Bearing health state online evaluation method and system based on morphological profile analysis and federal evolutionary hypergraph

Abstract

The application provides a bearing health state online evaluation method and system based on morphological profile analysis and federal evolution hypergraph, aiming at solving the problems of poor model self-adaptability, difficult cross-device knowledge migration and easy to be submerged early weak fault characteristics of the prior art under dynamic working conditions. The method captures the geometric profile evolution of bearing micro-damage by constructing a morphological multi-scale profile feature extraction engine, topologically preserving morphological decomposition of the vibration signal; adopts a Bayesian Poisson online learning algorithm to realize dynamic threshold adaptive updating and early warning of the health index; introduces an evolutionary hypergraph neural network to model the high-order multi-element fault propagation relationship between the bearing and the adjacent components; finally, through a federal edge collaborative framework, the incremental aggregation and knowledge migration of the cross-device model are realized under the premise of protecting data privacy. The application significantly improves the robustness of bearing fault diagnosis under variable working conditions and the sensitivity of early warning, and provides a lightweight and evolving solution for intelligent operation and maintenance in distributed industrial scenarios.

Description

Technical Field

[0001] This invention relates to the fields of mechanical fault diagnosis and industrial edge intelligence technology, and more specifically, to a method and system for online health assessment, early fault warning and cross-device knowledge transfer for rolling bearings in rotating machinery. Background Technology

[0002] As a core component of rotating machinery, the operational reliability of rolling bearings directly determines the overall lifespan and safety of the equipment. With the popularization of Industrial Internet of Things (IIoT) technology, bearing condition monitoring based on vibration signals has shifted from offline diagnosis to online prediction. However, in complex industrial scenarios, existing technologies still face the following technical bottlenecks: First, extracting weak fault features against a strong noise background is difficult. Early bearing faults (such as microcracks and pitting) generate weak impact signals that are easily masked by environmental noise and operating condition fluctuations. Traditional filtering methods and time-frequency analysis techniques struggle to effectively separate the impact components while preserving the signal topology, leading to a high false negative rate in early warning systems. While mathematical morphology, as a nonlinear signal processing method, has the advantage of preserving the signal's geometric contour, existing morphological filtering methods often use fixed structural elements and lack the ability to adaptively model the multi-scale morphological evolution of signals.

[0003] Secondly, fixed threshold strategies cannot adapt to dynamic changes in operating conditions. Existing health assessment systems typically set fixed alarm thresholds or rely on the statistical distribution of historical data. However, when equipment speed and load fluctuate, the baseline level of vibration energy changes accordingly, and fixed thresholds are prone to false alarms or missed alarms. There is an urgent need for an adaptive threshold mechanism that can sense changes in operating conditions online and dynamically update the confidence intervals of health indicators.

[0004] Secondly, there is a lack of multivariate high-order relationship modeling for fault propagation. Bearing failures are often not isolated events; their degradation process involves complex interactions with adjacent components (such as shafts, gears, and seals). Existing graph neural network methods are mostly limited to binary relationship modeling (i.e., edges between two nodes), making it difficult to express high-order fault propagation patterns caused by the combined effects of multiple components (e.g., bearing inner ring failure + gear meshing impact leading to bearing housing resonance). Although hypergraph neural networks possess high-order modeling capabilities, they have not yet been effectively introduced into bearing fault propagation analysis.

[0005] Finally, the issues of cross-device knowledge transfer and data silos are prominent. In large industrial enterprises, similar equipment is often distributed across different production lines or geographical locations. Due to limitations in data privacy and communication bandwidth, it is difficult to centrally train the entire lifecycle data of each device. How to achieve fault knowledge sharing and model co-evolution across devices while protecting data privacy is a major challenge currently facing industrial intelligence.

[0006] To address the aforementioned issues, this invention proposes a novel technical solution that organically integrates mathematical morphology contour analysis, Bayesian online learning, evolutionary hypergraph neural networks, and federated edge collaboration, providing an adaptive, evolvable, and privacy-preserving comprehensive solution for online assessment of bearing health status. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention provides an online assessment method and system for bearing health status based on morphological contour analysis and federated evolutionary hypergraphs, aiming to achieve sensitive capture of early bearing faults under varying operating conditions, dynamic threshold adaptive updating, multivariate fault propagation modeling, and cross-device privacy-preserving knowledge transfer.

[0008] (I) A method for processing partial discharge monitoring data of power equipment The method includes the following steps: Step S1: Morphological multi-scale contour feature extraction Mathematical morphology methods are used to perform multi-scale decomposition on the original vibration signal and extract morphological contour features that preserve the signal's geometric topology. S1.1 Adaptive Structural Element Generation: Based on the local extremum distribution and impulse period of the signal, a multi-scale structural element sequence is dynamically generated, wherein the scale parameter of the structural element is adaptively adjusted with the instantaneous frequency of the signal.

[0009] S1.2 Morphological Contour Decomposition: Using morphological opening-closing and closing-opening combination operations, the original signal is decomposed into the morphological contour main signal and the morphological residual signal. in, Represents the morphological opening operation. This represents the morphological closing operation, where g is the structuring element. The residual signal contains an impact component that matches the scale of the structuring element.

[0010] S1.3 Multi-scale morphological spectrum construction: Repeat the above decomposition under different scale structural elements to construct a morphological multi-scale spectrum, and extract morphological features such as spectral entropy and spectral kurtosis as the initial representation of the bearing's health status.

[0011] Step S2: Bayesian Poisson Online Learning and Adaptive Threshold Update Based on the morphological features extracted in step S1, a Bayesian Poisson online learning model is constructed to achieve dynamic threshold adaptive updating of health indicators. S2.1 Characteristic Poissonization Modeling: The impulse counting of the morphological residual signal is modeled as a Poisson process, with the impulse occurrence rate... It changes dynamically over time.

[0012] S2.2 Online Bayesian Update: The conjugate prior (Gamma distribution) is used for online Bayesian inference of the impact incidence rate. Whenever new observation data arrives, the posterior distribution is updated according to the Bayesian formula: S2.3 Dynamic Threshold Generation: Based on the posterior predicted distribution of the impact incidence rate, a dynamic confidence interval for the health indicator is generated. The upper bound of the confidence interval is used as the adaptive alarm threshold, and an early warning is triggered when the real-time impact incidence rate exceeds this threshold.

[0013] Step S3: High-order fault propagation modeling of evolutionary hypergraph neural networks Construct an evolutionary hypergraph neural network to model the high-order multivariate fault propagation relationships between the bearing and adjacent components: S3.1 Hypergraph Structure Construction: Using bearings and related components as nodes and failure modes involving multiple nodes as hyperedges, construct the initial hypergraph structure. A superedge can connect two or more nodes. For example, a superedge can connect three nodes at the same time: "bearing inner ring - rolling element - bearing housing", indicating a resonance failure mode under the combined action of the three.

[0014] S3.2 Hypergraph Structure Evolution: When a new morphological impact pattern is detected in step S1 or an early warning is triggered in step S2, the hypergraph structure is dynamically updated—adding nodes, adjusting hyperedge connections, or updating hyperedge weights. The evolution rule is driven by a change detection algorithm based on morphological features.

[0015] S3.3 Hypergraph Neural Network Inference: A hypergraph convolutional network is used to update the node states, and information between nodes is transmitted and aggregated through hyperedges. The final output is a heatmap showing the probability distribution of each node's health state and the fault propagation path.

[0016] Step S4: Cross-device knowledge transfer in federated edge collaboration Construct a federated edge learning framework to enable privacy-preserving knowledge sharing and model co-evolution among multiple devices: S4.1 Local Model Training: Each edge device (corresponding to one bearing monitoring terminal) independently trains the local model in steps S1-S3 using local data, and only uploads the model gradient or model parameters (not the original data) to the federated server.

[0017] S4.2 Differential Privacy Protection: Add random noise that satisfies differential privacy to the gradient before uploading to prevent the original data from being inferred from the gradient.

[0018] S4.3 Federated Aggregation and Knowledge Transfer: The federated server uses the FedAvg algorithm to perform weighted aggregation on the received local models to generate a global model. The global model is then distributed to each edge device to guide the next step of training of the local models, realizing knowledge transfer and collaborative evolution across devices.

[0019] S4.4 Incremental Model Update: When a new device is connected or the device's operating conditions change significantly, only a small amount of local data is needed to fine-tune the global model, which can quickly adapt to the new scenario without having to train from scratch.

[0020] (II) An online bearing health condition assessment system The system is used to implement the above method, including: 1. Signal Acquisition and Morphological Preprocessing Module: Includes an accelerometer, anti-aliasing filter, analog-to-digital converter, and morphological operation unit, used to acquire high-frequency vibration signals and extract multi-scale morphological contour features.

[0021] 2. Bayesian Online Learning and Adaptive Early Warning Module: This module includes a Poisson process modeling unit, an online Bayesian inference engine, and a dynamic threshold generator, used to achieve real-time updates of health indicators and trigger early warnings.

[0022] 3. Evolutionary Hypergraph Inference Module: This module includes a hypergraph construction unit, an evolutionary rule engine, and a hypergraph neural network computation unit. It is used to model high-order fault propagation relationships and output fault path heatmaps.

[0023] 4. Federated Edge Collaboration Module: Includes local model training unit, differential privacy noise adder, federated communication interface and model aggregation engine, used to realize privacy-preserving knowledge transfer and model co-evolution across devices.

[0024] (III) Innovations of this invention 1. Deep integration of morphological contour analysis and Bayesian Poisson online learning: The innovative approach combines the topology preservation property of mathematical morphology with the adaptive update capability of Bayesian online learning to construct a complete technical chain of "morphological feature extraction - Poisson impact counting - dynamic threshold early warning", which solves the dual problems of difficulty in capturing early weak faults in strong noise background and poor adaptability of fixed threshold.

[0025] 2. High-order fault propagation modeling using evolutionary hypergraph neural networks: Breaking through the limitation of traditional graph neural networks being restricted to binary relationships, a hypergraph structure is introduced to model high-order fault modes under the combined action of multiple components. A hypergraph evolution mechanism based on morphological change detection is designed, which enables the fault propagation model to be dynamically adjusted as the equipment deteriorates, significantly improving the accuracy of fault root cause tracing.

[0026] 3. A cross-device knowledge transfer framework for federated edge collaboration: To address data privacy and communication constraints in industrial scenarios, a knowledge transfer paradigm of "local training - differential privacy - federated aggregation - incremental fine-tuning" was designed. This enables the collaborative evolution of fault diagnosis knowledge among multiple devices without sharing the original data, providing a feasible solution for distributed industrial intelligence.

[0027] A method for constructing a health index based on morphological spectral entropy: A multi-scale morphological spectrum and its spectral entropy calculation method are proposed as a novel characterization index for bearing health status. This index is sensitive to early weak shocks and has natural robustness to operating condition fluctuations, making up for the shortcomings of traditional time-frequency domain indices under non-stationary operating conditions. Attached Figure Description

[0028] Figure 1 This is a diagram of the overall system architecture of the present invention, showing the interrelationships and data flow of the four major modules: morphological contour analysis, Bayesian online learning, evolutionary hypergraph inference, and federated edge collaboration.

[0029] Figure 2 The flowchart for morphological multi-scale contour feature extraction demonstrates the process of adaptive structuring element generation, morphological opening and closing operations, and multi-scale morphological spectrum construction.

[0030] Figure 3 This diagram illustrates Bayesian Poisson online learning and adaptive threshold update, demonstrating the evolution of the posterior distribution of the impact incidence rate and the generation mechanism of the dynamic threshold.

[0031] Figure 4 To illustrate the structure of an evolutionary hypergraph neural network, this paper demonstrates the hypergraph construction, hyperedge connections, evolutionary update mechanism, and reasoning process of hypergraph convolution.

[0032] Figure 5 The flowchart for cross-device knowledge transfer in federated edge collaboration demonstrates the complete process of local training, differential privacy noise addition, federated aggregation, and incremental fine-tuning. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Example 1: Early Warning of Bearings Based on Morphological Contour Analysis and Bayesian Online Learning This embodiment details how to use morphological contour analysis combined with Bayesian Poisson online learning to achieve sensitive detection and adaptive early warning of bearing faults.

[0035] Step 1: High-frequency vibration signal acquisition and preprocessing A PCB 356A16 accelerometer (sensitivity 10 mV / g) was mounted horizontally and vertically on the test bearing housing to continuously acquire vibration data at a sampling rate of 256 kHz. A double-buffering mechanism was used for data acquisition to ensure uninterrupted sampling. Preprocessing included: removal of the DC component and anti-aliasing filtering (cutoff frequency 100 kHz).

[0036] Step 2: Morphological multi-scale contour feature extraction uses flat structuring elements (straight line segments of length L), with L ranging from 3 to 51 sampling points (corresponding to impact pulse widths of approximately 0.012 ms to 0.2 ms). Morphological opening-closing and closing-opening combination operations are performed on the vibration signal of each basic window (duration 10 ms) to extract the morphological residual signal.

[0037] Taking scale L=21 as an example, the morphological contour decomposition process is as follows: ; ; in, This is the residual signal, which contains an impact component that matches the scale of the structural element.

[0038] Step 3: Multi-scale morphological spectrum construction. The above decomposition is repeated at 8 scales (L=3, 7, 11, 15, 21, 31, 41, 51), and the energy of the residual signal at each scale is calculated. A morphological multi-scale spectrum was constructed. Further calculation of the morphological spectral entropy was performed. ; The lower the spectral entropy, the more concentrated the energy is at a few scales, corresponding to a clear fault characteristic frequency; the higher the spectral entropy, the more dispersed the energy is, corresponding to a healthy state or complex disturbances.

[0039] Step 4: Bayesian Poisson Online Learning and Dynamic Threshold Generation: This involves calculating the number of morphological impacts detected per second. (Defined as the peak count of the residual signal exceeding the adaptive threshold) is modeled as a Poisson process: ; Impact incidence It changes slowly over time. The Gamma distribution is used as... Conjugate priors: ; in, , These are the shape parameter and the rate parameter.

[0040] The online Bayesian update rules are as follows: a priori: Observed Then, post-hoc update: ; ; To adapt to changes in operating conditions, a forgetting factor is introduced. Discounting historical information: ; ; Dynamic threshold Generated based on the 95th percentile of the posterior predicted distribution: ; in, It is the inverse cumulative distribution function of the negative binomial distribution (Poisson-Gamma composite distribution).

[0041] Real-time impact count Exceeding the dynamic threshold three times in a row This triggers an early warning.

[0042] Example 2: High-order fault propagation analysis based on evolutionary hypergraph neural network Based on Example 1, this embodiment further constructs an evolutionary hypergraph neural network to model the multivariate fault propagation relationship between the bearing and adjacent components.

[0043] Step 1: Hypergraph Structure Initialization. Key components in the monitoring system are used as nodes, including: bearing inner ring (IR), bearing outer ring (OR), rolling elements (RB), shaft (SH), bearing housing (HS), gear (GR), etc. Initial number of nodes. .

[0044] Hyperedge set Construction rules: Binary hyperedges (i.e. traditional edges): connect two physically contacting components, such as (IR, SH) or (OR, HS).

[0045] Triple and higher-order hyperedges: constructed based on prior fault mode knowledge. For example, hyperedges... = {IR, RB, OR} represents the load zone where the rolling element is in contact with both the inner and outer rings simultaneously; (Extra edge) = {RB, GR, SH} represents a combined failure mode in which the rolling elements mesh with the gears through shaft coupling.

[0046] Step Two: Hypergraph Evolution Mechanism When the morphological detection algorithm in Example 1 identifies a new impact pattern, it triggers hypergraph evolution: Add new nodes: If the impact mode frequency characteristics point to an unmodeled component (such as a cage), then add a new node dynamically.

[0047] Hyperedge weight update: Hyperedge weight Dynamic adjustment based on the correlation of morphological features of nodes within the hyperedge: ; in, Let be the morphological feature vector of node i. For reference fault mode characteristics.

[0048] Add a new hyperedge: When multiple nodes are detected to have abnormal feature synchronization, add a new hyperedge between them.

[0049] Step 3: Hypergraph Neural Network Inference uses a Hypergraph Convolutional Network for node state updates. The hypergraph convolutional layer is defined as follows: ; in, It is an incidence matrix. This indicates that node i belongs to hyperedge e; For the superedge weight; These are learnable parameters.

[0050] After two layers of hypergraph convolution, node features are input into a fully connected layer, which outputs a health status classification for each node (healthy, early degradation, severe failure). Simultaneously, based on an attention mechanism, the importance of hyperedges is calculated to generate a failure propagation heatmap, visualizing the propagation path of failures from the source node to other nodes.

[0051] Example 3: Cross-device knowledge transfer through federated edge collaboration This embodiment focuses on privacy-preserving knowledge sharing and model co-evolution in multi-device scenarios.

[0052] Scenario: A wind farm has 20 wind turbines, and the monitoring system of this invention is deployed on the main bearing of each turbine. The operating conditions of each turbine are similar but have slight differences, and due to data privacy policy restrictions, the raw vibration data cannot leave the local system.

[0053] Step 1: Local model initialization - the federated server issues initial global model parameters. Provided to all edge devices. The initial model is a pre-trained base model (trained on laboratory data or public datasets).

[0054] Step 2: Local model training and update in the... In round-fed communication, each edge device k utilizes local data D k Perform local training for E rounds, minimizing the local loss function: ; The second term is the KL divergence regularization term, which is used to constrain the local model from deviating too far from the global model and avoid catastrophic forgetting.

[0055] Step 3: Differential Privacy Noise Addition and Local Model Update Before uploading, add the following: Gaussian noise for differential privacy: ; Noise scale Clipping threshold based on gradient norm and privacy budget calculate.

[0056] Step 4: Federated Aggregation and Global Model Update The federated server collects the noisy local updates and aggregates them using a federated averaging algorithm. ; in, For equipment Local data volume This is the aggregate learning rate.

[0057] Step 5: Incremental Model Adaptation for New Equipment. When the 21st wind turbine is put into operation (without historical fault data), the federated server will adapt the current global model. The data is then distributed to the new equipment. The new equipment only needs to collect 24 hours of normal operating data and make minor adjustments to the global model (usually only 1-2 rounds) to quickly adapt to local operating conditions and achieve fault diagnosis capabilities with zero or few samples.

Claims

1. A method for processing partial discharge monitoring data of power equipment, characterized in that: The method includes the following steps: Step S1: Morphological multi-scale contour feature extraction Mathematical morphology methods are used to perform multi-scale decomposition on the original vibration signal and extract morphological contour features that preserve the signal's geometric topology. S1.1 Adaptive Structural Element Generation: Based on the local extremum distribution and impulse period of the signal, a multi-scale structural element sequence is dynamically generated, wherein the scale parameter of the structural element is adaptively adjusted with the instantaneous frequency of the signal. S1.2 Morphological Contour Decomposition: Using morphological opening-closing and closing-opening combination operations, the original signal is decomposed into the morphological contour main signal and the morphological residual signal. in, Represents the morphological opening operation. This represents the morphological closing operation, where g is the structuring element; the residual signal contains an impact component that matches the scale of the structuring element. S1.3 Multi-scale morphological spectrum construction: Repeat the above decomposition under different scale structural elements to construct a morphological multi-scale spectrum, and extract morphological features such as spectral entropy and spectral kurtosis as the initial representation of the bearing's health status. Step S2: Bayesian Poisson Online Learning and Adaptive Threshold Update Based on the morphological features extracted in step S1, a Bayesian Poisson online learning model is constructed to achieve dynamic threshold adaptive updating of health indicators. S2.1 Characteristic Poissonization Modeling: The impulse counting of the morphological residual signal is modeled as a Poisson process, with the impulse occurrence rate... It changes dynamically over time; S2.2 Online Bayesian Update: The conjugate prior (Gamma distribution) is used for online Bayesian inference of the impact incidence rate; whenever new observation data arrives, the posterior distribution is updated according to the Bayesian formula: S2.3 Dynamic Threshold Generation: Based on the posterior predicted distribution of the impact incidence rate, a dynamic confidence interval for the health indicator is generated; the upper bound of the confidence interval is used as the adaptive alarm threshold, and an early warning is triggered when the real-time impact incidence rate exceeds this threshold; Step S3: High-order fault propagation modeling of evolutionary hypergraph neural networks Construct an evolutionary hypergraph neural network to model the high-order multivariate fault propagation relationships between the bearing and adjacent components: S3.1 Hypergraph Structure Construction: Using bearings and related components as nodes and failure modes involving multiple nodes as hyperedges, construct the initial hypergraph structure. A superedge can connect two or more nodes. For example, a superedge can connect three nodes at the same time: "bearing inner ring - rolling element - bearing housing", indicating a resonance failure mode under the combined action of the three. S3.2 Hypergraph Structure Evolution: When a new morphological impact pattern is detected in step S1 or an early warning is triggered in step S2, the hypergraph structure is dynamically updated—adding nodes, adjusting hyperedge connections, or updating hyperedge weights; the evolution rules are driven by a change detection algorithm based on morphological features. S3.3 Hypergraph Neural Network Inference: A hypergraph convolutional network is used to update the node state, and information between nodes is transmitted and aggregated through hyperedges; finally, a heatmap of the health status probability distribution of each node and the fault propagation path is output. Step S4: Cross-device knowledge transfer in federated edge collaboration Construct a federated edge learning framework to enable privacy-preserving knowledge sharing and model co-evolution among multiple devices: S4.1 Local Model Training: Each edge device (corresponding to one bearing monitoring terminal) independently trains the local model in steps S1-S3 using local data, and only uploads the model gradient or model parameters (not the original data) to the federated server. S4.2 Differential Privacy Protection: Add random noise that satisfies differential privacy to the gradient before uploading to prevent the original data from being inferred from the gradient; S4.3 Federated Aggregation and Knowledge Transfer: The federated server uses the FedAvg algorithm to perform weighted aggregation on the received local models to generate a global model; the global model is distributed to each edge device to guide the next step of training of the local models, realizing knowledge transfer and collaborative evolution across devices; S4.4 Incremental Model Update: When a new device is connected or the device's operating conditions change significantly, only a small amount of local data is needed to fine-tune the global model, which can quickly adapt to the new scenario without having to train from scratch.

2. An online bearing health status assessment system that implements the power equipment partial discharge monitoring data processing method as described in claim 1, characterized in that: The system includes: Signal acquisition and morphological preprocessing module: including accelerometer, anti-aliasing filter, analog-to-digital converter and morphological operation unit, used to acquire high-frequency vibration signals and extract morphological multi-scale contour features; Bayesian Online Learning and Adaptive Early Warning Module: Includes a Poisson process modeling unit, an online Bayesian inference engine, and a dynamic threshold generator, used to achieve real-time updates of health indicators and early warning triggering; Evolutionary hypergraph reasoning module: includes hypergraph construction unit, evolutionary rule engine and hypergraph neural network computing unit, used to model high-order fault propagation relationships and output fault path heatmap; The Federation Edge Collaboration Module includes a local model training unit, a differential privacy noise adder, a federated communication interface, and a model aggregation engine, which are used to achieve privacy-preserving knowledge transfer and model co-evolution across devices.

Images