Health assessment method for coupling based on fusion scale-down test

Abstract

A coupling health assessment method based on fusion scaled-down testing is proposed. In the offline phase, source domain scaled-down model label datasets and source domain training sets are collected and generated. These datasets are used to train a multi-source domain semi-supervised scaled-down model fault state transfer network, which includes a multi-source domain feature extractor, classifier, domain classifier, and gradient inversion layer. In the online phase, the trained multi-source domain semi-supervised scaled-down model fault state transfer network is used for real-time coupling health assessment. This invention can simultaneously consider the motion, load, and interaction of various components in the actual system, as well as the elasticity and fatigue characteristics of materials in the actual system. It constructs a geometrically and materially fused scaled-down model, ensuring that the scaled-down model maintains a high degree of similarity to actual working conditions, providing accuracy for the scaled-down test. Furthermore, through transfer learning, the scaled-down test can be transferred to the actual system, enhancing the practicality and adaptability of the scaled-down model.

Description

Technical Field

[0001] This invention relates to a technology in the field of coupling manufacturing, specifically a coupling health assessment method based on a fusion scaling test. Background Technology

[0002] A coupling system is a transmission system that can transmit power and compensate for relative displacements at both ends of the transmission caused by various factors such as installation and operating conditions, while exerting minimal force and additional torque on both ends of the transmission during displacement compensation. Long-term operation in complex environments poses various risks of failure. These failures can lead to a significant decrease in transmission efficiency, increase the risk of equipment damage, noise and vibration, and may even cause safety accidents. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a coupling health assessment method based on fusion scaled-down tests. This method can simultaneously consider the motion, load, and interaction of various components in the actual system, as well as the elasticity and fatigue characteristics of materials in the actual system. It constructs a geometric and material fusion scaled-down model, which maintains a high degree of similarity to actual working conditions, improves the accuracy of scaled-down tests, and enhances the practicality and adaptability of the scaled-down model by transferring the scaled-down tests to the actual system through transfer learning.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a coupling health assessment method based on fusion scale-down testing. In the offline stage, source domain scale-down model label dataset and source domain training set are collected and generated to train a multi-source domain semi-supervised scale-down model fault state transfer network, which includes a multi-source domain feature extractor, classifier, domain classifier and gradient inversion layer (GRL). In the online stage, the trained multi-source domain semi-supervised scale-down model fault state transfer network is used to perform real-time coupling health assessment.

[0006] The source domain scaling model label dataset was obtained in the following way:

[0007] Step 1: Geometric scaling modeling of the coupling system was performed. By appropriately scaling the key geometric parameters of the system, the experimental model could be reduced in size while maintaining physical similarity. Next, material scaling modeling of the coupling system was performed. By appropriately scaling the key material parameters of the system, the experimental model could be manufactured using low-cost materials while maintaining physical similarity. Then, hybrid scaling modeling of the geometry and materials of the coupling system was performed. By appropriately scaling the key geometric and material parameters of the system, the scaling characteristics of both geometry and materials were combined, allowing the scaled model to simulate and analyze the performance of the coupling system under laboratory conditions.

[0008] The coupling system includes a diaphragm and a main shaft.

[0009] The key geometric parameters are: 2 diaphragms, diaphragm diameter of 382 mm, and spindle length of 726 mm.

[0010] The coupling system is made of titanium alloy.

[0011] The key material parameters are Young's modulus and Poisson's ratio.

[0012] The low-cost materials mentioned are PLA and ABS.

[0013] Step 2: Collect signal characteristics from the scaled-down test as the source task and the actual coupling system's condition diagnosis task as the target task.

[0014] The main difference between the target task and the source task lies in the data source. The target task data comes from the actual coupling system and lacks labels, while the source task data comes from scaled-down tests and has both normal and fault label information.

[0015] The signal characteristic described is a vibration signal.

[0016] The aforementioned condition diagnosis task refers to diagnosis through vibration signals.

[0017] The training mentioned refers to: redistributing the loss function of the fault state transition network of the multi-source domain semi-supervised scaled-down model, removing the prediction loss using unlabeled data, and pre-training using only the scaled-down model data of the source domain flexible shaft coupling system, with the pre-training learning rate dynamically adjusted.

[0018] The aforementioned redistribution of the loss function refers to: Ly′ = Ly + βLfault, increasing the loss function L. fault It is used to calculate the classification loss of normal and faulty data.

[0019] The aforementioned dynamic adjustment of the learning rate refers to: an initial learning rate of 2e 2 Learning rate according to Attenuation helps balance training efficiency and accuracy.

[0020] This invention relates to a system for implementing the above-mentioned method, comprising: a data acquisition and preprocessing unit that acquires operational data of the actual coupling system and the geometrically and materially fused scaled-down model at different time points, preprocesses the acquired data to generate input feature vectors suitable for the fault state transfer network of the multi-source domain semi-supervised scaled-down model, and a dataset containing features of the source task and the target task, for subsequent network training and transfer learning; and a transfer learning unit that replaces the general similarity criterion by aligning the feature spaces of the source domain scaled-down model data and the target domain data to reduce the differences between domains, adapt to new environments or tasks, and matches the feature patterns of the source domain and the target domain according to the feature data of the current target domain, and updates the model parameters in real time to generate the transferred feature representation and the corresponding results.

[0021] The aforementioned operational data is a vibration signal.

[0022] The preprocessing includes data cleaning, feature extraction, and standardization.

[0023] Technical effect

[0024] This invention addresses the problems of high cost, long cycle, low data utilization, and difficulty in mapping scaled-down models to actual systems in traditional coupling health assessment tests. Through a collaborative design involving geometric and material-integrated scaled-down modeling and multi-source domain semi-supervised transfer learning, it achieves more efficient, accurate, and practical coupling health assessment. By constructing a hybrid geometric and material scaled-down model and using low-cost 3D printing materials such as PLA and ABS instead of high-cost titanium alloys, the manufacturing and wear costs of test parts are significantly reduced while maintaining consistency with the prototype in core dynamic characteristics such as natural frequencies. Simultaneously, the scaled-down model is compact and has a short processing cycle, eliminating the need for a full-size test platform and significantly shortening the test iteration cycle, thus solving the pain point of traditional full-size tests being difficult to conduct on a large scale. Breaking through the limitations of traditional single geometric scaling, it comprehensively considers the synergistic influence of key geometric and material parameters of the coupling system, ensuring that the scaled-down model is highly similar to the actual system in terms of dynamic response and vibration characteristics. By using a multi-source domain feature extraction module and an adversarial training mechanism, the nonlinear mapping problem caused by geometric and material scaling is effectively overcome, and the feature space alignment between scaled-down experimental data and actual system data is achieved. By combining a pre-training strategy to optimize the loss function, the fault diagnosis of the actual system in the target domain can be completed using only the label data of the scaled-down model in the source domain, which greatly reduces the dependence on the labeled data of the actual system and improves the generalization ability and adaptability of the model. Attached Figure Description

[0025] Figure 1This is a flowchart of the present invention;

[0026] Figure 2 This is a diagram illustrating the effect of this implementation. Detailed Implementation

[0027] like Figure 1 As shown in this embodiment, a coupling health assessment method based on a fusion scaling test is provided, including:

[0028] Step 1: Establish a scaled-down model integrating geometry and materials, specifically including:

[0029] 1.1 Set the core parameters of the prototype coupling. The prototype is based on the titanium alloy flexible shaft coupling system commonly used in the aerospace field. Its key geometric and material parameters are as follows: Geometric parameters: number of diaphragms: 2, diaphragm diameter: 382 mm, main shaft length: 726 mm; Material parameters: titanium alloy, Young's modulus: 110 GPa, Poisson's ratio: 0.33; Core dynamic indicators: the natural frequencies and critical speeds of the first 10 modes of the prototype are determined by Ansys modal analysis.

[0030] 1.2 Geometric Scale-Down Modeling: Based on similarity theory, the key geometric parameters of the prototype were scaled to ensure the dynamic similarity between the scaled-down model and the prototype. To simulate the actual rotor system operating conditions, shaft connectors and bearing constraints were added to both ends of the scaled-down model to unify the total system length to 1200mm. The final geometric parameters of the scaled-down model were determined to be a diaphragm diameter of 64mm and a main shaft length of 90mm.

[0031] 1.3 Material Scale-Down Modeling: Low-cost 3D printing materials were used to replace titanium alloy. Similarity in mechanical properties was ensured through material parameter matching. PLA and ABS materials were selected. Ansys modal simulation was used to compare the Campbell plots of titanium alloy, PLA, and ABS materials, confirming that the dynamic response trends of the two 3D printing materials were consistent with those of titanium alloy, meeting the requirements of the scale-down test.

[0032] 1.4 Hybrid scale model fabrication: Combining geometric scaling parameters and material selection, the final scale model was fabricated using 3D printing technology. Modal testing was then performed on the printed scale model. Vibration signals were collected using an accelerometer to verify that the deviation between its natural frequency and the simulation results was ≤3%, ensuring the effectiveness of the model. The printing parameters are shown in Table 1.

[0033] Table 1

[0034] Step 2: Collect data on the characteristics of different operating conditions of the same coupling system and establish a dataset, specifically including:

[0035] 2.1 Select key operating parameters of the coupling system as feature acquisition objects, including: vibration signal (acquisition location: bearing seats at both ends of the scaled model and the drive end bearing of the actual system, sensor model: PCB 352C33, IMI industrial 2-pin) and speed signal (acquisition location: motor output shaft, sensor model: HENGSTLER RI58-O / 5000AR.12KB), with sampling frequency set to 50kHz for the scaled model and 1024Hz for the actual system.

[0036] 2.2 Data Acquisition and Preprocessing: After collecting historical operating data of the coupling system under different working conditions, the collected data is preprocessed. Specifically, after collecting data under different materials (PLA / ABS), speeds (900rpm / 1800rpm), misalignment conditions (0mm / 3mm / 5mm), and fault conditions (diaphragm crack), data cleaning (removing noise and outliers using the 3σ criterion), normalization (normalizing the vibration signal amplitude to the [-1,1] interval to unify the data scale) and feature extraction (such as time domain features: peak value, RMS value, kurtosis; frequency domain features: spectral peak value, harmonic components, etc.) is performed.

[0037] 2.3 Construction of feature space: Sort the data according to the combination of working conditions to ensure that the feature data of the source domain and the target domain have continuity and consistency in terms of working conditions.

[0038] Preferably, the feature data is stored and each data point is labeled with a corresponding operating condition label (material, rotational speed, misalignment, fault status) and feature label, which facilitates the training of the subsequent transfer learning model and the implementation of health assessment.

[0039] 2.4 Data processing and storage: Sort the data according to timestamps to ensure that the feature data of the source domain and the target domain have continuity and consistency in time.

[0040] Preferably, the feature data is stored and each data point is labeled with a corresponding timestamp, operating status (normal or faulty), and feature label, which facilitates the training of the reinforcement learning model and the implementation of transfer learning.

[0041] Step 3: Train a multi-source domain semi-supervised scaled-down model fault state transfer network using the dataset obtained in Step 2. This enables knowledge transfer from the scaled-down model to the actual system, thereby completing the coupling health assessment and outputting the assessment results. Specifically, this includes:

[0042] 3.1 A multi-source domain semi-supervised scaled-down model fault state transfer network is constructed. This network includes: a multi-source domain feature extractor, a classifier, a domain classifier, and a gradient inversion layer (GRL). The multi-source domain feature extractor is a 16-layer Bi-LSTM network (input size 128, hidden layers 32, positional encoding dimension 128) used for semi-supervised feature extraction of different materials, rotational speeds, and misalignment conditions. The classifier includes 4 layers of one-dimensional CNN and 2 layers of fully connected layers, with an output dimension of 2 (normal / fault), used for fault state classification. The domain classifier is a fully connected network used to distinguish whether features come from the source domain (scaled-down model) or the target domain (actual system). The gradient inversion layer is located between the feature extractor and the domain classifier and introduces negative gradients during backpropagation to achieve adversarial training.

[0043] 3.2 Model Pre-training: Pre-training was performed using only the source domain scaled-down model label data obtained in step 1, with the data volume being 10 times that of the target domain data; the adjusted loss function L was used. y =L y +βL fault , where: L y For classification loss, L fault The loss for normal / fault classification is set to β = 0.5. The training parameters are set as follows: 50 training epochs, initial learning rate 2e-2, weight decay 2e-3, batch size 4096, and AdamW optimizer is used. The learning rate is dynamically decayed according to α←0.95α-2e-3 to balance training efficiency and accuracy.

[0044] 3.3 Model Fine-tuning and Adversarial Training: The source domain scaled-down model label data obtained in step 1 is mixed with the target domain label data from the real experiment (the first 5% are normal and the last 5% are faulty) for model fine-tuning. Through adversarial learning between the feature extractor and the domain classifier, the feature extractor learns the common features of the source and target domains, achieving feature space alignment and making it impossible for the domain classifier to distinguish the source of features. The specific fine-tuning parameters are set as follows: 1000 training epochs, learning rate 1e-3, weight decay 1e-4, batch size 512, and an early stopping strategy is introduced (training stops if the validation set loss does not decrease for 30 consecutive iterations).

[0045] 3.4 Health Assessment Implementation: The trained network is deployed to the actual coupling monitoring system to collect vibration signals of the actual system in real time. After preprocessing in step 2.2, feature vectors are extracted and input into the model. The model outputs the fault probability distribution of the actual system. If the fault probability is ≥80%, it is judged as a fault state and an early warning is issued; if the fault probability is <80%, it is judged as a normal operating state.

[0046] The real-time acquisition of vibration signals from the actual system refers to: placing an accelerometer in the bearing housing at the drive end, installing the coupling system on a rotary test bench, and using a speed sensor to collect the motor output shaft speed data. The sampling frequency is 50kHz for the scaled-down model and 1024Hz for the actual system. The sampling duration is 360s for the scaled-down model and 60 minutes for the actual system. The coupling system data is shown in Table 2.

[0047] Table 2

[0048] The health assessment of the coupling system was conducted using this invention, and the experimental results are shown in Table 3.

[0049] Table 3

[0050] Compared with existing technologies, this invention significantly reduces experimental costs and cycles by integrating geometry and materials in scaled-down modeling, and improves the accuracy and reliability of scaled-down experiments. It achieves efficient knowledge transfer from scaled-down models to actual systems through multi-source domain semi-supervised transfer learning, solving the nonlinear mapping problem caused by geometric and material scaling and reducing dependence on labeled data from actual systems. It enhances the comprehensiveness and robustness of health assessment, simplifies experimental procedures and engineering application thresholds, and facilitates its widespread application in engineering scenarios in multiple fields such as aerospace and industrial automation.

[0051] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for assessing the health of couplings based on a scaled-down fusion test, characterized in that, By collecting and generating source domain scaled-down model label datasets and source domain training sets in the offline phase, the constructed multi-source domain semi-supervised scaled-down model fault state transfer network is trained. In the online phase, the trained multi-source domain semi-supervised scaled-down model fault state transfer network is used for real-time coupling health assessment. The multi-source domain semi-supervised scaled-down model fault state transfer network includes: a multi-source domain feature extractor, a classifier, a domain classifier, and a gradient inversion layer (GRL). Specifically: the multi-source domain feature extractor is a 16-layer Bi-LSTM network that performs semi-supervised extraction of features from different materials, rotational speeds, and misalignment conditions; the classifier includes four 1D CNN layers and two fully connected layers for fault state classification; the domain classifier is a fully connected network used to distinguish whether features originate from the scaled-down model or the actual system; and the gradient inversion layer is located between the feature extractor and the domain classifier and introduces negative gradients during backpropagation to achieve adversarial training.

2. The coupling health assessment method based on fusion scaling test according to claim 1, characterized in that, The source domain scaling model label dataset was obtained in the following way: Step 1: Geometric scaling modeling of the coupling system was performed. By appropriately scaling the key geometric parameters of the system, the experimental model could be reduced in size while maintaining physical similarity. Secondly, material scaling modeling of the coupling system was performed. By appropriately scaling the key material parameters of the system, the experimental model could be manufactured using low-cost materials while maintaining physical similarity. Next, hybrid scaling modeling of the geometry and materials of the coupling system was performed. By appropriately scaling the key geometric and material parameters of the system, the scaling characteristics of both geometry and materials were combined, enabling the scaled model to simulate and analyze the performance of the coupling system under laboratory conditions. Step 2: Collect signal characteristics from the scaled-down test as the source task and the actual coupling system's condition diagnosis task as the target task.

3. The coupling health assessment method based on fusion scaling test according to claim 2, characterized in that, The coupling system includes a diaphragm and a main shaft; The key geometric parameters mentioned are: the number of diaphragms is 2, the diaphragm diameter is 382mm, and the main shaft length is 726mm; The coupling system is made of titanium alloy; The key material parameters are Young's modulus and Poisson's ratio; The low-cost materials mentioned are PLA and ABS; The main difference between the target task and the source task lies in the difference in data source. The target task data comes from the actual coupling system and lacks labels, while the source task data comes from scaled-down tests and has both normal and fault label information. The signal characteristic described is a vibration signal; The aforementioned condition diagnosis task refers to diagnosis through vibration signals.

4. The coupling health assessment method based on fusion scaling test according to claim 1, characterized in that, The training mentioned refers to: redistributing the loss function of the fault state transition network of the multi-source domain semi-supervised scaled-down model, removing the prediction loss using unlabeled data, and pre-training using only the scaled-down model data of the source domain flexible shaft coupling system, with the pre-training learning rate dynamically adjusted.

5. The coupling health assessment method based on fusion scaling test according to claim 4, characterized in that, The aforementioned redistribution of the loss function refers to: Ly′ = Ly + βLfault, increasing the loss function L. fault , used to calculate the classification loss of normal and fault data; The aforementioned dynamic adjustment of the learning rate refers to: an initial learning rate of 2e 2 Learning rate according to Attenuation helps balance training efficiency and accuracy.

6. A coupling health assessment system based on a fusion scaling test, implementing the method of any one of claims 1-5, characterized in that, include: The data acquisition and preprocessing unit collects the actual coupling system and the geometric and material fusion scaled-down model's running data at different time points. The collected data is preprocessed to generate input feature vectors suitable for the fault state transfer network of the multi-source domain semi-supervised scaled-down model, as well as a dataset containing source task and target task features, for subsequent network training and transfer learning. Transfer learning units replace general similarity criteria. By aligning the feature spaces of scaled-down model data from the source domain and data from the target domain, they reduce the differences between domains, adapt to new environments or tasks, match feature patterns between the source and target domains based on the feature data of the current target domain, update model parameters in real time, and generate transferred feature representations and corresponding results. The aforementioned operational data is a vibration signal; The preprocessing includes data cleaning, feature extraction, and standardization.

Images