is separated from noise through the self-supervised graph contrast learning model and by using constructed positive sample pairs and negative sample pairs, and a bearing vibration signal is generated. The bearing vibration signal is input into a deep learning model for online learning to predict the remaining useful life, identify key harmonic components and execute a double noise reduction strategy based on the key harmonic components, the double noise reduction strategy comprises adjusting the motor speed to avoid a high noise resonance region and injecting compensation harmonics into a motor frequency converter, and noise reduction is performed through reverse torque pulsation. The learnable edge pruning module is used to assign adjustable weights to each edge of the bearing coherence graph, and dynamically reduce the weights of low correlation edges through an online gradient update algorithm, to adaptively sparsify the bearing coherence graph, and generate a bearing strong coherence graph; The self-supervised graph contrast learning model learns node representation of the bearing strong coherence graph by constructing positive sample pairs and negative sample pairs, so that the embedding vectors of nodes in positive sample pairs are close in the feature space, and the embedding vectors of nodes in negative sample pairs are far away, to separate the bearing mixed noise signal, and generate a bearing vibration signal.

6. The deep learning based vertical centrifugal pump motor non-driving end bearing noise reduction method of claim 1, wherein, The process of inputting the bearing vibration signal into the deep learning model for online learning to predict the remaining useful life and identify the key harmonic components is as follows: The deep learning model includes a convolutional neural network and a long short-term memory network, which are used to extract time domain features, frequency domain features and time-dependent features of the bearing vibration signal at the same time; Through an online learning mechanism, the deep learning model performs incremental training on historical vibration data and real-time collected bearing vibration signals, and outputs a bearing remaining useful life prediction value; Based on the bearing remaining useful life prediction value and vibration signal spectrum analysis, the key harmonic components related to the bearing degradation state are identified.

7. The deep learning based vertical centrifugal pump motor non-driving end bearing noise reduction method of claim 1, wherein, The double noise reduction strategy is: The double noise reduction strategy includes a first noise reduction strategy and a second noise reduction strategy; The first noise reduction strategy is to adjust the motor speed to dynamically avoid the high noise resonance region by real-time monitoring of the bearing vibration signal and the key harmonic components, and to reduce the coupling interference of mechanical noise and fluid noise; The second noise reduction strategy is to inject a compensation harmonic signal into the motor frequency converter, to generate an anti-phase torque ripple opposite in phase to the key harmonic components, to actively cancel the electromagnetic noise and residual mechanical vibration, and to reduce the bearing mixed noise.

8. The deep learning-based vertical centrifugal pump motor non-driving end bearing noise reduction method of claim 7, wherein, The process of injecting a compensation harmonic signal to generate an anti-phase torque ripple opposite in phase to the key harmonic components is as follows: The edge computing device calculates the amplitude and phase of the corresponding anti-phase harmonic according to the amplitude and phase information of the key harmonic components identified by the deep learning model; The anti-phase harmonic amplitude and phase are injected into the motor stator through the motor frequency converter, so that the generated anti-phase torque ripple and the original key harmonic components are mutually canceled in phase, and the amplitude and phase of the injected signal are continuously adjusted in real time to reduce noise.

Images