Escalator unsupervised diagnosis method and system
By using multi-dimensional feature extraction and anomaly diagnosis models, the problems of data collection and feature dimension in escalator fault diagnosis are solved, achieving highly accurate fault identification and status monitoring, and ensuring the comprehensiveness and interpretability of the diagnostic results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING KANGNI MECHANICAL & ELECTRICAL
- Filing Date
- 2024-04-30
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies for escalator fault diagnosis struggle to balance data acquisition frequency and completeness. This leads to reduced diagnostic effectiveness when the feature system is too large, and the differences between features of different dimensions are ignored, resulting in information loss and decreased diagnostic accuracy.
A multi-dimensional feature extraction method is adopted, including extraction of complete data, sliced data and time-frequency features. Combined with the LOF model and Kmeans model, anomaly diagnosis is performed on the vibration data of escalators. Healthy and undiagnosed datasets are constructed, and diagnostic results at different levels are output.
It enables precise capture of abrupt changes and stability anomalies in escalators, ensuring the integrity of the data feature system and the accuracy of diagnostic results, avoiding the loss of feature information and interference between dimensions, and improving the interpretability of the diagnosis.
Smart Images
Figure CN118445726B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a diagnostic method, and more particularly to an unsupervised diagnostic method and system for escalators. Background Technology
[0002] With the rapid advancement of urbanization in recent years, the scale of urban public transportation has also increased dramatically. Many public transportation hubs, such as high-speed rail stations and subway stations, now widely use escalators as a means of transporting passengers. Given the large passenger volume in public transportation facilities, the safety and reliability of escalators have become a major concern for public safety. Under prolonged heavy-load operation, escalator systems are experiencing increasingly severe wear and tear. Therefore, taking effective measures to monitor the health status of public transportation escalators has become an urgent issue to be addressed.
[0003] When conducting fault diagnosis, comprehensive and effective fault data covering the entire equipment operating cycle is crucial for successful diagnosis. However, in actual data collection, for equipment with long operating cycles, maintaining a consistent data collection frequency often makes extended data acquisition impossible. Furthermore, current commonly used mechanical structure vibration diagnosis algorithms utilize feature systems spanning multiple dimensions, including time, frequency, and time-frequency domains. The typical approach is to group features from all domains into a single feature system and uniformly perform dimensionality reduction or direct training. Therefore, when the number of features is too large, the diagnostic performance of the directly trained model decreases. On the other hand, uniformly and indiscriminately reducing the dimensionality of features can easily overlook the differences between various domains, leading to the loss of valuable information. Summary of the Invention
[0004] Purpose of the invention: The purpose of this invention is to provide an unsupervised diagnostic method and system for escalators that can both ensure the integrity of the feature system of the collected data and improve the accuracy of the diagnostic results.
[0005] Technical solution: The present invention provides an unsupervised diagnostic method for escalators, comprising the following steps:
[0006] Sensors were installed at multiple measuring points on the escalator to obtain vibration data of the escalator;
[0007] Multiple key time periods for collecting vibration data are set, and data collected at a certain time period at a certain measuring point is extracted to obtain single-point, single-time-period data.
[0008] Multi-dimensional feature extraction is performed on single-point, single-time-period data to obtain single-dimensional feature data of multiple different dimensions for single-point, single-time-period data;
[0009] Extract single-dimensional feature data of a single measuring point and a single time period for multiple days within the normal operation period of the escalator, and construct a health dataset of that dimension. Extract single-dimensional feature data of the same dimension of the same measuring point and the same time period for multiple days within the date to be diagnosed, and construct a corresponding dataset to be diagnosed.
[0010] Perform anomaly diagnosis on the diagnostic dataset and the health dataset, and output diagnostic results at different levels.
[0011] Furthermore, the multi-dimensional feature extraction includes:
[0012] Complete data feature extraction: Time domain features, frequency domain features, and time-frequency domain features are directly extracted from single measurement point and single time period data to obtain first-dimensional feature data, second-dimensional feature data, and third-dimensional feature data.
[0013] The slice data feature extraction first slices the single measurement point and single time period data in the time series dimension to obtain multiple short-time signals. Then, the short-time signals are subjected to time domain feature extraction and frequency domain feature extraction to obtain fourth-dimensional feature data and fifth-dimensional feature data.
[0014] Multiple extraction of time-frequency features involves performing time-frequency analysis on single-point, single-period data to obtain the time spectrum. Then, the obtained time spectrum is sliced in the time-series dimension to obtain multiple short-time time spectra. Finally, time-domain features are extracted from the short-time time spectra to obtain the sixth-dimensional feature data.
[0015] Furthermore, the anomaly diagnosis of the dataset to be diagnosed and the health dataset refers to:
[0016] First, the LOF model is used to identify anomalies in the dataset to be diagnosed, and the diagnostic score for each single-dimensional feature data is output.
[0017] Set a threshold threLOF for diagnosing motivational anomalies. Data to be diagnosed with a score ≥ threLOF is labeled as motivational anomalies, and data to be diagnosed with a score < threLOF is labeled as valid samples.
[0018] The K-means model is used to perform cluster analysis on the health dataset and valid samples, and the accuracy of the diagnostic results is output.
[0019] Furthermore, the diagnostic results at different levels include:
[0020] For single-dimensional feature diagnosis results, a threshold of stimulated abnormal data volume, stimulatePer1, is set. When the proportion of data in the dataset to be diagnosed that is marked as stimulated abnormal is ≥stimuPer1, the single-dimensional feature data is determined to be stimulated abnormal. A threshold of threKmeans is set for stability abnormal diagnosis. When accuracy ≥ threKmeans, the single-dimensional feature data is determined to be stable abnormal. When accuracy < threKmeans, the single-dimensional feature data is determined to be normal.
[0021] For single-time period diagnostic results, if the proportion of single-dimensional feature data judged as having abnormal stimulation within a certain time period is ≥stimuPer2, then the time period is judged as an abnormal stimulation period; if the proportion of single-dimensional feature data judged as having abnormal stability within the time period is ≥stablePer1, then the time period is judged as an abnormal stability period; otherwise, the time period is judged as a normal time period.
[0022] For single-point diagnostic results, if any single-period diagnostic result of a certain point exceeds posNum1 for a continuous period and is determined to be an excitation abnormal period, then the point is determined to be an excitation abnormal point. If any single-period diagnostic result of a certain point exceeds posNum2 for a continuous period and is determined to be a stability abnormal period, then the point is determined to be a stability abnormal point. Otherwise, the point is determined to be a normal point.
[0023] Furthermore, the multiple key time periods for collecting vibration data refer to time periods with passenger flow tidal characteristics and escalator equipment start-stop characteristics.
[0024] Furthermore, before performing anomaly diagnosis on the dataset to be diagnosed, the quantity of data to be diagnosed in the dataset to be diagnosed is first evaluated. If the quantity of data to be diagnosed is greater than numLOF, then LOF diagnosis is performed; otherwise, it is determined that the data quantity is insufficient.
[0025] Furthermore, before the effective samples are input into the K-means model, the number of effective samples is first evaluated. If the number of effective samples is greater than numKmeans, then the samples are input into the K-means model; otherwise, the data volume is deemed insufficient.
[0026] Based on the same inventive concept, the present invention also provides an unsupervised diagnostic system for escalators, comprising:
[0027] The monitoring module is capable of collecting vibration data from multiple measuring points on the escalator within a specified time period;
[0028] The data extraction module is used to classify and process the collected vibration data to obtain vibration data at different measuring points and at different times.
[0029] The data processing module is used to extract multi-dimensional features from vibration data at different measuring points and at different times, and obtain multiple single-dimensional feature data of vibration data at different measuring points and at different times.
[0030] The anomaly diagnosis module is used to analyze and diagnose the single-dimensional feature data of each measurement point, time period, and measurement point within multiple days of the diagnosis date, and output the diagnosis results of each single-dimensional feature, time period, and measurement point.
[0031] Furthermore, the multi-dimensional feature extraction includes:
[0032] Complete data feature extraction involves directly extracting time-domain, frequency-domain, and time-frequency-domain features from data of a certain measurement point over a certain period of time to obtain first-dimensional feature data, second-dimensional feature data, and third-dimensional feature data.
[0033] The slice data feature extraction first slices the data of a certain measurement point for a certain period of time in the time series dimension to obtain multiple short-time signals. Then, the short-time signals are subjected to time domain feature extraction and frequency domain feature extraction to obtain fourth-dimensional feature data and fifth-dimensional feature data.
[0034] Multiple extraction of time-frequency features involves performing time-frequency analysis on data from a certain measurement point over a certain period to obtain the time spectrum. Then, the obtained time spectrum is sliced in the time-series dimension to obtain multiple short-time time spectra. Finally, time-domain features are extracted from the short-time time spectra to obtain the sixth-dimensional feature data.
[0035] Furthermore, the anomaly diagnosis module includes:
[0036] The database construction submodule is used to construct a health dataset consisting of single-dimensional feature data of each time period at each measuring point within multiple days of the normal operation date of the escalator, and to construct a diagnosis dataset consisting of single-dimensional feature data of each time period at each measuring point within multiple days of the diagnosis date corresponding to the health dataset.
[0037] The LOF diagnostic submodule is used to identify abnormal data in the dataset to be diagnosed;
[0038] The Kmeans diagnostic submodule is used to perform cluster analysis on the health data in the health dataset and the valid data filtered by the LOF diagnostic submodule.
[0039] The analysis submodule is used to comprehensively evaluate the diagnostic results of the LOF diagnostic submodule and the Kmeans diagnostic submodule, and output the diagnostic results of each single-dimensional feature data, each time period, and each measurement point.
[0040] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: The diagnostic results of the present invention have a high accuracy rate and can accurately capture the sudden abnormal phenomena and stability abnormal phenomena of escalators; The present invention not only ensures the comprehensiveness of the data feature system, but also avoids the reduction in diagnostic effect caused by the simultaneous diagnosis of multi-dimensional features, the loss of effective information of each feature dimension, and the information interference between different feature dimensions, while avoiding the ambiguity of feature concepts caused by feature fusion, thus ensuring the interpretability of features. Attached Figure Description
[0041] Figure 1 This is a schematic diagram of the method structure of the present invention.
[0042] Figure 2 This is a schematic diagram of the multi-dimensional feature extraction structure of the present invention.
[0043] Figure 3 This is a schematic diagram of the structure for determining the excitation anomaly of the present invention.
[0044] Figure 4 This is a schematic diagram of the stability anomaly determination structure of the present invention.
[0045] Figure 5 This is a schematic diagram of the single-time-period anomaly determination structure of the present invention.
[0046] Figure 6 This is a schematic diagram of the single-point segment anomaly determination structure of the present invention.
[0047] Figure 7 This is a schematic diagram of the system structure of the present invention. Detailed Implementation
[0048] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0049] Example 1
[0050] like Figures 1 to 6 As shown, the unsupervised diagnostic method for escalators according to the present invention includes the following steps:
[0051] (1) Sensors are installed at multiple measuring points on the escalator to obtain vibration data of the escalator. Among them, multiple measuring points are selected from key components or components with high failure rates of the escalator. Therefore, when the escalator fails, the specific fault point and faulty component can be accurately located, and the direction of maintenance of the escalator can be determined. The common key measuring points of the escalator are shown in Table 1. In actual application, the measuring points can be added or reduced according to actual needs.
[0052] Table 1. Measurement Points for Escalators
[0053] Serial Number measuring point 1 motor 2 gearbox 3 Motor base 4 truss 5 Lower stairwell bearing - left 6 Lower stairwell bearing - right 7 Main drive wheel - left
[0054] (2) Set multiple key time periods for collecting vibration data, extract data collected at a certain time period at a certain measuring point, and obtain single measuring point single time period data; The operating conditions of escalators are mainly related to passenger flow and equipment start-up and shutdown. Except for special dates such as holidays, the passenger flow of public transportation venues where escalators are located is mostly related to morning peak, evening peak, long-distance passenger transport time, etc. These passenger flow changes generally have a periodicity on a daily basis. Therefore, in a day, select key time periods with large differences in operating conditions to collect vibration data according to passenger flow tides and equipment start-up and shutdown times. In this embodiment, five time periods as shown in Table 2 are selected as the time periods for collecting vibration data.
[0055] Table 2 Time periods for collecting vibration data
[0056] Time period Start time End time illustrate Period 1 5:00 6:59 Equipment startup Period 2 7:00 8:59 Morning rush hour Time period 3 12:00 13:59 Peak Period 4 17:00 18:59 Evening rush hour Period 5 22:00 23:00 Equipment stopped
[0057] (3) Perform multi-dimensional feature extraction on single measurement point and single time period data to obtain multiple single-dimensional feature data of different dimensions of single measurement point and single time period data; among which, multi-dimensional feature extraction includes complete data feature extraction, slice data feature extraction and time-frequency feature multiple extraction. Through the above three extraction methods, a total of six different dimensions of single-dimensional feature data of single measurement point and single time period data are obtained.
[0058] (31) Whole data feature extraction refers to the direct extraction of time domain features, frequency domain features, and time-frequency domain features from single measurement point and single time period data to obtain first-dimensional feature data, second-dimensional feature data, and third-dimensional feature data.
[0059] (32) Slice data feature extraction: First, slice the single measurement point and single time period data in the time series dimension to obtain multiple short-time signals. Then, extract the time domain features and frequency domain features of the short-time signals to obtain the fourth dimension feature data and the fifth dimension feature data.
[0060] (33) Multiple extraction of time and frequency features: Time and frequency analysis is performed on single measurement point and single time period data to obtain time spectrum. Then, the obtained time spectrum is sliced in the time series dimension to obtain multiple short time spectrums. Finally, time domain features are extracted from the short time spectrum to obtain the sixth dimension feature data.
[0061] To improve the accuracy of diagnostic results, data preprocessing is performed before multi-dimensional feature extraction of single-point, single-time-period data. Preprocessing includes noise reduction, filtering, and screening out invalid points and signal anomalies.
[0062] Escalators have long operating cycles, making it difficult to collect data for extended periods while maintaining the collection frequency. The multi-feature extraction method described above can simultaneously extract features from both complete and segmented data, ensuring that the extracted feature information covers both low-frequency data and high-frequency, short-term data.
[0063] (4) Extract single-dimensional feature data of a single measurement point and a single time period for multiple days within the normal operation date of the escalator, and construct a health dataset of that dimension. Extract single-dimensional feature data of the same dimension of the same measurement point and the same time period for multiple days within the date to be diagnosed, and construct a corresponding dataset to be diagnosed. Construct the health dataset and the dataset to be diagnosed to facilitate subsequent diagnosis.
[0064] (5) Perform anomaly diagnosis on the dataset to be diagnosed and the healthy dataset, and output diagnostic results at different levels; the specific diagnostic process is as follows:
[0065] (51) First, the number of data to be diagnosed in the dataset to be diagnosed is evaluated. If the number of data to be diagnosed is greater than numLOF, then LOF diagnosis is performed; otherwise, it is determined that the amount of data is insufficient. LOF (Local Outlier Factor) refers to the outlier detection algorithm. The specific value of numLOF can be set according to the actual situation of the escalator and combined with expert experience.
[0066] (52) The LOF model performs anomaly identification on the dataset to be diagnosed and outputs the diagnostic score for each single-dimensional feature data.
[0067] (53) Set the threshold threLOF for diagnosing stimuli abnormalities. Data to be diagnosed with score ≥ threLOF are labeled as stimuli abnormalities, and data to be diagnosed with score < threLOF are labeled as valid samples. The specific value of threLOF can be set according to the actual situation of the escalator and combined with expert experience.
[0068] (54) Evaluate the number of valid samples. If the number of valid samples is greater than numKmeans, input the Kmeans model; otherwise, it is determined that the data volume is insufficient. Kmeans refers to the clustering algorithm. The specific value of numKmeans can be set according to the actual situation of the escalator and combined with expert experience.
[0069] (55) The model performs cluster analysis on the health dataset and valid samples input into the Kmeans model and outputs the diagnostic accuracy.
[0070] (56) Output the single-dimensional feature diagnosis results, set the threshold for the proportion of stimuli abnormal data, and when the proportion of the data to be diagnosed that is marked as stimuli abnormal in the dataset to be diagnosed is ≥stimuPer1, the single-dimensional feature data is determined to be stimuli abnormal. Set the threshold for stability abnormal diagnosis, threKmeans. When accuracy ≥ threKmeans, the single-dimensional feature data is determined to be stability abnormal. When accuracy < threKmeans, the single-dimensional feature data is determined to be normal. The specific values of stimuPer1 and threKmeans can be set according to the actual situation of the escalator and combined with expert experience.
[0071] (57) Output the diagnosis results for a single time period. If the proportion of single-dimensional feature data that is judged to be abnormal in a certain time period is ≥stimuPer2, then the time period is judged to be an abnormal in excitation period. If the proportion of single-dimensional feature data that is judged to be abnormal in stability is ≥stablePer1, then the time period is judged to be an abnormal in stability period. Otherwise, the time period is judged to be a normal time period. The specific values of stimulatePer2 and stablePer1 can be set according to the actual situation of the escalator and combined with expert experience.
[0072] (58) Output single-point diagnostic results. If the diagnostic results of any single time period of a certain point exceed posNum1 for a continuous period of time, it is determined to be an excitation abnormal period. If the diagnostic results of any single time period of a certain point exceed posNum2 for a continuous period of time, it is determined to be a stability abnormal period. Otherwise, the point is determined to be a normal point. The specific values of posNum1 and posNum2 can be set according to the actual situation of the escalator and combined with expert experience.
[0073] This anomaly diagnosis method ultimately outputs various levels of diagnostic results, such as insufficient data volume, abnormal excitation of single-dimensional feature data, abnormal stability of single-dimensional feature data, normal single-dimensional feature data, abnormal excitation of single time period, abnormal stability of single time period, normal single time period, abnormal excitation of single measurement point, abnormal stability of single measurement point, and normal single measurement point. This allows staff to understand the operating status of the escalator from different perspectives, which is beneficial for staff to maintain the escalator.
[0074] The above-mentioned anomaly diagnosis method first analyzes the single-dimensional feature data of different dimensions independently, and then performs fusion diagnosis on the evaluation results of multiple feature dimensions through a diagnostic result fusion strategy. This diagnostic method not only ensures the comprehensiveness of the feature system, but also avoids the loss of effective information of each feature dimension, avoids information interference between different feature dimensions, avoids the ambiguity of feature concepts caused by feature fusion, and ensures the interpretability of features.
[0075] Example 2
[0076] like Figures 1 to 7 As shown, the unsupervised diagnostic system for escalators according to the present invention includes a monitoring module, a data extraction module, a data processing module, and an anomaly diagnosis module. The monitoring module can collect vibration data from multiple measuring points of the escalator within a specified time period. The data extraction module is used to classify and process the collected vibration data to obtain vibration data from different measuring points at different time periods. The data processing module can perform multi-dimensional feature extraction on the vibration data from different measuring points at different time periods to obtain multiple single-dimensional feature data of different dimensions of vibration data from different measuring points at different time periods. The anomaly diagnosis module can analyze and diagnose the single-dimensional feature data of each measuring point and each time period within multiple days of the diagnosis date, and output the diagnostic results of each single-dimensional feature, each time period, and each measuring point.
[0077] The multi-dimensional feature extraction includes: complete data feature extraction, which directly extracts time-domain, frequency-domain, and time-frequency-domain features from data of a specific measurement point over a specific time period to obtain first-dimensional, second-dimensional, and third-dimensional feature data; sliced data feature extraction, which first slices the data of a specific measurement point over a specific time period in the time-series dimension to obtain multiple short-time signals, and then extracts time-domain and frequency-domain features from these short-time signals to obtain fourth-dimensional and fifth-dimensional feature data; and multiple time-frequency feature extraction, which performs time-frequency analysis on the data of a specific measurement point over a specific time period to obtain a time spectrum, then slices the obtained time spectrum in the time-series dimension to obtain multiple short-time time spectra, and finally extracts time-domain features from these short-time time spectra to obtain a sixth-dimensional feature data. This multi-dimensional feature extraction can obtain single-dimensional feature data in various dimensions, facilitating independent analysis of each single-dimensional feature data by the subsequent anomaly diagnosis module and ensuring the integrity of the features represented by the collected data.
[0078] The anomaly diagnosis module includes a database construction submodule, an LOF diagnosis submodule, a K-means diagnosis submodule, and an analysis submodule.
[0079] The database construction submodule is used to construct a health dataset consisting of single-dimensional feature data of each time period at each measuring point within multiple days of the escalator's normal operation, and to construct a diagnosis dataset consisting of single-dimensional feature data of each time period at each measuring point within multiple days of the corresponding diagnosis date of the health dataset.
[0080] The LOF diagnostic submodule is used to identify abnormal data in the dataset to be diagnosed. LOF diagnosis is used to capture stimuli (sudden) anomalies in escalators. Data that is diagnosed as normal by LOF will enter the Kmeans diagnostic submodule for secondary diagnosis in order to prevent the occurrence of group-wide stimuli.
[0081] The Kmeans diagnostic submodule is used to perform cluster analysis on health data in the health dataset and valid data filtered by the LOF diagnostic submodule, thereby capturing abnormal phenomena in escalator stability.
[0082] The analysis submodule is used to comprehensively evaluate the diagnostic results of the LOF diagnostic submodule and the Kmeans diagnostic submodule, and output the diagnostic results of each single-dimensional feature data, each time period, and each measurement point. When analyzing the diagnostic results of each single-dimensional feature data, the determination of stimulating anomalies is based on the proportion of LOF diagnostics that classifies them as stimulating anomalies. When the proportion exceeds a preset value, the single-dimensional feature data is classified as stimulating anomaly. Whether the single-dimensional feature data is stable anomaly is determined by the K-means diagnostic output. When analyzing the diagnostic results of each measurement point, the proportion of single-dimensional feature data classified as stimulating anomalies and stable anomalies for that measurement point is calculated. When the proportion exceeds the corresponding preset value, the measurement point is classified as stimulating anomaly and / or stable anomaly. When analyzing the diagnostic results of each time period, the diagnostic results of all time periods for that measurement point are calculated. If any single time period diagnostic result for that measurement point is classified as a period of stimulating anomalies for more than N consecutive days, the measurement point is classified as a period of stimulating anomalies. If any single time period diagnostic result for that measurement point is classified as a period of stable anomalies for more than M consecutive days, the measurement point is classified as a period of stable anomalies. N and M are preset values.
[0083] The anomaly diagnosis module first analyzes the single-dimensional feature data of different dimensions independently, and then performs fusion diagnosis on the evaluation results of multiple feature dimensions through a diagnosis result fusion strategy. This diagnostic method not only ensures the comprehensiveness of the feature system, but also avoids the reduced diagnostic effect, loss of effective information of each feature dimension, and information interference between different feature dimensions caused by simultaneous diagnosis of multiple features. At the same time, it avoids the ambiguity of feature concepts caused by feature fusion and ensures the interpretability of features.
Claims
1. An unsupervised diagnostic method for escalators, characterized in that: Includes the following steps, Sensors were installed at multiple measuring points on the escalator to obtain vibration data of the escalator; Multiple key time periods for collecting vibration data are set, and data collected at a certain time period at a certain measuring point is extracted to obtain single-point, single-time-period data. Multi-dimensional feature extraction is performed on single-point, single-time-period data to obtain single-dimensional feature data of multiple different dimensions for single-point, single-time-period data; Multi-dimensional feature extraction includes: Complete data feature extraction: Directly extract time-domain features, frequency-domain features, and time-frequency-domain features from single-point, single-time-period data to obtain first-dimensional feature data, second-dimensional feature data, and third-dimensional feature data; The slice data feature extraction first slices the single measurement point and single time period data in the time series dimension to obtain multiple short-time signals. Then, the short-time signals are subjected to time domain feature extraction and frequency domain feature extraction to obtain fourth-dimensional feature data and fifth-dimensional feature data. Multiple extraction of time-frequency features: Time-frequency analysis is performed on single measurement point and single time period data to obtain time spectrum. Then, the obtained time spectrum is sliced in the time series dimension to obtain multiple short-time time spectra. Finally, time domain features are extracted from the short-time time spectra to obtain the sixth dimension feature data. Extract single-dimensional feature data of a single measuring point and a single time period for multiple days within the normal operation period of the escalator, and construct a health dataset of that dimension. Extract single-dimensional feature data of the same dimension of the same measuring point and the same time period for multiple days within the date to be diagnosed, and construct a corresponding dataset to be diagnosed. Perform anomaly diagnosis on the diagnostic dataset and the health dataset, and output diagnostic results at different levels; where performing anomaly diagnosis on the diagnostic dataset and the health dataset refers to: First, the LOF model is used to identify anomalies in the dataset to be diagnosed, and the diagnostic score for each single-dimensional feature data is output. Set a threshold threLOF for diagnosing motivational anomalies. Data to be diagnosed with a score ≥ threLOF is labeled as motivational anomalies, and data to be diagnosed with a score < threLOF is labeled as valid samples. The K-means model is used to perform cluster analysis on the health dataset and valid samples, and the accuracy of the diagnostic results is output.
2. The unsupervised diagnostic method for escalators according to claim 1, characterized in that: The different levels of diagnostic results include: For single-dimensional feature diagnosis results, a threshold of stimulated abnormal data volume, stimulatePer1, is set. When the proportion of data in the dataset to be diagnosed that is marked as stimulated abnormal is ≥stimuPer1, the single-dimensional feature data is determined to be stimulated abnormal. A threshold of threKmeans is set for stability abnormal diagnosis. When accuracy ≥ threKmeans, the single-dimensional feature data is determined to be stable abnormal. When accuracy < threKmeans, the single-dimensional feature data is determined to be normal. For single-time period diagnostic results, if the proportion of single-dimensional feature data judged as having abnormal stimulation within a certain time period is ≥stimuPer2, then the time period is judged as an abnormal stimulation period; if the proportion of single-dimensional feature data judged as having abnormal stability within the time period is ≥stablePer1, then the time period is judged as an abnormal stability period; otherwise, the time period is judged as a normal time period. For single-point diagnostic results, if any single-period diagnostic result of a certain point exceeds posNum1 for a continuous period and is determined to be an excitation abnormal period, then the point is determined to be an excitation abnormal point. If any single-period diagnostic result of a certain point exceeds posNum2 for a continuous period and is determined to be a stability abnormal period, then the point is determined to be a stability abnormal point. Otherwise, the point is determined to be a normal point.
3. The unsupervised diagnostic method for escalators according to claim 1, characterized in that: The key time periods for collecting vibration data refer to the time periods with passenger flow tidal characteristics and escalator equipment start-stop characteristics.
4. The unsupervised diagnostic method for escalators according to claim 1, characterized in that: Before performing anomaly diagnosis on the dataset to be diagnosed, the quantity of data to be diagnosed in the dataset is first evaluated. If the quantity of data to be diagnosed is greater than numLOF, then LOF diagnosis is performed; otherwise, it is determined that the data quantity is insufficient.
5. The unsupervised diagnostic method for escalators according to claim 1, characterized in that: Before inputting the valid samples into the K-means model, the number of valid samples is first evaluated. If the number of valid samples is greater than numKmeans, the samples are input into the K-means model; otherwise, the data volume is considered insufficient.
6. An unsupervised diagnostic system for escalators, characterized in that: include: The monitoring module is capable of collecting vibration data from multiple measuring points on the escalator within a specified time period; The data extraction module is used to classify and process the collected vibration data to obtain vibration data at different measuring points and at different times. The data processing module is used to extract multi-dimensional features from vibration data at different measuring points and at different times, and obtain multiple single-dimensional feature data of vibration data at different measuring points and at different times. The multi-dimensional feature extraction includes: Complete data feature extraction involves directly extracting time-domain, frequency-domain, and time-frequency-domain features from data of a certain measurement point over a certain period of time to obtain first-dimensional feature data, second-dimensional feature data, and third-dimensional feature data. The slice data feature extraction first slices the data of a certain measurement point for a certain period of time in the time series dimension to obtain multiple short-time signals. Then, the short-time signals are subjected to time domain feature extraction and frequency domain feature extraction to obtain fourth-dimensional feature data and fifth-dimensional feature data. The system employs multiple extraction of time-frequency features. It performs time-frequency analysis on data from a specific measurement point over a specific time period to obtain a time spectrum. Then, it slices the obtained time spectrum along the time-series dimension to obtain multiple short-time time spectra. Finally, it extracts time-domain features from these short-time time spectra to obtain the sixth-dimensional feature data. The anomaly diagnosis module analyzes and diagnoses various single-dimensional feature data from different measurement points and time periods over multiple days leading up to the diagnosis date, and outputs diagnostic results for each single-dimensional feature, time period, and measurement point. The anomaly diagnosis module includes: The database construction submodule is used to construct a health dataset consisting of single-dimensional feature data of each time period at each measuring point within multiple days of the normal operation date of the escalator, and to construct a diagnosis dataset consisting of single-dimensional feature data of each time period at each measuring point within multiple days of the diagnosis date corresponding to the health dataset. The LOF diagnostic submodule is used to identify abnormal data in the dataset to be diagnosed; The Kmeans diagnostic submodule is used to perform cluster analysis on the health data in the health dataset and the valid data filtered by the LOF diagnostic submodule. The analysis submodule is used to comprehensively evaluate the diagnostic results of the LOF diagnostic submodule and the Kmeans diagnostic submodule, and output the diagnostic results of each single-dimensional feature data, each time period, and each measurement point.