Online operation state evaluation method for main driving chain of escalator

Abstract

The online operation state evaluation method for the main driving chain of the escalator comprises the steps that an image collection system is built, and state images of the main driving chain are continuously collected in the historical operation process of the escalator; the collected state image is processed, and a transverse vibration signal generated in the running process of the driving chain is extracted; performing sample labeling, signal denoising, signal fragmentation and data standardization processing on the collected original transverse vibration signals to obtain a real historical data sample set for model training; a real historical data sample set is used as training data, and a data enhancement model based on a time sequence generation framework is constructed and trained; inputting a random noise vector into the trained data enhancement model, generating different types of abnormal samples, and forming a health state evaluation training set by real historical data samples and the abnormal samples; and establishing and training a health state assessment model based on the health state assessment training set.

Description

Technical Field

[0001] This invention belongs to the field of special equipment operation status assessment technology, and relates to an online operation status assessment method for the main drive chain of an escalator. Background Technology

[0002] Special equipment is defined as equipment involving life safety and posing significant risks. Escalators, as a type of elevator, fall under the category of special equipment and are widely used in densely populated public places. In the power transmission process of escalators, the main drive chain, as the core transmission component, not only undertakes the crucial functions of power transmission and step circulation, but is also a vital foundation of the entire safety protection system. Its driving stability and braking reliability directly determine the overall performance and safety level of the escalator. However, during long-term operation, the main drive chain gradually degrades in performance due to factors such as mechanical wear and fatigue loads, and may even experience risks of chain elongation and breakage. Therefore, condition monitoring and scientific evaluation of the escalator's main drive chain are of great significance for ensuring safe equipment operation and preventing accidents.

[0003] Current maintenance strategies for escalator main drive chains typically involve planned, periodic maintenance. This method relies heavily on regular inspections and maintenance, and the evaluation results depend heavily on the analytical skills and experience of the personnel, lacking effective utilization of real-time equipment operating data. Furthermore, this intermittent maintenance easily leads to over-maintenance or under-maintenance, resulting in either wasted resources or overlooked potential safety hazards. In recent years, with the continuous development of deep learning algorithms, data-driven equipment condition assessment technologies have gained increasing attention. These methods collect operating signals through sensors installed on the equipment and use algorithms to intelligently assess the equipment's condition, fully utilizing the representative data of the equipment's condition and effectively improving the real-time performance and objectivity of condition monitoring.

[0004] Effectively collecting status characterization data for the main drive chain of escalators remains challenging due to its operational characteristics. Directly mounting sensors on the chain may interfere with its normal operation or even accelerate wear; while fixing sensors to static components such as couplings makes the acquired data susceptible to interference from other components, hindering accurate reflection of the drive chain's true state. Furthermore, efficiently extracting and utilizing the chain's characteristic parameters during operation is a key challenge for achieving accurate assessment.

[0005] The technical solution proposed in Chinese patent application "CN202310237002.1 Safety Detection Method for Escalator Drive Chain Based on Machine Vision" focuses on using a purely visual method for detection. This solution aims to optimize the quality of acquired images through image preprocessing and other techniques. In the anomaly diagnosis stage, it directly matches real-time images with an initial template of the healthy chain to determine the status. However, this diagnostic method is relatively simple: on the one hand, noise is easily introduced during image stitching, and the diagnostic results are easily affected by equipment performance and environmental changes, resulting in insufficient robustness; on the other hand, this solution still does not overcome the limitation of traditional maintenance models requiring downtime for inspection, failing to achieve online real-time monitoring. Therefore, compared to manual periodic maintenance, its main advantage lies only in improving the objectivity of detection, without significant breakthroughs in diagnostic intelligence and operational efficiency.

[0006] The technical solution proposed in Chinese patent "CN202210562832.7 A Fault Diagnosis Method for Escalator Drive Chain Based on Vibration Feature Enhancement" focuses on achieving fault diagnosis through a combination of signal processing and deep learning. This solution employs a data-driven approach, heavily relying on vibration signals generated during escalator drive chain operation as input. However, its signal acquisition method has limitations: it monitors vibration using only a single accelerometer mounted on one side of the sprocket, essentially providing an indirect measurement of the drive chain's vibration behavior. Due to the complex structure of escalator systems and the simultaneous operation of multiple mechanical components, the acquired vibration signals inevitably contain multi-source vibration interference from the sprocket, drive shaft, and step chain, introducing significant noise at the signal acquisition stage. This interference from signal sources significantly restricts the overall diagnostic performance. Summary of the Invention

[0007] To address the aforementioned technical problems, the purpose of this invention is to provide an online operational status assessment method for the main drive chain of an escalator.

[0008] The present invention provides a method for online operation status evaluation of the main drive chain of an escalator, comprising:

[0009] Step 1: Set up an image acquisition system to continuously acquire status images of the main drive chain during the escalator's historical operation;

[0010] Step 2: Process the acquired state images to extract the lateral vibration signals generated during the operation of the drive chain;

[0011] Step 3: Perform sample labeling, signal denoising, signal segmentation, and data standardization on the collected raw lateral vibration signals to obtain a real historical data sample set for model training;

[0012] Step 4: Using real historical data sample sets as training data, construct and train a data augmentation model based on a time series generation framework;

[0013] Step 5: Input random noise vectors into the trained data augmentation model to generate abnormal samples of different categories. The health status assessment training set is composed of real historical data samples and abnormal samples.

[0014] Step 6: Based on the health status assessment training set, establish and train the health status assessment model.

[0015] This invention provides an online operational status assessment method for the main drive chain of escalators, enabling dynamic acquisition of vibration signals from the escalator's main drive chain during operation. A data-driven approach is used to fully analyze the spatiotemporal characteristics of the acquired signals, and a data augmentation model is introduced to effectively address the problem of uneven data distribution under real-world operating conditions, improving the model's generalization ability and convergence speed. This invention effectively overcomes the technical bottlenecks of traditional downtime maintenance strategies, which are unable to effectively cope with emergencies and heavily rely on the subjective experience of personnel. Attached Figure Description

[0016] Figure 1 This is a flowchart of an online operation status evaluation method for the main drive chain of an escalator according to the present invention. Detailed Implementation

[0017] like Figure 1 As shown, the present invention provides an online operational status assessment method for the main drive chain of an escalator, comprising:

[0018] Step 1: Set up an image acquisition system to continuously acquire status images of the main drive chain during the escalator's historical operation.

[0019] The image acquisition system in step 1 includes a miniature CMOS industrial camera, a light source, and a light-colored background plate. The miniature CMOS industrial camera is equipped with a short focal length, distortion-free wide-angle lens; it captures images from the side of the tensioned edge of the main drive chain. This configuration ensures a wider field of view within the limited installation space, ensuring that multiple chain plates are captured in a single frame image, providing sufficient information for subsequent extraction of the chain's spatial features. Since escalator drive chains typically operate in dim or unlit environments, a dedicated light source is required. The light source illuminates perpendicularly to the chain plate surface to reduce shadow interference. To further enhance the distinction between the target and background in the image, the light-colored background plate is placed behind the chain plate, ensuring that the photographed area is positioned between the miniature industrial camera and the light-colored background plate, creating a high-contrast imaging effect that makes the chain link edges clear, facilitating subsequent vibration signal extraction.

[0020] During installation, the camera's tilt is adjusted to ensure the image acquisition area is parallel to the overall trend of the chain tension side. This ensures that the vibration signal components extracted from the acquired sequence images are primarily lateral vibrations, meeting the requirements for subsequent analysis.

[0021] Step 2: Process the acquired state images to extract the lateral vibration signals generated during the operation of the drive train, specifically:

[0022] Step 2.1: Use the Canny algorithm to extract edges from the chain image. The specific process includes: image grayscale conversion, Gaussian filtering, gradient calculation, non-maximum suppression, double threshold detection, and edge tracking.

[0023] Step 2.1.1: Convert the RGB values ​​of the chain image to a grayscale value using a weighted average method. The specific formula is as follows:

[0024]

[0025] in: This represents the calculated grayscale value. , , These represent the pixel values ​​for the red, green, and blue channels, respectively.

[0026] Step 2.1.2: Smooth the image using a Gaussian filter to reduce the impact of noise on edge detection. The filtering formula is as follows:

[0027]

[0028] in, This refers to the coordinate offset of the current node relative to the center point in the Gaussian filter kernel; Point The weight values ​​of the Gaussian filter kernel; The standard deviation represents the Gaussian distribution and is used to control the smoothness of the filter.

[0029] Step 2.1.3: Calculate the gradient of the image in the x and y directions using the Sobel operator, and then calculate the magnitude and direction of the gradient. The Sobel operator formula is as follows:

[0030]

[0031] in, and These represent the gradient values ​​in the horizontal and vertical directions, respectively. This represents the convolution operator. This indicates the grayscale image to be processed.

[0032] The formulas for the magnitude and direction of the gradient are as follows:

[0033]

[0034] in, Point gradient magnitude at that point Point The gradient direction at that point.

[0035] Step 2.1.4: If the gradient magnitude of the current pixel is not at its maximum value in the gradient direction, suppress it. The specific steps are as follows:

[0036] First, the gradient direction is approximated as 0°, 45°, 90°, and 135°. Then, the gradient magnitude of the current pixel is compared with that of the adjacent pixels along the gradient direction. Only the pixel with the largest gradient magnitude is retained, and the rest are set to 0.

[0037] Step 2.1.5: Set two thresholds, a high threshold and a low threshold, and classify pixels into strong edges, weak edges, and non-edges based on the gradient magnitude:

[0038]

[0039] in, Point What type is the location determined to be? Indicates the point after nonmaximum suppression processing Gradient magnitude at; Indicates a high threshold; Indicates a low threshold. These represent strong edges, weak edges, and non-edges, respectively.

[0040] Step 2.1.6: Process weak edges to ensure edge continuity, specifically as follows:

[0041] (1) If a weak edge is connected to a strong edge, it is considered a real edge;

[0042] (2) If a weak edge is not connected to a strong edge, it is considered noise and suppressed.

[0043] (3) By examining weak edge pixels and their neighboring pixels, if at least one of them is a strong edge pixel, the weak edge point is saved; otherwise, it is removed.

[0044] Step 2.2: Based on the edge extraction results, extract the upper and lower boundaries of the chain by scanning along the horizontal axis: perform a bidirectional traversal along each column of pixels, and define the first edge point encountered as the upper and lower boundaries, thereby constructing a complete boundary contour. The midpoint is then obtained based on the upper and lower boundaries, which is the chain shape fitting curve, specifically:

[0045] Step 2.2.1: Based on the binary image generated after edge extraction, contour extraction is performed using the horizontal scanning method, specifically as follows:

[0046] For each column of the image: the upper boundary point is the coordinate of the first non-zero pixel encountered while scanning down the column, which is then taken as the upper boundary point of the chain. The lower boundary point is determined by scanning the column from bottom to top, and the coordinates of the first non-zero pixel encountered are taken as the lower boundary point of the chain. After traversing all columns, the upper and lower boundaries of the chain are obtained.

[0047] Step 2.2.2: Given that the chain plate structure is approximately symmetrical in the vertical direction, its morphological curve is characterized by calculating the geometric central axes of the upper and lower boundaries. Specifically, for each column... The ordinate of the axis point Calculated by the following formula:

[0048]

[0049] in: For a specific column The vertical coordinate of the chain shape curve obtained from the above calculation.

[0050] Step 2.3: Set equally spaced vertical measurement lines on each image. Construct a time-domain signal characterizing the chain's dynamic vibration using the coordinate sequence of the intersection points of the vertical measurement lines and the chain morphology fitting curve. This allows for the quantitative analysis of vibration characteristics. This step aims to extract the spatial discrete points of the chain morphology curve from a single frame image and then construct the vibration signal using a multi-frame image sequence. Specifically:

[0051] Step 2.3.1: In the image coordinate system, set multiple vertical measurement lines at equal intervals along the horizontal axis; these measurement lines are evenly distributed in space and are used to discretize the chain morphology.

[0052] Step 2.3.2: For each frame of image, calculate the coordinates of the intersection point of each vertical measurement line and the chain shape curve. The ordinate value of each intersection point represents the measurement point position of the chain at that moment.

[0053] Step 2.3.3: Extract and arrange the ordinate values ​​of the same spatial measurement point in the continuous multi-frame images in chronological order. The data of each measurement point constitutes an independent displacement-time series, which is the vibration signal at that measurement point.

[0054] Step 3: Perform sample labeling, signal denoising, signal segmentation, and data standardization on the collected raw vibration signals to obtain a real historical data sample set for model training, specifically:

[0055] Step 3.1: Based on the equipment operation records and status monitoring data, label the original lateral vibration signal with the corresponding health status category, including: healthy, deteriorated, and faulty.

[0056] Step 3.2: During vibration signal acquisition, high-frequency noise components are introduced into the original signal due to environmental interference and other factors. To improve signal quality, wavelet threshold denoising is used to preprocess the original transverse vibration signal. This method effectively separates noise from the true vibration components by selecting appropriate wavelet basis functions and threshold rules.

[0057] Step 3.3: To transform the continuous vibration signal into a sample sequence suitable for model training, a fixed-length sliding window is used to segment the denoised vibration signal. This operation divides the long time-series signal into multiple short signal segments of equal length. This increases the number of training samples and also helps the model capture the local temporal dependencies of the signal.

[0058] Step 3.4: To eliminate the impact of differences in signal amplitude dimensions on model training, the signal segments are subjected to minimax normalization, linearly mapping their amplitudes to the [-1, 1] interval. This accelerates the model convergence process and improves training stability and generalization ability. A real historical data sample set is then constructed for supervised learning.

[0059] Step 4: Using real historical data samples as training data, construct and train a data augmentation model based on the time series generation framework, specifically as follows:

[0060] Step 4.1: Construct a data augmentation model based on TimeGAN and incorporating a category-aware mechanism. The model consists of two parts: a multi-branch generation unit and a classifier.

[0061] Multi-branch generation unit: Based on the three categories of "health", "deterioration" and "failure" defined in the equipment health status assessment task, three parallel and parameter-independent generation units are constructed. Each generation unit is based on the TimeGAN framework, learns the temporal dynamic characteristics and distribution patterns of historical data samples of the corresponding category, and generates time series data corresponding to a health status category ("health", "deterioration" and "failure") to achieve categorized sample generation.

[0062] Classifier: To enhance the class discriminativeness of the generated samples, a classifier is introduced. This classifier adopts the same TCN-GAT combined architecture as the state evaluation model described in step 6. Its function is to classify the generated samples and calculate the classification loss, so as to encourage the generation unit to generate samples with obvious classification features.

[0063] Step 4.2: First, the classifier is trained using a historical data sample set to enable it to have good state recognition capabilities, which will serve as the discrimination benchmark for the subsequent generation process.

[0064] Step 4.3: While performing adversarial training on each generation unit using the historical data sample set, the generated samples, along with their preset class labels, are simultaneously input into the trained classifier for discrimination. The classification loss is calculated, and this loss is used as an auxiliary supervision signal integrated into the overall loss function. This prompts the generator to strengthen class discrimination features while fitting the data distribution, thereby generating augmented samples that are more conducive to the training of the classification model. The overall loss function of the data augmentation model consists of four parts: the reconstruction loss of the TimeGAN generation unit under each category, the unsupervised adversarial loss, the supervised loss, and the classification loss calculated by the classifier. The overall loss function L is expressed as:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071] Where L is the loss of the entire generative model, and N represents the total number of N types of signals generated, i.e., the total number of generative units. The newly added classification loss is used to improve model fitting and classification performance. This represents the true label of the m-th generated sample. This represents the probability that the classifier predicts the m-th generated sample belongs to class k, and M represents the number of samples generated by each generator in a training batch. Let be the loss function of the generation unit corresponding to the k-th category, where , , These are three hyperparameters, used to balance the weights of unsupervised adversarial loss, supervised loss, and classification loss in the total loss; The reconstruction loss of the k-th generating unit is used to train the autoencoder, ensuring that it can accurately reconstruct the input data. This represents the original static features in the k-th generating unit; This represents the reconstructed static features in the k-th generating unit; This represents the original time series features of the k-th generating unit; This represents the time-series features reconstructed from the k-th generating unit; The balanced unsupervised adversarial loss for the k-th generation unit is used to train the generator and discriminator, ensuring that the distribution of generated data is consistent with that of real data. and These represent the discrimination results for static features and time series features in the k-th generation unit, respectively; This represents the discrimination result of the discriminator in the k-th generation unit on the reconstructed time series features; The supervised loss for the k-th generation unit is used to train the supervisor, ensuring that the generated data follows the temporal dynamics of the real data. For real data at time t The corresponding latent representation; This is the global latent vector obtained by encoding conditional information in the k-th generating unit; Let be the random noise vector at time step t in the k-th generation unit; This represents the prediction of the potential state at the current time t.

[0072] Step 5: Input a random noise vector into the trained data augmentation model to generate abnormal samples of different categories. The health status assessment training set consists of real historical data samples and abnormal samples, specifically:

[0073] Step 5.1: Using the trained data augmentation model, select the generation units corresponding to unhealthy states, including "deterioration" and "fault", and input random noise vectors into them respectively to generate abnormal samples of the corresponding categories.

[0074] Step 5.2: By controlling the number of samples generated for each anomaly category, the number of samples generated is precisely made up to fill the gap in the number of anomaly samples in the original real historical data sample set, thus achieving a balance between the number of normal and anomaly samples.

[0075] Step 5.3: The health status assessment training set is composed of real historical data samples and abnormal samples, providing a high-quality data foundation for the subsequent full learning of the health status assessment model.

[0076] Step 6: Based on the health status assessment training set, establish and train the health status assessment model.

[0077] The model is a multi-layer network structure that integrates temporal and spatial features, including: a temporal feature extraction layer, a graph structure construction layer, and a spatial feature extraction layer.

[0078] Temporal Feature Extraction Layer: This layer uses a temporal convolutional neural network (TCN) to extract temporal features from the high-dimensional vibration signals in the input health status assessment training set. The TCN can effectively capture the multi-scale temporal features of the input signal and output a high-dimensional feature sequence with the same dimension as the input high-dimensional vibration signal.

[0079] Graph Structure Construction Layer: The high-dimensional feature sequence output by the temporal feature extraction layer is converted into graph structure data according to time steps. In this graph, the high-dimensional feature sequence at each time step is defined as a node, and fully connected edges are established between all nodes to construct a fully connected topology graph. This structure spatializes the temporal features, laying the foundation for subsequent network layers to extract the spatial dependencies between nodes.

[0080] Spatial Feature Extraction Layer: This layer is built on the Graph Attention Network (GAT). The fully connected topology graph generated in the previous stage is used as input. The attention mechanism is used to adaptively learn the correlation strength between nodes in the graph, i.e., the temporal features of each measurement point. By weighted aggregation of neighbor node information, GAT can effectively capture the implicit spatial dependencies between measurement points, thereby extracting spatiotemporal comprehensive features that integrate temporal dynamics and topological associations.

[0081] The health status assessment model uses the cross-entropy loss function. After training, it is integrated and deployed on a host computer hardware and software platform. This system can interface with a real-time image acquisition unit to perform online sensing and assessment of the operating status of the escalator's main drive chain, ultimately achieving automated and intelligent operation and maintenance from data acquisition and feature analysis to status recognition.

[0082] Key points of this invention:

[0083] 1. Due to the structural and operational characteristics of roller chains differing from conventional transmission equipment, it is difficult to install traditional contact sensors. Forcibly installing contact sensors may interfere with the normal operating characteristics of the drive chain, leading to unforeseen operational risks. Furthermore, fixing sensors in other locations for indirect measurement introduces significant vibration noise from other components, reducing the specificity and accuracy of the condition assessment. Therefore, this solution employs a non-contact acquisition method to effectively extract vibration signals.

[0084] 2. In real-world operating environments, equipment malfunctions are typically repaired or replaced promptly, resulting in a significantly smaller number of signal samples under abnormal conditions compared to those under normal conditions, leading to an imbalanced data distribution. To address this, the present invention introduces a generative model to synthesize high-quality abnormal state data, thus compensating for the sample gap.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the ideas of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for online operation status evaluation of the main drive chain of an escalator, characterized in that, include: Step 1: Set up an image acquisition system to continuously acquire status images of the main drive chain during the escalator's historical operation; Step 2: Process the acquired state images to extract the lateral vibration signals generated during the operation of the drive chain; Step 3: Perform sample labeling, signal denoising, signal segmentation, and data standardization on the collected raw lateral vibration signals to obtain a real historical data sample set for model training; Step 4: Using real historical data sample sets as training data, construct and train a data augmentation model based on a time series generation framework; Step 5: Input random noise vectors into the trained data augmentation model to generate abnormal samples of different categories. The health status assessment training set is composed of real historical data samples and abnormal samples. Step 6: Based on the health status assessment training set, establish and train the health status assessment model.

2. The online operation status evaluation method for the main drive chain of an escalator according to claim 1, characterized in that, The image acquisition system in step 1 includes: a miniature CMOS industrial camera, a light source, and a light-colored background plate; The micro CMOS industrial camera is equipped with a short focal length, distortion-free wide-angle lens; the micro CMOS industrial camera takes pictures from the side of the tensioning edge of the main drive chain; The light source is positioned to illuminate the chain plate surface perpendicularly, thereby reducing shadow interference. The light-colored background plate is placed behind the chain plate, ensuring that the area being photographed is positioned between the miniature industrial camera and the light-colored background plate, forming a high-contrast imaging effect, making the edges of the chain links clear, and facilitating the subsequent extraction of vibration signals.

3. The online operation status evaluation method for the main drive chain of an escalator according to claim 1, characterized in that, Step 2 specifically involves: Step 2.1: Use the Canny algorithm to extract edges from the chain image. The specific process includes: image grayscale conversion, Gaussian filtering, gradient calculation, non-maximum suppression, double threshold detection, and edge tracking. Step 2.2: Based on the edge extraction results, the upper and lower boundaries of the chain are extracted by scanning along the horizontal axis. The midpoint is obtained based on the upper and lower boundaries, which is the chain shape fitting curve. Step 2.3: Set equally spaced vertical measurement lines on each graph. Construct a time-domain signal characterizing the dynamic vibration of the chain by using the coordinate sequence of the intersection points of the vertical measurement lines and the fitting curve of the chain morphology, thereby realizing the quantitative analysis of vibration characteristics.

4. The online operation status evaluation method for the main drive chain of an escalator according to claim 3, characterized in that, Step 2.1 specifically involves: Step 2.1.1: Convert the RGB values ​​of the chain image to a grayscale value using a weighted average method. The specific formula is as follows: in: This represents the calculated grayscale value. , , These represent the pixel values ​​for the red, green, and blue channels, respectively. Step 2.1.2: Smooth the image using a Gaussian filter to reduce the impact of noise on edge detection. The filtering formula is as follows: in, This refers to the coordinate offset of the current node relative to the center point in the Gaussian filter kernel; Point The weight values ​​of the Gaussian filter kernel; The standard deviation represents the Gaussian distribution and is used to control the smoothness of the filter. Step 2.1.3: Calculate the gradient of the image in the x and y directions using the Sobel operator, and then calculate the magnitude and direction of the gradient. The Sobel operator formula is as follows: in, and These represent the gradient values ​​in the horizontal and vertical directions, respectively. This represents the convolution operator. This represents the grayscale image to be processed; The formulas for the magnitude and direction of the gradient are as follows: in, Point gradient magnitude at that point Point The gradient direction at that location; Step 2.1.4: If the gradient magnitude of the current pixel is not at its maximum value in the gradient direction, suppress it. The specific steps are as follows: First, the gradient direction is approximated as 0°, 45°, 90°, and 135°. Then, the gradient magnitude of the current pixel is compared with that of its neighboring pixels along the gradient direction. Only the pixel with the largest gradient magnitude is retained, and the rest are set to 0. Step 2.1.5: Set two thresholds, a high threshold and a low threshold, and classify pixels into strong edges, weak edges, and non-edges based on the gradient magnitude: in, Point What type is the location determined to be? Indicates the point after nonmaximum suppression processing Gradient magnitude at; Indicates a high threshold; Indicates a low threshold. These represent strong edges, weak edges, and non-edges, respectively. Step 2.1.6: Process weak edges to ensure edge continuity, specifically as follows: (1) If a weak edge is connected to a strong edge, it is considered a real edge; (2) If a weak edge is not connected to a strong edge, it is considered noise and suppressed. (3) By examining weak edge pixels and their neighboring pixels, if at least one of them is a strong edge pixel, the weak edge point is saved; otherwise, it is removed.

5. The online operation status evaluation method for the main drive chain of an escalator according to claim 3, characterized in that, Step 2.2 specifically involves: Step 2.2.1: Based on the binary image generated after edge extraction, contour extraction is performed using the horizontal scanning method, specifically as follows: For each column of the image: the upper boundary point is the coordinate of the first non-zero pixel encountered while scanning down the column, which is then taken as the upper boundary point of the chain. The lower boundary point is determined by scanning the column from bottom to top, and the coordinates of the first non-zero pixel encountered are taken as the lower boundary point of the chain. After traversing all columns, the upper and lower boundaries of the chain are obtained. Step 2.2.2: Given that the chain plate structure is approximately symmetrical in the vertical direction, its morphological curve is characterized by calculating the geometric central axes of the upper and lower boundaries. Specifically, for each column... The ordinate of the axis point Calculated by the following formula: in: For a specific column The vertical coordinate of the chain shape curve obtained from the above calculation.

6. The online operation status evaluation method for the main drive chain of an escalator according to claim 3, characterized in that, Step 2.3 specifically involves: Step 2.3.1: In the image coordinate system, set multiple vertical measurement lines at equal intervals along the horizontal axis; these measurement lines are evenly distributed in space and are used to discretize the chain morphology. Step 2.3.2: For each frame of image, calculate the coordinates of the intersection point of each vertical measurement line and the chain shape curve. The ordinate value of each intersection point represents the measurement point position of the chain at that moment. Step 2.3.3: Extract and arrange the ordinate values ​​of the same spatial measurement point in the continuous multi-frame images in chronological order. The data of each measurement point constitutes an independent displacement-time series, which is the vibration signal at that measurement point.

7. The online operation status evaluation method for the main drive chain of an escalator according to claim 1, characterized in that, Step 3 specifically involves: Step 3.1: Based on the equipment operation records and status monitoring data, label the original vibration signals with corresponding health status tags, including: healthy, deteriorated, and faulty; Step 3.2: The original transverse vibration signal is preprocessed using wavelet threshold denoising method to separate noise from the real vibration components; Step 3.3: In order to convert the continuous vibration signal into a sample sequence suitable for model training, a fixed-length sliding window is used to segment the denoised vibration signal. This operation divides the long time series signal into multiple short sequence signal segments of equal length. Step 3.4: Perform max-min normalization on the signal segment to linearly map its amplitude to the interval [-1, 1], and construct a real historical data sample set that can be used for supervised learning.

8. The online operation status evaluation method for the main drive chain of an escalator according to claim 1, characterized in that, Step 4 specifically involves: Step 4.1: Construct a data augmentation model based on TimeGAN and incorporating a category-aware mechanism. The model consists of two parts: a multi-branch generation unit and a classifier. Multi-branch generation unit: Based on the three categories of "health", "deterioration" and "failure" defined in the equipment health status assessment task, three parallel and parameter-independent generation units are constructed; each generation unit is based on the TimeGAN framework and learns the temporal dynamic features and distribution patterns of historical data samples of the corresponding category to achieve categorized sample generation; Classifier: To enhance the class discriminativeness of the generated samples, a classifier is introduced. This classifier adopts the same TCN-GAT combination architecture as the state evaluation model described in step 6. Its function is to classify the generated samples and calculate the classification loss, so as to encourage the generation unit to generate samples with obvious classification features. Step 4.2: First, the classifier is trained using a historical data sample set to enable it to have good state recognition ability, which serves as the judgment benchmark for the subsequent generation process; Step 4.3: While performing adversarial training on each generated unit using the historical data sample set, the generated samples that are sent to the discriminator for discrimination are simultaneously input into the trained classifier along with their preset class labels, and the classification loss is calculated. The overall loss function L of the data augmentation model is expressed as: Where L is the loss of the entire generative model, and N represents the total number of N types of signals generated, i.e., the total number of generative units. For the newly added classification loss, This represents the true label of the m-th generated sample. This represents the probability that the classifier predicts the m-th generated sample belongs to class k, and M represents the number of samples generated by each generator in a training batch. Let be the loss function of the generation unit corresponding to the k-th category, where , , These are three hyperparameters, used to balance the weights of unsupervised adversarial loss, supervised loss, and classification loss in the total loss; The reconstruction loss for the k-th generating unit is used to train the autoencoder. This represents the original static features in the k-th generating unit. This represents the reconstructed static features in the k-th generating unit. This represents the original time series features of the k-th generating unit; This represents the time-series features reconstructed from the k-th generating unit; The balanced unsupervised adversarial loss for the k-th generating unit is used to train the generator and discriminator. and These represent the discrimination results for static features and time series features in the k-th generation unit, respectively. This represents the discrimination result of the discriminator in the k-th generation unit on the reconstructed time series features; The supervised loss for the k-th generating unit is used to train the supervisor. For real data at time t The corresponding latent representation; This is the global latent vector obtained by encoding conditional information in the k-th generating unit; Let be the random noise vector at time step t in the k-th generation unit; This represents the prediction of the potential state at the current time t.

9. The online operation status evaluation method for the main drive chain of an escalator according to claim 1, characterized in that, Step 5 specifically involves: Step 5.1: Using the trained data augmentation model, select the generation units corresponding to unhealthy states, including "deterioration" and "fault", and input random noise vectors into them respectively to generate abnormal samples of the corresponding categories; Step 5.2: By controlling the number of samples generated for each anomaly category, the number of samples generated is precisely made up to fill the gap in the number of anomaly samples in the original real historical data sample set, so as to achieve a balance between the number of normal samples and anomaly samples. Step 5.3: The health status assessment training set is composed of real historical data samples and abnormal samples, providing a high-quality data foundation for the subsequent full learning of the health status assessment model.

10. The online operation status evaluation method for the main drive chain of an escalator according to claim 1, characterized in that, The health status assessment model in step 6 is specifically as follows: Temporal Feature Extraction Layer: This layer uses a temporal convolutional neural network (TCN) to extract temporal features from the high-dimensional vibration signals in the input health status assessment training set. The TCN can effectively capture the multi-scale temporal features of the input signal and output a high-dimensional feature sequence with the same dimension as the input high-dimensional vibration signal. Graph Structure Construction Layer: The high-dimensional feature sequence output by the temporal feature extraction layer is converted into graph structure data according to time steps. In this graph, the high-dimensional feature sequence at each time step is defined as a node, and fully connected edges are established between all nodes to construct a fully connected topology graph. Spatial Feature Extraction Layer: This layer is built on the Graph Attention Network (GAT). The fully connected topology graph generated in the previous stage is used as input. The attention mechanism is used to adaptively learn the correlation strength between nodes in the graph, i.e., the temporal features of each measurement point. By weighted aggregation of neighbor node information, GAT can effectively capture the implicit spatial dependencies between measurement points, thereby extracting spatiotemporal comprehensive features that integrate temporal dynamics and topological associations.

Images