Sensorless roller chain state monitoring method and device

Abstract

The invention discloses a sensorless roller chain state monitoring method and device. The method comprises the steps that S1, system state monitoring data are collected from the interior of a motor driver of a driving system; s2, slicing the original torque data of the driver according to a motor position signal in the system state monitoring data to obtain a plurality of torque data segments; s3, cyclic spectrum analysis is carried out on the torque data segment, and frequency domain characteristics representing the periodic operation state of the system are obtained; s4, inputting the frequency domain features into a graph attention network for representation learning and sequence reconstruction to obtain a prediction result; and S5, evaluating the health state of the roller chain according to the difference between the prediction result and the frequency domain feature. Through an innovative sensorless monitoring scheme, the problems that in traditional roller chain state monitoring, sensors are difficult to install, the cost is high, and data are seriously interfered by noise are effectively solved. And by utilizing the existing torque and position signals in the motor driver, the system complexity and the maintenance cost are obviously reduced.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring of rotating machinery systems, and specifically to a sensorless method and apparatus for monitoring the condition of roller chains. Background Technology

[0002] With the advancement of technology, the reliability and safety of equipment are crucial for intelligent manufacturing. As the manufacturing industry accelerates towards intelligence and precision, rotating machinery, as a core transmission component, directly determines the operational efficiency, product quality, and safety of the entire production system. Traditional passive maintenance strategies only take action after mechanical failure, while preventative maintenance requires regular inspections and heavily relies on the professional experience of engineers. In contrast, predictive maintenance based on condition monitoring analyzes machine operating data to continuously assess the machine's health and accurately predict its remaining service life. This not only prevents sudden machine failures but also ensures high production quality, optimizes maintenance plans, and reduces economic losses.

[0003] In recent years, deep learning technology has been widely used in the field of mechanical fault diagnosis and has achieved remarkable results. However, condition monitoring still faces some challenges that need to be addressed.

[0004] (1) Data acquisition costs are high in most condition monitoring methods. For continuously moving components such as roller chains, the installation of contact sensors (such as vibration sensors) is extremely challenging, while non-contact sensors (such as acoustic sensors) are easily affected by environmental noise. Furthermore, prior knowledge or complex simulation analysis is required to determine the sensor installation location to ensure the collection of relevant data. The large amount of data generated during long-term monitoring puts enormous pressure on data storage and management.

[0005] (2) Existing research has largely focused on bearings and gears, while other critical rotating components, such as roller chains, which are also widely used, have received little attention, and their failure modes differ significantly. In roller chains, the load is distributed across multiple links, resulting in uneven wear, while the load on bearings and gears is more concentrated. Bearings and gears typically operate in a closed and well-lubricated environment, while roller chains often operate in the opposite way. Furthermore, roller chains gradually elongate due to wear on pins and sleeves, making their failure modes more uncertain.

[0006] (3) Existing monitoring methods typically extract health indicators in the time domain, frequency domain, or time-frequency domain, which often require complex data preprocessing and their effectiveness is heavily dependent on specific working conditions. Traditional machine learning methods rely on manual feature engineering such as data preprocessing and feature extraction, while deep learning models may overfit to noise and are difficult to capture global and local dependencies effectively at the same time.

[0007] Therefore, it is necessary to develop a new mechanical condition monitoring method that can effectively track the degradation process of roller chains, and is low in cost and does not require additional sensors, thus overcoming the inconveniences and defects of the existing technologies. Summary of the Invention

[0008] Based on this, this invention patent establishes a sensorless roller chain condition monitoring method based on multivariate cyclic graph attention network and dynamic time warping. This method utilizes sensorless data and physical information graph attention network to achieve accurate and robust fault prediction under different working conditions.

[0009] To achieve the above objectives, the present invention provides the following solution:

[0010] A sensorless roller chain condition monitoring method includes the following steps:

[0011] S1. Collect system status monitoring data from the motor driver of the drive system;

[0012] S2. Based on the motor position signal in the system status monitoring data, slice the original torque data of the driver to obtain several torque data segments;

[0013] S3. Perform cyclic spectrum analysis on the torque data segment to obtain the frequency domain characteristics characterizing the periodic operating state of the system;

[0014] S4. Input the frequency domain features into the graph attention network for representation learning and sequence reconstruction to obtain the prediction result;

[0015] S5. Evaluate the health status of the roller chain based on the difference between the prediction result and the frequency domain characteristics.

[0016] Preferably, in step S1, when collecting system status monitoring data, key parameters are also defined, including sampling frequency and measurement interval.

[0017] Preferably, S2 specifically includes: dividing the original torque data into multiple logically independent data segments according to the chain rotation cycle based on the motor position signal, so as to isolate the torque information within each chain rotation cycle.

[0018] Preferably, S3 specifically includes:

[0019] Cyclic spectrum analysis was applied to each torque data segment to transform it from the time domain to the frequency domain and extract cyclic stationary features.

[0020] The cyclic spectrum analysis is a two-dimensional Fourier transform of the instantaneous autocorrelation function.

[0021] Preferably, S4 specifically includes:

[0022] The frequency domain features are used as nodes to construct a graph structure;

[0023] The graph structure is represented and learned by using the self-attention mechanism in a graph attention network.

[0024] Sequence reconstruction is performed based on the learned representation, and the prediction result is output.

[0025] Preferably, S4 further includes: obtaining the reconstruction error by calculating the node-by-node difference between the output prediction result of the graph attention network and the input frequency domain features.

[0026] Preferably, S5 specifically includes: using a dynamic time warping algorithm to calculate the sequence difference between the prediction result and the frequency domain features, and using the calculated difference as an indicator to evaluate the degradation of the roller chain.

[0027] The present invention also provides a sensorless roller chain condition monitoring device, the device being used to implement the above method, comprising: a monitoring module, a processing module, an analysis module, a reconstruction module, and an evaluation module;

[0028] The monitoring module is used to collect system status monitoring data from the motor driver of the drive system;

[0029] The processing module is used to slice the original torque data of the driver according to the motor position signal in the system status monitoring data to obtain several torque data segments.

[0030] The analysis module is used to perform cyclic spectrum analysis on the torque data segment to obtain frequency domain characteristics that characterize the periodic operating state of the system.

[0031] The reconstruction module is used to input the frequency domain features into the graph attention network for representation learning and sequence reconstruction to obtain the prediction result;

[0032] The evaluation module is used to assess the health status of the roller chain based on the difference between the prediction result and the frequency domain characteristics.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] This invention effectively overcomes the problems of difficult sensor installation, high cost, and significant noise interference in traditional roller chain condition monitoring through an innovative sensorless monitoring scheme. Utilizing the existing torque and position signals within the motor driver, no additional sensing devices are required, significantly reducing system complexity and maintenance costs. Frequency domain features extracted through cyclic spectrum analysis accurately capture the periodic degradation information during roller chain operation. Combined with the representation learning and sequence reconstruction capabilities of graph attention networks, the model's ability to capture global and local dependencies is enhanced, improving its adaptability and robustness to different operating conditions. Attached Figure Description

[0035] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0037] Figure 2 This is a flowchart illustrating the method according to an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram illustrating the similarity between two time series X and Y in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the roller chain wear test platform in an embodiment of the present invention; Figure 4 In the diagram, (a) represents the designed experimental setup, (b) represents the new chain, (c) represents the rusted chain, (d) represents the pitting defect, and (e) represents the flaking defect.

[0040] Figure 5 This is a schematic diagram of a health assessment in an early stage of an embodiment of the present invention;

[0041] Figure 6 This is a comparison chart of various degradation assessments (HI) for condition 1 in this embodiment of the invention. Figure 6 In this context, (a) represents kurtosis (K), (b) represents root mean square (RMS), (c) represents Reni entropy (RE), and (d) represents the fusion index (F). P );

[0042] Figure 7 This is a graph showing the HI curves obtained by each model after predicting degradation in working condition 1 of this embodiment of the invention. Figure 7 In the model, (a) represents the FNN-D model, (b) represents the CNN-D model, (c) represents the BiLSTM-D model, (d) represents the LSTM-D model, (e) represents the GRU-D model, (f) represents the TFM-D model, and (g) represents the MCGAN-D model.

[0043] Figure 8 This is a comparison chart showing the predicted degradation degree of different models under three working conditions in an embodiment of the present invention; Figure 8 In the diagram, (a) represents a degradation assessment comparison of chain 1 under condition 1, (b) represents a degradation assessment comparison of chain 5 under condition 2, and (c) represents a degradation assessment comparison of chain 7 under condition 31. Detailed Implementation

[0044] 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.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Example 1

[0047] This embodiment provides a sensorless roller chain condition monitoring method, the steps of which include:

[0048] S1. Collect system status monitoring data from the motor driver of the drive system.

[0049] System condition monitoring data is collected from within the motor driver of the drive system to assess the health of the roller chain. Key parameters, including sampling frequency and measurement interval, are defined during data acquisition.

[0050] S2. Based on the motor position signal in the system status monitoring data, slice the original torque data of the driver to obtain several torque data segments.

[0051] The raw torque data of the driver is sliced ​​using the motor position signal obtained in S1. This torque data reflects the changes in the health status of the roller chain mounted on the sprocket and driven by the electric motor. Based on the physical characteristics of the mechanical system, the torque data is divided into multiple logically independent data segments to isolate the relevant information for each chain rotation.

[0052] Furthermore, in S2, since roller chains are used in long-distance applications, the sprocket meshing with the chain completes N rotations for each full work cycle. The instantaneous torque fluctuations caused by the polygonal effect make direct processing of the raw data challenging. Therefore, a data segmentation strategy is proposed, treating each segment as a virtual sensor to capture torque changes at specific rotations. The original time series (denoted as...) is then processed... Divide into sub-signals, and represent as a sequence. Where the superscript t represents the signal x in the time domain, and the subscript k represents the index of the segment, as shown below:

[0053]

[0054] This segmentation method significantly reduces the periodic drift caused by polygon effects, resulting in higher periodic consistency and signal comparability in subsequent CSA feature extraction.

[0055] S3. Perform cyclic spectrum analysis on the torque data segment to obtain the frequency domain characteristics that characterize the periodic operating state of the system.

[0056] To capture the intrinsic characteristics of the roller chain, cyclic spectrum analysis (CSA) was applied to each torque data obtained in S2, transforming it from the time domain to the frequency domain to extract cyclic stationary features characterizing the periodic operating state of the system.

[0057] Furthermore, in S3, to enhance graph representation learning, CSA (Constant Autocorrelation Function) is used to extract features of cyclic stationarity. The two-dimensional Fourier transform is defined as:

[0058]

[0059] Where W is the duration of time. Yes, it's a Fourier transform. The cyclic frequency α is related to the fault frequency, and the spectral frequency f is related to the carrier frequency. CSA uses α as the frequency domain representation variable and integrates the amplitude along the f direction to obtain the enhanced envelope spectrum. The transformation process is as follows:

[0060]

[0061] This method does not require preprocessing of the original signal and can directly extract rotational features from the motor drive torque signal, effectively reflecting the cyclic feature changes caused by roller chain meshing impact, friction, etc.

[0062] S4. Input the frequency domain features into the graph attention network for representation learning and sequence reconstruction to obtain the prediction results.

[0063] The transformed frequency domain data is input into a multivariate graph attention network for representation learning and sequence reconstruction. This network consists of several key components: data embedding, graph construction, self-attention mechanism, and reconstruction error.

[0064] Furthermore, in S4, N nodes are... The graph structure is constructed, and representation learning and sequence reconstruction are performed through a graph attention network (GAT). A graph attention network (GAT) is a specialized type of graph neural network (GNN) that utilizes graph structure and node features to perform context-aware reasoning through a self-attention mechanism. By dynamically assigning importance weights to neighboring nodes, GAT can effectively capture the intrinsic features of graph structure data.

[0065] Graph attention networks are mainly composed of the following key components:

[0066] Data embedding: The N frequency domain feature nodes output from step S3 are used as multivariate inputs and simultaneously input into the graph neural network model MCGAN, which can process the data of each chain rotation differently.

[0067] Graph structure construction: Each feature node Defined as a graph One of the vertices, where Given a set of nodes, establish connecting edges between the nodes. This represents the potential dependencies between different nodes. The structural information of the graph is encoded by the adjacency matrix A, where each element... Represents node o i and o j Does an edge exist between them?

[0068] Graph Self-Attention Prediction: Self-attention mechanisms assign different weights to different input elements to indicate their relative importance. The core idea is to learn the correlations between different inputs to achieve efficient information aggregation and extraction. It typically involves three vectors: query (Q), key (K), and value (V). Given an input sequence... , where x i Let represent a d-dimensional vector containing the i-th element. The input data is first projected onto three matrices Q, K, and V. The calculation process for these projections is as follows:

[0069]

[0070] in , and This represents each element x applied to the input X. i The three linear mappings are used. Attention weights are calculated by applying the dot product of the Q and K matrices, followed by a scaling factor and a softmax function. This scaling factor balances the gradient and numerical range, ensuring stable training. The final score represents the attention weights, determining how much attention each element receives, as shown below:

[0071]

[0072] Where P represents the output attention weight, d k This is the dimension of the input vector. Divide by Scaling the dot product operation helps accelerate the learning process. This normalization step alleviates the gradient vanishing problem during backpropagation. Finally, multiplying the input vector by its corresponding attention weights allows the model to focus on the most relevant parts of the input sequence. This process enhances the model's ability to focus on the most informative information and its robustness to noise in the data.

[0073] Reconstruction Error: To achieve a health assessment of the roller chain's operating state, the degree of system degradation is quantified by calculating the reconstruction error between the model input and output. The prediction results output by the graph attention network are then used. With corresponding input features By performing node-by-node comparisons, the difference at each snapshot represents the reconstruction error. By tracking the reconstruction error over time, intelligent health indicators that change over time can be constructed, enabling monitoring of the health status and assessment of degradation trends throughout the entire lifecycle of the roller chain system.

[0074] S5. Assess the health status of the roller chain based on the difference between the prediction results and the frequency domain characteristics.

[0075] The dynamic time warping algorithm is used to quantitatively calculate the difference or distance between the output sequence predicted by the graph network in S4 and the original input sequence in S3, serving as the basis for evaluating degradation. Detailed methods are as follows... Figure 2 As shown.

[0076] Furthermore, a method using Dynamic Time Warping (DTW) is employed to calculate the similarity between two time series. For example... Figure 3 As shown, the sequence and These are two waveforms with similar overall shapes but different lengths and misaligned on the time axis, X being a sample x. k Point y in Y k+1 DTW aligns two sequences by finding an optimal curved path W and calculating the minimum cumulative distance along that path. To achieve this alignment, DTW constructs a... A grid matrix, where each element Representing point x i and y i The distance between X and Y. A smaller distance indicates greater similarity. Optimal alignment is determined by identifying the path that traverses this matrix, minimizing the cumulative distance between X and Y. Its calculation expression is:

[0077]

[0078] Among them, w k It is the k-th element on the curved path W. Through calculation, the time series is well aligned, thus achieving the minimum cumulative distance.

[0079] The method flow of this embodiment is as follows: Figure 1 As shown.

[0080] Example 2

[0081] To verify the effectiveness of the method of this invention, this embodiment designed a specialized roller chain wear test platform to observe the wear and degradation process of the roller chain, such as... Figure 4 As shown in (a). The platform includes two SEW-Eurodrive induction motors and drivers, two sprockets and four bearings, and can test two roller chains simultaneously. Figure 4 (b)-(e) depict the new chain, the chain that rusted after 300 hours of testing, the chain with pitting defects, and the chain with peeling defects, respectively. A total of nine identical roller chains were tested under three different operating conditions, as detailed in Table 1.

[0082] Table 1

[0083]

[0084] The torque and position data of the drive are recorded at a frequency of 2 kHz. The sprocket has 20 teeth that mesh with 100 chain links, and the total chain length is 94.2975 mm. Therefore, a complete rotation of the chain corresponds to 5 rotations of the sprocket or motor shaft (i.e., N=5).

[0085] To comprehensively evaluate the performance of the MCGAN-D method, this paper designed two experiments: a comparative experiment and an ablation experiment.

[0086] (1) Comparative experiment

[0087] To demonstrate the superior performance of this invention, this embodiment uses a variety of deep learning models combined with DTW as comparison models, including feedforward neural networks (FNN), convolutional neural networks (CNN), long short-term memory networks (LSTM), bidirectional LSTM (BiLSTM), gated recurrent units (GRU), and Transformer (TFM).

[0088] To evaluate the Health Indicator (HI) extracted by the model, this paper uses a trend indicator (Trd) to measure the correlation between feature y and time index τ throughout the degradation process, as shown in the following formula:

[0089] .

[0090] like Figure 5As shown, health assessment starting from the early stages is crucial for ensuring the reliable operation of rotating machinery and preventing unexpected downtime in the later stages of mechanical failure. Based on this, this embodiment proposes a weighted performance assessment method to enable HI (Health Assessment) to better capture early trends. It divides the entire lifecycle of the machinery into an initial stage and a severe failure stage, used to distinguish between the healthy state and the early stage of failure initiation (Trd1) and the late stage of operation (Trd2). When calculating the final score, a higher weight (0.8) is assigned to Trd1, while a lower weight (0.2) is assigned to Trd2. The calculation formula is as follows:

[0091] .

[0092] In the comparative experiment, this embodiment conducted qualitative and quantitative analyses of HI. The qualitative analysis examined the three operating conditions listed in Table 1, using several statistical indicators, including mean, standard deviation, variance, kurtosis (K), skewness (Skw), root mean square (RMS), maximum value (Max), minimum value (Min), peak-to-peak (P2P), Shannon entropy (SE), Raney entropy (RE), and circuit breaker index (FP). Figure 6 The chart comparing the various indicators for operating condition 1 shows that these indicators exhibit significant fluctuations, resulting in a low trend. Figure 7 The chart shows the HI curves obtained from the degradation predictions of various models in Condition 1. It can be seen that the prediction results are relatively consistent, with a relatively stable trend in the healthy stage and an upward trend in the later stages of the degradation process. GRU-D and MCGAN-D performed better, especially MCGAN-D, which was more accurate in predicting chain 3.

[0093] To more clearly compare the differences between the various models, HI was normalized based on the maximum amplitude. Figure 8This chart compares the degradation predictions of different models under three scenarios. The vertical axis represents the normalized HI of each model, and the horizontal axis represents the number of data snapshots. In scenario 1, most models show similar patterns before 1200 snapshots, while MCGAN-D stands out, showing a slight upward trend from the beginning and a sharp increase around 1200 snapshots. In scenario 2, compared to other models, MCGAN-D shows a clearer trend before 1400 snapshots, especially showing a more significant increase after 800 snapshots. This upward trend already exists at 600 snapshots, much earlier than the observations of other models around 1000 snapshots. In scenario 3, most models, except for GRU-D and MCGAN-D, predict similar results. The curves of GRU-D and MCGAN-D almost overlap, but between 200 and 1180 snapshots, MCGAN-D shows a clearer trend than GRU-D. The above phenomena show that MCGAN-D performs best among all deep learning models, not only providing the most effective health predictions in the early stages, but also maintaining a stable prediction trend throughout the degradation process.

[0094] In quantitative analysis, weighted performance evaluation methods are used to evaluate both statistical indicators and model performance.

[0095] Table 2

[0096] .

[0097] Table 2 presents a quantitative comparison of health assessment using statistical indicators, selecting 11 classic statistical indicators as benchmarks. It can be seen that traditional indicators generally score low, with both positive and negative results, making it difficult to accurately reflect the degenerative process. In contrast, the health indicators constructed by MCGAN-D maintain a positive trend across all chains, with an average score of 0.728, significantly outperforming the best statistical indicator among traditional indicators, peak-to-peak value.

[0098] Table 3

[0099] .

[0100] Table 3 presents a quantitative comparison of health assessments using deep learning models, selecting six representative deep learning models combined with DTW for comparison. Experimental results show that MCGAN-D achieved the best performance on seven out of nine test chains, with an average score of 0.728, demonstrating excellent early degradation trend tracking capabilities.

[0101] (2) Ablation test

[0102] To verify the crucial role of DTW in constructing health indicators, this paper designed an ablation experiment, replacing DTW with root mean square error (RMSE) and bias score (DS), respectively. These methods quantitatively evaluate the performance of the model in reconstructing data by quantifying the difference between the real and predicted sequences. The formulas for RMSE and DS are as follows:

[0103]

[0104] Among them, y i It's a genuine RUL label. is the predicted RUL, and n is the length of the predicted data.

[0105]

[0106] in, and This represents the observed and predicted values ​​at time step t, where i represents the index of the predicted sequence. and These are the median and interquartile range, respectively.

[0107] The experimental results are shown in Table 3. The results show that the average score of MCGAN-D using DTW is 0.728, while the scores of MCGAN-R using RMSE and MCGAN-DS using DS are 0.615 and 0.635, respectively, confirming the superiority of DTW in dealing with time series distortion and noise.

[0108] In summary, the MCGAN-D method proposed in this invention demonstrates significant advantages in roller chain condition monitoring. The constructed health index features strong early warning capabilities, smooth and monotonic trends, and strong adaptability to different working conditions, providing an effective solution for sensorless condition monitoring of mechanical systems.

[0109] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A sensorless roller chain condition monitoring method, characterized in that, Includes the following steps: S1. Collect system status monitoring data from the motor driver of the drive system; S2. Based on the motor position signal in the system status monitoring data, slice the original torque data of the driver to obtain several torque data segments; S3. Perform cyclic spectrum analysis on the torque data segment to obtain the frequency domain characteristics characterizing the periodic operating state of the system; S4. Input the frequency domain features into the graph attention network for representation learning and sequence reconstruction to obtain the prediction result; S5. Evaluate the health status of the roller chain based on the difference between the prediction result and the frequency domain characteristics.

2. The sensorless roller chain condition monitoring method according to claim 1, characterized in that, In step S1, when collecting system status monitoring data, key parameters are also defined, including sampling frequency and measurement interval.

3. The sensorless roller chain condition monitoring method according to claim 1, characterized in that, S2 specifically includes: dividing the original torque data into multiple logically independent data segments according to the chain rotation cycle based on the motor position signal, so as to isolate the torque information within each chain rotation cycle.

4. The sensorless roller chain condition monitoring method according to claim 1, characterized in that, S3 specifically includes: Cyclic spectrum analysis was applied to each torque data segment to transform it from the time domain to the frequency domain and extract cyclic stationary features. The cyclic spectrum analysis is a two-dimensional Fourier transform of the instantaneous autocorrelation function.

5. The sensorless roller chain condition monitoring method according to claim 1, characterized in that, S4 specifically includes: The frequency domain features are used as nodes to construct a graph structure; The graph structure is represented and learned by using the self-attention mechanism in a graph attention network. Sequence reconstruction is performed based on the learned representation, and the prediction result is output.

6. The sensorless roller chain condition monitoring method according to claim 5, characterized in that, S4 further includes: obtaining the reconstruction error by calculating the node-by-node difference between the output prediction result of the graph attention network and the input frequency domain features.

7. The sensorless roller chain condition monitoring method according to claim 6, characterized in that, S5 specifically includes: using a dynamic time warping algorithm to calculate the sequence difference between the prediction result and the frequency domain features, and using the calculated difference as an indicator to evaluate the degradation of the roller chain.

8. A sensorless roller chain condition monitoring device, the system being used to implement the method according to any one of claims 1-7, characterized in that, include: Monitoring module, processing module, analysis module, reconstruction module, and evaluation module; The monitoring module is used to collect system status monitoring data from the motor driver of the drive system; The processing module is used to slice the original torque data of the driver according to the motor position signal in the system status monitoring data to obtain several torque data segments. The analysis module is used to perform cyclic spectrum analysis on the torque data segment to obtain frequency domain characteristics that characterize the periodic operating state of the system. The reconstruction module is used to input the frequency domain features into the graph attention network for representation learning and sequence reconstruction to obtain the prediction result; The evaluation module is used to assess the health status of the roller chain based on the difference between the prediction result and the frequency domain characteristics.

Images