Deep learning-based vertical centrifugal pump motor non-driving end bearing noise reduction method

Abstract

The present application 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, comprising collecting multi-source mixed noise through an electrical sensor deployed on a bearing seat, and constructing a bearing coherence graph representing signal relationships; on an edge computing device, a blind source decoupling algorithm integrating learnable edge pruning and self-supervised graph contrast learning is used to adaptively sparsify and learn the representation of the coherence graph, and separate out pure bearing vibration signals; the bearing vibration signals are input into a deep learning model for online analysis, to predict the remaining effective life of the bearing and identify key harmonic components; a double noise reduction strategy is executed, the motor speed is adjusted to avoid the resonance region, and compensation harmonics are actively injected into the frequency converter to generate counter-phase torque pulsation, achieving targeted suppression of noise. The present application reduces the noise of the vertical centrifugal pump motor non-driving end bearing through a 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] An edge computing device is constructed to execute 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 of the bearing coherence graph to generate a bearing strong coherence graph. Through a self-supervised graph comparison learning model, noise separation is performed on the bearing strong coherence graph using constructed positive and negative sample pairs 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 effective life, identify key harmonic components, and implement a dual noise reduction strategy based on the key harmonic components, including adjusting the motor speed to avoid the high noise resonance area and injecting compensation harmonics into the motor inverter, and reducing noise through anti-phase torque pulsation.

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

[0012] Electrical sensors are deployed in the bearing housing of the non-drive end of the vertical centrifugal pump motor. The electrical sensors include voltage sensors, current sensors and electromagnetic induction sensors. They respectively collect the stator current signal, stator voltage signal and electromagnetic induction signal of the non-drive end bearing of the vertical centrifugal pump motor, and perform synchronous sampling and time domain alignment to generate a mixed bearing noise signal including mechanical noise, electromagnetic noise and fluid noise.

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

[0014] The specific process of constructing the bearing coherence map based on the characterization relationship of the bearing noise signal is as follows:

[0015] Amplitude spectrum, phase spectrum, and envelope features are extracted from the bearing mixed noise signal; the amplitude spectrum and phase spectrum are extracted using Fast Fourier Transform, and the envelope features are extracted using Hilbert Transform based on the amplitude spectrum and phase spectrum;

[0016] Based on the amplitude spectrum, phase spectrum, and envelope characteristics, the cross-correlation coefficient and coherence function between electrical sensors are calculated, and amplitude similarity, phase consistency, and envelope correlation are generated and fused into edge weights. Using mechanical noise, electromagnetic noise, and fluid noise as nodes and the fused edge weights as weighted edges, a bearing coherence graph is constructed.

[0017] The construction of the edge computing device performs self-supervised blind source decoupling on the bearing coherence graph in the following specific manner:

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

[0019] The separation process identifies the differential edge weight distribution of mechanical vibration characteristics, electromagnetic induction characteristics, and fluid disturbance characteristics in the graph structure, and peels off the corresponding noise components one by one to finally obtain the bearing vibration signal related to the bearing operating state.

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

[0021] 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-relevance edges 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 contrastive learning model constructs positive and negative sample pairs to learn node representations of the bearing strong coherence graph. This makes the embedding vectors of nodes in the positive sample pairs closer together in the feature space, while the embedding vectors of nodes in the negative sample pairs are further apart, thus separating the mixed noise signal of the bearing and generating the bearing vibration signal.

[0023] The specific process of inputting the bearing vibration signal into a deep learning model for online learning to predict the remaining effective life and identify key harmonic components is as follows:

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

[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 predicted value of the remaining effective life of the bearing.

[0026] Based on the prediction results and vibration signal spectrum analysis, key harmonic components related to bearing deterioration were identified.

[0027] The dual noise reduction strategy is as follows:

[0028] The dual 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 and dynamically avoid the high-noise resonance zone by real-time monitoring of bearing vibration signals and key harmonic components, thereby reducing the coupling interference between mechanical noise and fluid noise.

[0030] The second noise reduction strategy is to inject a compensation harmonic signal into the motor inverter. By generating an anti-phase torque pulsation that is opposite in phase to the key harmonic components, the electromagnetic noise and residual mechanical vibration are actively canceled, thereby reducing the mixed noise of the bearing.

[0031] The specific process of generating an anti-phase torque pulsation with the phase opposite to the key harmonic component by injecting the compensation harmonic signal is as follows:

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

[0033] The amplitude and phase of the anti-phase harmonics are injected into the motor stator through the motor frequency converter, so that the generated anti-phase torque pulsation cancels out the original key harmonic components in phase. Noise reduction is achieved by continuously adjusting the amplitude and phase of the injected signal in real time.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention transforms a complex timing problem into a structured graph problem by deploying multiple electrical sensors and constructing a signal coherence graph. Utilizing a self-supervised blind source decoupling algorithm deployed on an edge computing device, through learnable edge pruning and graph comparison learning, it can automatically and without human intervention accurately extract vibration signals strongly correlated with bearing conditions from mixed noise superimposed from multiple sources such as mechanical, electromagnetic, and fluid sources. This method overcomes the limitations of traditional signal processing methods that rely on manual feature design and are difficult to handle non-stationary signals, significantly improving the separation accuracy of the target signal and the adaptive capability of the algorithm.

[0036] 2. This invention not only focuses on noise reduction, but also uses the separated pure bearing vibration signal for in-depth analysis. Through the constructed convolutional neural network and long short-term memory network deep learning model, it can simultaneously extract the time, frequency domain and time-dependent features of the signal. Combined with the online incremental training mechanism, it can not only accurately identify the key harmonic components directly related to the bearing deterioration state, but also predict the remaining effective life of the bearing in real time. This upgrades the method into a set of equipment health management methods with both diagnostic and predictive functions.

[0037] 3. Based on the key harmonics identified by the deep learning model, this invention implements an innovative dual noise reduction strategy. On the one hand, it avoids the resonance zone by adjusting the motor speed, thereby reducing coupling noise macroscopically. On the other hand, it generates anti-phase harmonics in real time through edge computing devices and injects them into the motor via the frequency converter to generate anti-phase torque pulsation, thereby achieving precise active cancellation of key electromagnetic and mechanical noise. This closed-loop control method, which combines avoidance and cancellation, has a clearer noise reduction target, more significant effect, and stronger adaptability compared to single passive vibration isolation or active noise reduction methods. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a deep learning-based method for noise reduction in the non-drive end bearings of a vertical centrifugal pump motor.

[0039] Figure 2 This is a schematic diagram of the self-supervised blind source decoupling algorithm structure for a deep learning-based method of noise reduction in the non-drive end bearing of a vertical centrifugal pump motor.

[0040] Figure 3 A flowchart illustrating the process of noise reduction for bearing vibration signals in a deep learning-based method for noise reduction of bearings at the non-drive end of a vertical centrifugal pump motor. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1:

[0043] Please see Figure 1 This invention provides a noise reduction method for the non-drive end bearing of a vertical centrifugal pump motor based on deep learning. The technical solution is as follows:

[0044] A deep learning-based method for noise reduction of the non-drive end bearings of a vertical centrifugal pump motor includes:

[0045] 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.

[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 includes an integrated learnable edge pruning module to perform online adaptive sparsification of the data coherence map to generate a bearing strong coherence map. Through a self-supervised graph comparison learning model, the constructed positive and negative sample pairs are used to separate noise from the bearing strong coherence map 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 effective life, identify key harmonic components, and implement a dual noise reduction strategy based on the key harmonic components, including adjusting the motor speed to avoid the high noise resonance area and injecting compensation harmonics into the motor inverter, 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 housing on the non-drive end of its motor. This sensor set includes a Hall effect clamp-on 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, closely attached to the bearing housing housing, for picking up local leakage magnetic field fluctuations caused by bearing operation.

[0049] During data acquisition, the sampling mode was set to synchronous sampling to ensure that the signals from the voltage, current, and electromagnetic induction channels were acquired at the hardware level simultaneously, thereby minimizing the scanning delay between channels. To capture the high-frequency characteristics that may be generated by bearing failure, the sampling frequency was set to 50kHz and the sampling precision to 24 bits to ensure sufficient dynamic range and signal resolution.

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

[0051] Based on the sample point offset corresponding to the peak position of the cross-correlation function, the entire time series data of voltage and electromagnetic induction signals are shifted and compensated. After completing this alignment step, a multi-dimensional data matrix is ​​finally obtained, where each column represents the stator current, stator voltage and electromagnetic induction signals that have been accurately aligned on the time base, which constitute the bearing mixed noise signal, including mechanical noise, electromagnetic noise and fluid noise.

[0052] By jointly deploying current, voltage, and electromagnetic induction sensors, coupled information related to the bearing's operating state was successfully captured from multiple physical dimensions, providing a far more comprehensive and richer picture than data from a single sensor. Simultaneously, a dual strategy combining hardware synchronous sampling and software fine alignment based on cross-correlation functions ensured extremely high time synchronization accuracy and data fidelity among multiple signals. This provided a high-quality, phase-distortion-free data foundation for subsequent precise separation of mixed mechanical, electromagnetic, and fluid noise components, as well as for achieving highly reliable fault diagnosis.

[0053] It is constructed based on the inherent characterization relationship of the bearing mixed noise signal;

[0054] Each time series of the acquired stator current, stator voltage, and electromagnetic induction signals is segmented. For example, the signal is divided into multiple data frames containing 4096 sampling points with a 50% overlap. A Hanning window is applied to each data frame to suppress spectral leakage. Feature extraction is performed on each windowed data frame. The signal of each frame is transformed from the time domain to the frequency domain by applying a Fast Fourier Transform to obtain its corresponding complex spectrum. Based on the complex spectrum, the amplitude spectrum (i.e., the modulus of the complex number) and the phase spectrum (i.e., the argument of the complex number) are extracted. Simultaneously, a Hilbert transform is applied to each original time-domain data frame to construct its analytic signal, and the modulus of the analytic signal is taken to obtain the instantaneous envelope features of the signal.

[0055] The amplitude similarity of any two sensor signals is quantified by calculating the cross-correlation coefficient between their corresponding amplitude spectra; the phase consistency is quantified by calculating the mean of the squared amplitude coherence function of any two sensor signals within the main frequency band; and the envelope correlation is quantified by calculating the Pearson correlation coefficient between the envelope signals of any two sensor signals. These calculations are performed on all data frames and averaged. By quantifying the signal relationship from three dimensions—amplitude, phase, and envelope—a comprehensive and in-depth analysis of dynamic correlation is achieved. At the same time, frame-by-frame averaging and other processing techniques ensure the statistical robustness of these quantification indicators, providing a highly reliable data foundation for constructing an accurate coherence graph model.

[0056] The graph is defined with three nodes representing mechanical noise, electromagnetic noise, and fluid noise, respectively. The amplitude similarity, phase consistency, and envelope correlation calculated in the previous step are fused by weighting. In this embodiment, the weights are set to 0.3, 0.4, and 0.3 to form a single fused edge weight. This weight is assigned between nodes that represent the corresponding physical processes. For example, the weights calculated from the current and voltage signals define the edges between the electromagnetic noise and fluid noise nodes. This constructs a bearing coherence graph containing three nodes and three weighted edges, which visually represents the interaction strength between different noise sources in a structured way.

[0057] The bearing coherence graph constructed in this embodiment abstracts and transforms the high-dimensional, unstructured time-series signal relationships into a low-dimensional, structured graphical model. Nodes represent physical noise sources, and weighted edges incorporating multi-dimensional features quantify the interaction strength between sources. This provides an intuitive, system-level perspective for understanding complex noise coupling mechanisms. This structured data representation makes complex dynamic relationships machine-readable, providing ideal input and a solid foundation for subsequent applications of graph algorithms for accurate noise source decoupling.

[0058] Executing a self-supervised blind source decoupling algorithm on edge computing devices based on bearing coherence graphs, referencing Figure 2The self-supervised blind source decoupling algorithm is deployed on an edge computing unit (e.g., NVIDIA Jetson AGXXavier) close to a vertical centrifugal pump. The edge computing unit has GPU acceleration capabilities, which can meet the real-time computing requirements of the graph neural network model. The 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 learnable edge pruning module dynamically optimizes the input bearing coherence graph. This module is essentially a small neural network. Based on the initial features of the nodes in the graph, it calculates an importance weight (between 0 and 1) for each edge. For any two nodes, the initial feature vectors of these two nodes are concatenated to form a combined feature vector with doubled dimensions, completely containing all the initial information of the two nodes connected by that edge. This combined feature vector is then input into a multilayer perceptron, which contains a hidden layer with a ReLU activation function and an output layer. The output layer has only one neuron, and its input is mapped to a weight between 0 and 1 using a Sigmoid activation function.

[0060] The weights are multiplied by the original fused edge weights to dynamically adjust the connection strength of the edges in the graph. During training, the model will learn to prune, that is, weaken or even ignore edges that are irrelevant to the target task, so that the graph attention network can focus more on the graph structure that represents the coupling relationship of key noise sources and generate a sparse bearing coherent graph.

[0061] The graph attention network encoder is trained using self-supervised graph contrastive learning to learn effective feature representations for each noise source node. Two related but distinct views are constructed by applying two different random data augmentations to the input strongly coherent graph (e.g., randomly discarding edges and masking node features). For the same noise source node, its representations in the two views constitute positive sample pairs, while the representations of different nodes constitute negative sample pairs. The goal of this self-supervised graph contrastive learning is to optimize the contrastive loss function so that the vector representations of positive sample pairs in the feature space are as similar as possible, while the vector representations of negative sample pairs are as far apart as possible.

[0062] After self-supervised pre-training, the graph attention network encoder generates highly discriminative feature vectors for three nodes: mechanical noise, electromagnetic noise, and fluid noise, performing noise separation and target signal reconstruction. The trained model is used to identify the edge weight distribution and feature patterns in the graph dominated by electromagnetic effects and fluid disturbances. A decoder (e.g., a multilayer perceptron) is used to reconstruct the corresponding noise components from the original mixed noise signal based on the feature vectors of the electromagnetic noise and fluid noise nodes. By subtracting these identified and reconstructed noise components from the original mixed noise signal, the interference can be decoupled, and finally, a pure bearing vibration signal that is most directly related to the bearing's own operating state can be obtained.

[0063] The self-supervised graph contrastive learning framework enables accurate separation of highly coupled multi-source noise signals without the need for manual sample labeling. Combined with a learnable edge pruning module, it can adaptively identify and focus on key noise coupling paths. The graph attention network learns the essential distinguishing features of each noise source. The entire decoupling model is efficiently deployed on edge computing units, achieving low-latency, real-time processing at the data source.

[0064] Specifically, the learnable edge pruning module adaptively sparsifies the input bearing coherence graph. For each edge in the graph, the module generates a gated weight in the range of (0, 1) through a small multilayer perceptron network. During training, end-to-end learning is performed through an online gradient update algorithm based on the Adam optimizer, enabling it to assign lower weights to edges with low relevance 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 retains only the most critical coupling relationships.

[0065] The goal of the self-supervised graph contrastive learning model is to efficiently learn node representations of a bearing strongly coherent graph without human labels. It constructs positive and negative sample pairs through data augmentation by performing two independent, random perturbation operations on the input strongly coherent graph. For example, it randomly discards edges in the graph with a 20% probability and randomly masks some initial feature dimensions of nodes with a 15% probability, thereby generating two views with similar but slightly different content. In each iteration of model training, these two augmented views are simultaneously input into a graph attention network encoder with shared weights to calculate the embedding vector of each node in both views. For the same noise source node (such as a mechanical noise node), the two embedding vectors obtained in view A and view B constitute a positive sample pair, while the embedding vectors of any different nodes constitute a negative sample pair.

[0066] The self-supervised graph contrastive learning model is optimized using a normalized temperature-scaling cross-entropy loss function. The goal of this loss function is to bring all positive sample pairs (i.e., the distance between the embedding vectors of the same node in different views) closer together in the feature space through gradient descent, while simultaneously pushing away all negative sample pairs (i.e., the distance between the embedding vectors of different nodes). Through contrastive learning on a large amount of unlabeled data, the encoder can ultimately learn to generate linearly separable or highly separable embedding vectors for three fundamentally different nodes: mechanical noise, electromagnetic noise, and fluid noise. This achieves a deep representation of the inherent separability of mixed noise signals, laying the foundation for the subsequent decoder to reconstruct and separate the pure bearing vibration signal.

[0067] By combining learnable edge pruning with self-supervised contrastive learning, high-quality, 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 correlation information in the graph, allowing the model to focus on the most critical system dynamics. On the other hand, data-augmented contrastive learning forces the model to discover and learn the most essential and discriminative invariant features of each noise source, enabling the generation of highly separable node embedding vectors in the feature space without any manual labels.

[0068] See Figure 3 The bearing vibration signal is input into a deep learning model to perform online condition assessment, life prediction and key harmonic identification. The deep learning model adopts a hybrid architecture that combines convolutional neural network and long short-term memory network, which aims to make full use of the multi-dimensional information in the signal.

[0069] The continuous bearing vibration signal is segmented into a fixed-length input sequence (e.g., each segment contains 2048 sampling points). This sequence is then processed by a one-dimensional convolutional neural network (CNN) at the front end of the model. The CNN consists of stacked convolutional and pooling layers. The convolutional layers act as learnable feature extractors, automatically capturing local, high-frequency pulse features (time-domain features) and hidden frequency-domain patterns (frequency-domain features) related to bearing faults from the original time-domain waveform. The pooling layers reduce the dimensionality of the extracted features, enhancing their translation invariance. The output of the CNN 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 into the back end of the model, a long short-term memory network consisting of two stacked layers, which learns and remembers the temporal dependencies between features. By modeling the gradual evolution of the feature sequence, the LSTM network can capture the degradation trend of the bearing throughout its entire life cycle from health to failure. The output of the LSTM is connected to an independent fully connected layer head for regression prediction, outputting the specific value of the bearing's remaining effective life.

[0071] The fully connected layer head is used for the remaining effective life regression prediction task. The network is a multilayer perceptron, whose structure includes a hidden layer that receives 256-dimensional input and outputs 128 neurons. This hidden layer uses ReLU as the activation function and is connected to an output layer containing only one neuron and using a linear activation function. This ensures that the final output of the model is an unbounded, continuous real value that can directly correspond to the predicted remaining effective 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 effective lifetime and the true label for the output of the regression head; by backpropagating this loss, the Adam optimizer is used to update all parameters of the shared backbone network (CNN and LSTM) and the fully connected layer head at the same time.

[0073] The training and application of the deep learning model includes an online learning mechanism. Before deployment, it undergoes thorough offline pre-training using historical datasets covering various working conditions and the entire lifecycle. After deployment to the edge computing device, the model continuously collects new, confirmed vibration signals while making real-time predictions. When significant changes in the state are detected periodically, the newly collected data is used 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 and continuously improve the accuracy of predictions. When the model's predicted value drops significantly, it automatically performs high-resolution spectral analysis (such as envelope spectrum analysis) on the vibration signal segment that triggers the judgment, identifying the frequency components with the highest energy that match the theoretical frequency of bearing failure. These are the key harmonic components used for subsequent active noise reduction control.

[0074] By deeply integrating multi-task learning and online fine-tuning mechanisms, this method achieves comprehensive intelligent diagnosis from state assessment and lifetime prediction to fault feature localization. It outputs a prediction of the device's future remaining effective lifetime and can also adaptively optimize the model through online learning mechanisms to match the degradation mode of specific devices, continuously improving diagnostic accuracy and closely integrating diagnostic results with control requirements. When a fault is predicted, it can automatically and accurately identify key harmonic components that are noise sources, thus providing a direct and executable control basis for subsequent implementation of targeted active noise reduction strategies.

[0075] Once the deep learning model predicts an early bearing failure and successfully identifies 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 to 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 performing slow acceleration and deceleration scanning on the motor, the mechanical or fluid resonance frequency points of the pump system are calibrated and stored as resonance zones in the edge device. When the key harmonic frequency identified in real time is close to the system resonance zone at the current speed, the edge computing device will start the avoidance algorithm. The avoidance algorithm calculates the optimal target speed within the speed fluctuation range allowed by the process flow. This speed can make the key harmonic frequency deviate from the resonance zone. The edge device sends a new speed command to the motor inverter, thereby macroscopically reducing the mechanical and fluid noise coupling amplification effect caused by resonance.

[0077] After the motor stabilizes at the new target speed, the second noise reduction strategy is activated to actively inject compensation harmonics; the edge computing device calculates a compensation signal for each harmonic that needs to be canceled in real time based on the key harmonic components. The frequency of the compensation signal is the same as that of the key harmonic, its phase is precisely set to be out of phase, and its amplitude is calculated according to the preset model of the motor torque response to ensure that a sufficiently large reverse torque pulsation can be generated.

[0078] The edge device digitally synthesizes all calculated compensation harmonic signals into a composite compensation waveform, which is then sent to the motor inverter of the vertical centrifugal pump via a high-speed communication interface. The direct torque control of the motor inverter modulates the waveform and injects it into the stator voltage or current command of the motor, generating torque pulsations in the air gap of the motor that are out of phase with the noise source. This actively and specifically cancels out specific harmonic noise generated by electromagnetic effects or residual mechanical vibrations, achieving fine-grained suppression of mixed noise. The entire dual strategy operates continuously within a closed-loop control framework, constantly fine-tuning the speed and compensation signal by monitoring the noise reduction effect in real time to achieve optimal noise reduction performance.

[0079] Efficient and robust targeted noise reduction is achieved through coordinated control combining macroscopic avoidance and microscopic cancellation. The first noise reduction strategy avoids the resonance zone macroscopically by intelligently adjusting the rotational speed, creating a stable operating environment for active noise reduction and preventing the malignant amplification of noise. On this basis, the second strategy achieves source-level, targeted active cancellation of key noise harmonics by injecting compensation harmonics into the frequency converter. The entire strategy operates in a real-time closed-loop framework and can continuously and adaptively optimize control parameters according to the actual noise reduction effect, ensuring the best noise reduction performance under complex and variable operating conditions.

[0080] This embodiment proposes a closed-loop intelligent noise reduction method that achieves seamless integration from signal perception and intelligent separation to active control. By creatively transforming multi-source electrical signals into structured coherent graphs and utilizing a self-supervised graph neural network, it achieves accurate separation of target vibration signals without the need for manual labeling, solving the problem of traditional methods struggling to handle coupled noise. This method can not only perform in-depth fault diagnosis and life prediction based on the separated signals, but also execute a dual collaborative noise reduction strategy combining speed avoidance and harmonic injection based on the diagnosis results. This achieves adaptive and targeted suppression of abnormal noise, resulting in excellent effects that address both the symptoms and the root cause.

[0081] Example 2:

[0082] In this embodiment, the present invention provides a deep learning-based noise reduction method for the non-drive end bearing of a vertical centrifugal pump motor. This method is applied to a vertical centrifugal pump used in a petrochemical plant, driven by a frequency converter, where the non-drive end bearing of the motor shows signs of early wear. First, a set of electrical sensors is deployed outside the bearing housing, including a clamp-on current sensor, a parallel voltage sensor, and an electromagnetic induction sensor closely attached to the housing. Three signals are acquired synchronously at a sampling rate of 50kHz, and precise time-domain alignment is achieved through a correction algorithm based on the cross-correlation function, generating a mixed noise signal containing mechanical, electromagnetic, and fluid noise.

[0083] After acquiring the mixed noise signal, the edge computing device deployed next to the pump processes the signal in real time. The device first extracts the amplitude spectrum, phase spectrum, and envelope features from the signal through fast Fourier transform and Hilbert transform, and calculates the correlation and coherence between the signals based on these features, and fuses them into weights to construct a dynamic bearing coherence map with mechanical noise, electromagnetic noise, and fluid noise as nodes. The self-supervised blind source decoupling algorithm integrated in the device starts to execute, and its internal learnable edge pruning module adaptively sparsifies the coherence map to generate a strong coherence map. The self-supervised graph contrastive learning model constructs positive and negative sample pairs to perform representation learning on the strong coherence map, 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 determines that the bearing has entered the early failure stage by analyzing the temporal characteristics of the vibration signal, and its predicted remaining effective life shows an accelerated downward trend. At the same time, by performing spectral analysis on the signal segment that triggers the fault judgment, key harmonic components that match the fault characteristic frequency of the bearing outer ring are identified.

[0085] Once the critical harmonic component is identified, the edge computing device immediately initiates a dual noise reduction strategy. First, the first strategy is executed: the device queries a pre-stored system resonance zone map and finds that the critical harmonic frequency at the current speed is close to a resonance point. It then calculates and sends a command to the motor inverter, fine-tuning the motor speed by 4%, successfully avoiding the risk of resonance amplification. After the new speed stabilizes, the second strategy is immediately executed. Based on the amplitude and phase of the identified critical harmonic, the edge computing device calculates a compensation harmonic signal with completely opposite parameters in real time and injects this digital signal into the inverter via a high-speed bus. The inverter then generates an anti-phase torque pulsation to actively cancel out the specific harmonic noise generated by the bearing fault. The entire process runs continuously in a closed loop, with fine-tuning made by constantly monitoring the noise reduction effect, significantly reducing abnormal bearing noise.

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

Claims

1. A method for noise reduction of the non-drive end bearing of a vertical centrifugal pump motor based on deep learning, characterized in that, include: An electrical sensor is deployed in the bearing housing at the non-drive end of the vertical centrifugal pump motor to collect the mixed bearing noise and perform time-domain alignment to generate a mixed bearing noise signal. Based on the characterization relationship of the mixed bearing noise signal, a bearing coherence diagram is constructed. The specific process is as follows: Amplitude spectrum, phase spectrum, and envelope features are extracted from the bearing mixed noise signal. The amplitude spectrum and phase spectrum are extracted using Fast Fourier Transform, and the envelope features are extracted using Hilbert Transform based on the amplitude spectrum and phase spectrum. Based on the amplitude spectrum, phase spectrum, and envelope characteristics, the cross-correlation coefficient and coherence function between electrical sensors are calculated, and amplitude similarity, phase consistency, and envelope correlation are generated and fused into edge weights. Using mechanical noise, electromagnetic noise, and fluid noise as nodes and the fused edge weights as weighted edges, a bearing coherence graph is constructed. An edge computing device is constructed to execute 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 of the bearing coherence graph to generate a bearing strong coherence graph. Through a self-supervised graph comparison learning model, noise separation is performed on the bearing strong coherence graph using constructed positive and negative sample pairs to generate a bearing vibration signal. The bearing vibration signal is input into a deep learning model for online learning to predict the remaining effective life, identify key harmonic components, and execute a dual noise reduction strategy based on the key harmonic components. The dual 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 and dynamically avoid high noise resonance areas by real-time monitoring of bearing vibration signals and key harmonic components, thereby reducing the coupling interference of mechanical noise and fluid noise. The second noise reduction strategy is to inject a compensation harmonic signal into the motor inverter. By generating an anti-phase torque pulsation that is opposite in phase to the key harmonic components, the electromagnetic noise and residual mechanical vibration are actively canceled, thereby reducing the mixed noise of the bearing.

2. The noise reduction method for the non-drive end bearing of a vertical centrifugal pump motor based on deep learning according to claim 1, characterized in that, The process for collecting mixed bearing noise is as follows: Electrical sensors are deployed in the bearing housing of the non-drive end of the vertical centrifugal pump motor. The electrical sensors include voltage sensors, current sensors and electromagnetic induction sensors. They respectively collect the stator current signal, stator voltage signal and electromagnetic induction signal of the non-drive end bearing of the vertical centrifugal pump motor, and perform synchronous sampling and time domain alignment to generate a mixed bearing noise signal including mechanical noise, electromagnetic noise and fluid noise. The synchronous sampling and time-domain alignment eliminates the phase difference and sampling delay of signals acquired by different sensors through a time offset correction method based on cross-correlation function.

3. The noise reduction method for the non-drive end bearing of a vertical centrifugal pump motor based on deep learning according to claim 1, characterized in that, The construction of the edge computing device performs self-supervised blind source decoupling on the bearing coherence graph in the following specific manner: A self-supervised blind source decoupling algorithm is deployed in an edge computing device near the vertical centrifugal pump motor. The self-supervised blind source decoupling algorithm integrates a learnable edge pruning module and a self-supervised graph contrastive learning model to separate mechanical noise, electromagnetic noise and fluid noise. The separation process identifies the differential edge weight distribution of mechanical vibration characteristics, electromagnetic induction characteristics, and fluid disturbance characteristics in the graph structure, and then peels off the corresponding noise components one by one to obtain the bearing vibration signal related to the bearing operating state.

4. The method for noise reduction of the non-drive end bearing of a vertical centrifugal pump motor based on deep learning according to claim 3, characterized in that, The self-supervised blind source decoupling specifically refers to: 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-relevance 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 contrastive learning model constructs positive and negative sample pairs to learn node representations of the bearing strong coherence graph. This makes the embedding vectors of nodes in the positive sample pairs closer together in the feature space, while the embedding vectors of nodes in the negative sample pairs are further apart, thus separating the mixed noise signal of the bearing and generating the bearing vibration signal.

5. The method for noise reduction of the non-drive end bearing of a vertical centrifugal pump motor based on deep learning according to claim 1, characterized in that, The specific process of inputting the bearing vibration signal into a deep learning model for online learning to predict the remaining effective life and identify 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 simultaneously extract the time-domain features, frequency-domain features and time-dependent features of the bearing vibration signal; 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 predicted value of the remaining effective life of the bearing. Based on the predicted remaining effective life of the bearing and the vibration signal spectrum analysis, key harmonic components related to the bearing deterioration state were identified.

6. The method for noise reduction of the non-drive end bearing of a vertical centrifugal pump motor based on deep learning according to claim 1, characterized in that, The specific process of generating an anti-phase torque pulsation with the phase opposite to the key harmonic component by injecting the compensation harmonic signal is as follows: The edge computing device calculates the amplitude and phase of the corresponding anti-phase harmonics in real time based on the amplitude and phase information of the key harmonic components identified by the deep learning model. The amplitude and phase of the anti-phase harmonics are injected into the motor stator through the motor frequency converter, so that the generated anti-phase torque pulsation cancels out the original key harmonic components in phase. Noise reduction is achieved by continuously adjusting the amplitude and phase of the injected signal in real time.

Images