Escalator fault detection method and system with data self-confirmation function

Abstract

The invention belongs to the technical field of escalator fault monitoring and data processing, and relates to an escalator fault detection method and system with data self-confirmation, and the method comprises the steps: collecting key part sensor data, and generating an original time sequence; carrying out differential operation on the data and merging the data with original data to form sensor diagnosis composite data; judging whether the sensor is abnormal or not by using the sensor state diagnosis model, and generating a state confirmation report; when the sensors are normal, pure vibration data streams are extracted; calculating multiple groups of feature scale parameters, segmenting and compressing the vibration data, and generating a multi-scale feature fusion vector; after position codes are added, the escalator state diagnosis model is input, health state initial judgment is obtained, and finally a comprehensive diagnosis conclusion is output. According to the method, the problems that false alarm is easily generated when the sensor has a fault, and early fault features are lost and the diagnosis precision is limited due to the fact that a traditional method is difficult to adapt to multi-band signals are solved.

Description

Technical Field

[0001] This application belongs to the technical field of escalator fault monitoring and data processing, and relates to an escalator fault detection method and system with data self-verification. Background Technology

[0002] Escalators, as large-scale passenger-carrying special equipment widely used in public transportation and commercial venues, directly impact the personal and property safety of passengers through the operational safety of their mechanical systems. Escalators typically consist of several key components, including a motor, gearbox, drive shaft, tensioning device, and step rails. Operating under high load and continuous operation for extended periods, they are prone to malfunctions such as bearing wear, gear failure, or loose mechanical connections. To achieve real-time monitoring of equipment status, the industry typically deploys various types of sensors, such as temperature and vibration sensors, in key transmission components. By collecting and analyzing the physical signals of the equipment, abnormalities can be identified, thus enabling early warning of malfunctions.

[0003] Currently, the commonly used escalator monitoring solutions in the industry mainly rely on signal analysis methods based on fixed threshold discrimination or basic statistical feature extraction. These traditional solutions typically assume that the data collected by the front-end sensors is absolutely true and reliable. They directly filter and denoise the raw signal, calculate the root mean square value, peak value, or perform spectral analysis, and compare the calculation results with preset industry standard thresholds. If the limit is exceeded, a fault is determined. Although some solutions introduce machine learning algorithms, they still mostly use raw time-domain data as input and rarely consider the influence of mechanical structural parameters on the signal's characteristic frequencies.

[0004] However, due to the general lack of independent diagnostic mechanisms for the health status of sensors in existing technologies, when sensors experience inherent faults such as zero-point drift, decreased sensitivity, constant bias, or signal stagnation, their abnormal output data is easily misjudged as escalator mechanical faults, leading to frequent false alarms. Furthermore, because escalator drive chains contain multi-frequency motion components, traditional single-scale or fixed-resolution signal processing methods struggle to meet the feature extraction requirements of components operating at different speeds. This can easily result in the loss of weak signals characterizing early faults during data processing, thus limiting the diagnostic accuracy of complex mechanical systems. Summary of the Invention

[0005] In a first aspect, the present invention provides a method for detecting escalator faults with self-verification of data, employing the following technical solution: An escalator fault detection method with data self-verification includes the following steps: S1. Collect sensor data installed on key transmission components of escalators, and convert the continuously collected temperature and vibration simulation signals within a specified time period into a raw sensor time series containing the readings and timestamps of each sensor. S2. Receive the raw sensor time series, perform differential operations on it to extract the dynamic change characteristics of the data, and merge it with the raw data to generate sensor diagnostic composite data; S3. Input the sensor diagnostic composite data into the pre-trained sensor status diagnostic model. The model will determine whether there are any abnormalities in the data caused by sensor malfunctions and generate a sensor status confirmation report. S4. Analyze the sensor status confirmation report. When it is confirmed that all sensors are in normal working condition, extract the vibration data portion from the original sensor time series and generate a clean vibration data stream. S5. Based on the inherent rotation frequency and data sampling frequency of each moving part of the escalator, calculate multiple sets of characteristic scale parameters, and use these characteristic scale parameters to segment and compress the pure vibration data stream to generate a multi-scale feature fusion vector. S6. Add position encoding to the multi-scale feature fusion vector and input the result into the pre-trained escalator status diagnosis model to generate an initial judgment of the escalator health status. S7. Integrate the sensor status confirmation report with the preliminary assessment of the escalator's health status to generate the final comprehensive equipment diagnostic conclusion.

[0006] A further aspect of the present invention generates sensor diagnostic composite data, comprising the following steps: Set a relative change scale to define the time interval for calculating dynamic changes in data; Subtract the data point at the current moment in the original sensor time series from the data point that is T moments later on the time axis to obtain the relative change sequence; The relative change sequence is aligned with the original sensor time series according to the timestamps of the data points, and the values ​​at the corresponding positions are added point by point. The merged sequence is then output as the sensor diagnostic composite data.

[0007] In a further embodiment of the present invention, the sensor state diagnosis model in step S3 is a classification model built on a neural network, which can identify at least the following sensor fault types: bias fault, drift fault, spike fault, accuracy degradation fault, and jamming fault. The discrimination process includes: dividing the sensor diagnostic composite data into multiple data segments, inputting them into the model to calculate the probability distribution of each fault type; if a fault is identified, marking the faulty sensor number, location, and fault type in the sensor status confirmation report; if no fault is identified, marking the sensor status as normal.

[0008] A further aspect of the present invention generates a clean vibration data stream, comprising the following steps: Check the status markers in the sensor status confirmation report and only perform data extraction if all markers are normal. Based on the preset data storage structure mapping table, the reading indexes corresponding to all temperature sensors are identified and removed from the original sensor time series, while vibration signal segments corresponding to all vibration sensors are retained and extracted. The extracted vibration signal segments are spliced ​​together according to the physical installation order of the sensors on the escalator to form a one-dimensional pure vibration data stream.

[0009] A further aspect of this invention involves calculating multiple sets of feature scale parameters, including the following steps: Based on the motor speed and bearing structure parameters of the escalator, calculate the theoretical rotational frequency of key moving parts in the escalator transmission system. The theoretical rotational frequency includes at least the inner ring rotational frequency, the cage rotational frequency, the rolling element rotational frequency, the cage rotational frequency relative to the inner ring, the rolling element rolling over the inner ring surface, and the rolling element rolling over the outer ring surface. Divide the data sampling frequency by the calculated theoretical rotation frequency to obtain the dimensionless integer value, which is the characteristic scale parameter.

[0010] A further aspect of this invention generates a multi-scale feature fusion vector, comprising the following steps: For each set of calculated characteristic scale parameters, the pure vibration data stream is divided into several continuous data segments according to the length defined by the parameter; Extract the maximum value within each data segment to generate a compressed feature sequence at that scale; By concatenating the compressed feature sequences generated at all scales, a multi-scale feature fusion vector is constructed.

[0011] In a further embodiment of the present invention, the position encoding in step S6 is calculated using sine and cosine functions and is used to assign temporal position information to the data points in the multi-scale feature fusion vector. The escalator status diagnosis model is built on the Transformer architecture and includes a multi-head self-attention mechanism module, which is used to capture the long-distance dependencies within the multi-scale feature fusion vector and output the classification result of whether the escalator is currently in a normal state or a specific mechanical fault state as a preliminary judgment of the escalator's health status.

[0012] A further aspect of the present invention generates a comprehensive diagnostic conclusion for the equipment, comprising the following steps: Determine if a fault marker exists in the sensor status confirmation report; If a fault marker exists, the sensor fault information in the report is extracted as the comprehensive diagnostic conclusion for the equipment, and the judgment of the escalator status is suspended. If no fault marker is found, extract the escalator fault information or normal status information from the initial health status assessment of the escalator as the comprehensive diagnostic conclusion for the equipment.

[0013] In a further embodiment of the present invention, the location of the sensor collected in step S1 includes at least the following: the motor load end housing, the motor non-load end housing, the gearbox input end housing, the gearbox output end housing, the main drive bearing housing, the handrail belt bearing housing, the step rail, and the tension bracket bearing housing of the escalator. Arrange temperature and vibration combined sensors or single-channel vibration sensors at the location, and synchronously collect and generate the original sensor time series at a preset sampling frequency.

[0014] Secondly, the present invention provides an escalator fault detection system with self-verification of data, employing the following technical solution: An escalator fault detection system with self-verification of data includes the following modules: The data acquisition and sequence generation module is used to acquire sensor data installed on key transmission components of escalators and convert the continuously acquired temperature and vibration analog signals within a specified time period into a raw sensor time series containing the readings and acquisition timestamps of each sensor. The composite input construction module is used to receive the raw sensor time series, perform differential operations on it to extract the dynamic change features of the data, and merge it with the raw data to generate sensor diagnostic composite data. The sensor status diagnosis module is used to input the sensor diagnosis composite data into a pre-trained sensor status diagnosis model. The model determines whether there are any abnormalities in the data caused by sensor malfunctions and generates a sensor status confirmation report. The data cleaning and extraction module is used to parse the sensor status confirmation report. When it is confirmed that all sensors are in normal working condition, the vibration data part of the original sensor time series is extracted to generate a clean vibration data stream. The multi-scale feature fusion module calculates multiple sets of feature scale parameters based on the inherent rotation frequency and data sampling frequency of each moving part of the escalator, and uses these feature scale parameters to segment and compress the pure vibration data stream to generate a multi-scale feature fusion vector. The escalator health diagnosis module is used to add position encoding to the multi-scale feature fusion vector and input the result into the pre-trained escalator status diagnosis model to generate an initial judgment of the escalator health status. The comprehensive diagnostic output module integrates the sensor status confirmation report with the preliminary assessment of the escalator's health status to generate the final comprehensive equipment diagnostic conclusion.

[0015] In summary, the present invention has the following beneficial technical effects: 1. This invention receives raw sensor time series data and performs differential operations to extract dynamic change features. This extracted data is then merged with the original data to generate composite sensor diagnostic data. A pre-trained sensor state diagnostic model is used to independently identify the data source. This process prioritizes and eliminates sensor data with bias, drift, or jamming faults before diagnosing escalator mechanical faults. By implementing a pre-diagnostic data self-verification step, it reduces false signal interference caused by sensor hardware failures, improves the purity of subsequent input data, and lowers the probability of false alarms related to escalator faults.

[0016] 2. Based on physical structural parameters such as escalator motor speed and bearing type, this invention calculates the theoretical rotational frequency of each moving component and combines the sampling frequency with these characteristic frequencies to generate multiple sets of characteristic scale parameters. Using these parameters, the pure vibration data stream is segmented and compressed at multiple levels to generate a multi-scale feature fusion vector. This feature extraction method based on a physical model enables the processed data vector to simultaneously contain feature information adapted to different frequency bands such as high-speed shaft rotation and low-speed chain operation, reducing the limitations of single-scale analysis when processing multi-frequency coupled mechanical signals and enhancing the system's ability to capture weak fault features in complex transmission chains.

[0017] 3. This invention introduces position encoding after generating multi-scale feature fusion vectors to supplement the temporal dimension information of the sequence data. Combined with a deep learning-based escalator condition diagnosis model, and utilizing the nonlinear mapping capability of neural networks, along with feature vectors containing positional information, it can more deeply uncover the temporal dependencies in vibration signals. In this way, it can not only determine whether the escalator is faulty, but also identify specific fault types (such as bearing wear types) and fault locations based on the learned fault modes, improving the granularity and accuracy of fault diagnosis results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0020] Figure 2 Structural schematic diagrams of embodiments of this application are disclosed. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The following is in conjunction with the appendix Figure 1 - Figure 2 A preferred description of the present invention is provided below.

[0023] See attached document Figure 1 This invention proposes a method for detecting escalator faults with data self-verification, comprising the following steps: S1. Collect sensor data installed on key transmission components of escalators, and convert the continuously collected temperature and vibration simulation signals within a specified time period into a raw sensor time series containing the readings and timestamps of each sensor. S2. Receive the raw sensor time series, perform differential operations on it to extract the dynamic change characteristics of the data, and merge it with the raw data to generate sensor diagnostic composite data; S3. Input the sensor diagnostic composite data into the pre-trained sensor status diagnostic model. The model will determine whether there are any abnormalities in the data caused by sensor malfunctions and generate a sensor status confirmation report. S4. Analyze the sensor status confirmation report. When it is confirmed that all sensors are in normal working condition, extract the vibration data portion from the original sensor time series and generate a clean vibration data stream. S5. Based on the inherent rotation frequency and data sampling frequency of each moving part of the escalator, calculate multiple sets of characteristic scale parameters, and use these characteristic scale parameters to segment and compress the pure vibration data stream to generate a multi-scale feature fusion vector. S6. Add position encoding to the multi-scale feature fusion vector and input the result into the pre-trained escalator status diagnosis model to generate an initial judgment of the escalator health status. S7. Integrate the sensor status confirmation report with the preliminary assessment of the escalator's health status to generate the final comprehensive equipment diagnostic conclusion.

[0024] Specifically, technicians first install monitoring devices along the escalator's power transmission path. These devices must cover the motor's load-side housing, unload-side housing, gearbox input housing, gearbox output housing, the left and right sides of the main drive bearing housing, the left and right sides of the handrail bearing housing, the left and right sides of the step rails, and the left and right sides of the tension bracket bearing housing. These monitoring devices consist of temperature and vibration combination sensors and single-channel vibration sensors. After activating the sensors, the vibration signal sampling frequency is set to 1000Hz, and the data acquisition period is set to 10 seconds. Within each acquisition period, the sensors convert temperature changes and mechanical vibrations in the physical environment into analog electrical signals. The acquisition card receives these analog electrical signals and converts them into digital format. For the temperature and vibration combination sensor, one temperature value and ten thousand vibration values ​​are extracted within one period; for the vibration sensor, ten thousand vibration values ​​are extracted within one period. Subsequently, the escalator's unique serial number and the millisecond-level timestamp of the start time of this acquisition are read. The digital readings from each sensor are then concatenated according to the predetermined sensor number sequence to form a continuous numerical array. Finally, the escalator number, the data collection timestamp, and the sensor value array are packaged and combined to construct the original sensor time series.

[0025] The raw sensor time series is a structured object containing device identification information, time dimension information, and complete physical monitoring data. The data portion of this series is a one-dimensional floating-point array of length 120010, including data from all monitoring points within 10 seconds. The temperature and vibration combined sensor is an integrated sensor capable of simultaneously measuring surface temperature through a thermistor and vibration acceleration through a piezoelectric element. The data acquisition card is the signal processing hardware connecting the sensor and the computer system, responsible for performing high-precision analog-to-digital conversion tasks.

[0026] For example, for the escalator numbered ES_2F_03, its data acquisition timestamp is recorded as 1744680313000. The raw sensor time series generated at this time point contains an array Data consisting of 120010 floating-point numbers. The first bit of this array represents the temperature value of the motor load end casing; the second to the 10001st bits represent the vibration acceleration sequence corresponding to that position within 10 seconds; the 10002nd bit represents the temperature value of the motor non-load end casing; the 10003rd to the 20002nd bits represent its corresponding vibration acceleration sequence; and so on, until the 120010th bit at the end of the array, completely recording the operating conditions of the 12 sensor nodes within a specific 10-second interval. In one embodiment of the present invention, step S2 includes the following steps: Set a relative change scale to define the time interval for calculating the dynamic changes of data; subtract the data point at the current moment in the original sensor time series from the data point that is lagging behind by T moments on the time axis to obtain the relative change sequence; align the relative change sequence with the original sensor time series according to the timestamps of the data points, and add the values ​​at the corresponding positions point by point, and output the merged sequence as the sensor diagnostic composite data.

[0027] Specifically, technicians first set the relative change scale T, which defines the time span step for calculating dynamic features in a discrete time series. The original sensor time series generated in step S1 is received, and temporal difference operations are performed to extract dynamic change features. During the operation, the data points from the start position 1 to T are kept unchanged or filled with zeros; for each data point with position i greater than T, the difference between the current value xi and the value xi−T T time steps prior is calculated, thus generating a relative change sequence Δxi of the same length as the original sequence. Subsequently, a feature fusion operation is performed, aligning the relative change sequence with the original sensor time series according to timestamps, and adding the values ​​at corresponding positions point by point, i.e., calculating the sum of the original reading and the dynamic difference value for each sampling point, and outputting the merged sequence data as sensor diagnostic composite data. This process enhances the data's ability to express the non-stationary change trend of the sensor signal by superimposing first-order or multi-order difference information.

[0028] This step involves difference extraction and linear superposition of discrete time series, as shown in the following formula: in: Represents the original sensor time series. Values ​​at each sampling point; This indicates the relative scale of change, and is set based on the defined step size for capturing motion information in the data. For example, setting it to 1 corresponds to the first-order difference. This represents the calculated relative change sequence; Indicates the sensor diagnostic composite input number Point value.

[0029] The relative change scale is a hyperparameter in positive integer form, used to determine the size of the time lag window when calculating the dynamic trend of the data. Sensor diagnostic composite data is a numerical sequence that integrates sensor static readings and dynamic rate of change characteristics. Its data structure is consistent with the original sensor time series, serving as the input vector for subsequent neural network models.

[0030] For example, for the raw sensor time series of the escalator ES_2F_03, a relative change scale is set. The value is 1. Three consecutive vibration acceleration data points are selected from the sequence: 0.0232, 0.0235, and 0.0238. For the first data point, 0.0232, its position is not greater than... The corresponding difference value is 0, and the composite input value is 0.0232 + 0 = 0.0232. For the second data point 0.0235, its relative change is calculated as 0.0235 - 0.0232 = 0.0003, and the composite input value is 0.0235 + 0.0003 = 0.0238. For the third data point 0.0238, its relative change is calculated as 0.0238 - 0.0235 = 0.0003, and the composite input value is 0.0238 + 0.0003 = 0.0241. This logic processes all the data to generate the final sensor diagnostic composite data.

[0031] In one embodiment of the present invention, step S3 includes the following steps: The sensor status diagnosis model in step S3 is a classification model built on a neural network. It can identify at least the following sensor fault types: bias fault, drift fault, spike fault, accuracy degradation fault, and jamming fault. The discrimination process includes: dividing the sensor diagnostic composite data into multiple data segments, inputting them into the model to calculate the probability distribution of each fault type, and if a fault is identified, marking the faulty sensor number, location, and fault type in the sensor status confirmation report; if no fault is identified, marking the sensor status as normal.

[0032] Specifically, technicians first split the received sensor diagnostic composite data according to sensor channels, extracting a data sequence of 10,000 bytes for each vibration sensor. Based on the preset dimensions of the model input layer, such as a batch size of 10 and a single sample length of 1,000 bytes, the long sequence of each sensor is divided into 10 sub-segments of 1,000 bytes each. These sub-segments are then input into the sensor state diagnostic model. Internally, the model first performs feature mapping on the input vector through a linear layer, then processes the data through a normalization layer and an activation function layer to standardize the data distribution and introduce non-linear features. Finally, a Softmax layer calculates the probability distribution of the data belonging to each preset state category. The model selects the category with the highest probability as the diagnostic result for that data segment. The diagnostic results of all segments are collected. If any sensor's state category is identified as a fault (such as bias, drift, etc.), the sensor's number, installation location, and specific fault type are recorded in the sensor state confirmation report; if all sensor diagnostic results are normal, the system state is marked as fully normal in the report.

[0033] This step involves the forward propagation calculations within the neural network, including linear transformation, statistical normalization, and probability calculations. Here, based on statistical principles, it is modified to the form of a sum of squared deviations, as shown in the following formula: Linear mapping and input processing: Layer Normalization Statistic Calculation: Normalization transformation and activation: Probability calculation and classification judgment: in, That is, the aforementioned ; This represents the weight transformation of a linear layer; Output for linear layers; The sequence length (e.g., 1000); and These are the mean and standard deviation of the input sequence, respectively. and These are the affine parameters that the model can learn; To modify the activation function of the linear unit; Total number of state categories; For the model output layer corresponding to the first Logits of categories; For belonging to the first The probability of the category; This is the state index for the final determination.

[0034] It should be noted that the sensor condition diagnostic model is a pre-trained deep learning neural network. In this embodiment, its architecture is a multilayer perceptron (MLP) containing one or more hidden layers. Its output layer can distinguish between eight states: normal state, bias fault, drift fault, spike fault, accuracy degradation fault, jamming fault, data loss fault, and random fault. The sensor condition confirmation report is a structured data document used to store detailed identification information and fault attributes of all abnormal sensors discovered during the diagnostic process, or to store system health confirmation markers when no faults are found.

[0035] For example, for the sensor diagnostic composite data of escalator ES_2F_03, the 10,000 data points of the handrail belt bearing housing (left side) sensor (number 7) are divided into 10 input vectors. After inputting into the sensor status diagnostic model, the model calculates that the probability of the current data belonging to the "bias fault" category is 0.98, which is higher than other categories. Therefore, the model determines that the sensor is abnormal. A sensor status confirmation report is then generated, which records the following under the corresponding entry: SensorID is 7, SensorPosition is "handrail belt bearing housing (left side)", SensorType is "temperature and vibration combined sensor", and FaultDescription is marked as "bias fault".

[0036] In one embodiment of the present invention, step S4 includes the following steps: Check the status markers in the sensor status confirmation report and perform data extraction only when all markers are normal. Based on the preset data storage structure mapping table, identify and remove the reading indices corresponding to all temperature sensors from the original sensor time series, and retain and extract the vibration signal segments corresponding to all vibration sensors. The extracted vibration signal segments are spliced ​​together according to the physical installation order of the sensors on the escalator to form a one-dimensional pure vibration data stream.

[0037] Specifically, technicians first analyze the sensor status confirmation report output in step S3, iterating through the diagnostic status fields of each sensor node recorded in the report. The program executes logical judgments; if any sensor's fault description field is found to be non-empty or its status is marked as abnormal, the collected data within the current time window is considered unreliable, the subsequent processing flow is terminated, and the event is recorded in the maintenance log. Only when the report confirms that the status of sensors at all installation locations is normal, such as the 12 monitoring points at the motor, gearbox, bearing housing, and rail, is the data cleaning program activated.

[0038] At this point, the original sensor time series corresponding to the report is retrieved. Based on the preset escalator data storage structure mapping table, such as filtering the 120010-bit Data array bit by bit, all temperature reading indices corresponding to the temperature and vibration combined sensors are identified and removed (e.g., bit 1, bit 10002, etc.), retaining only the index range of the vibration acceleration signal (e.g., bits 2 to 10001, bits 10003 to 20002, etc.). For sensor channels that only measure vibration, such as those at the escalator rails, their corresponding continuous index ranges are directly retained. Finally, all extracted vibration signal segments are spliced ​​together end-to-end according to a predetermined order of sensor physical installation locations (e.g., from the motor end to the tension bracket end), constructing a one-dimensional continuous floating-point array containing only vibration acceleration values—a pure vibration data stream—as a reliable basis for subsequent multi-scale feature analysis.

[0039] The sensor status confirmation report is a verification file containing health status markers for all sensors, serving as the gating basis for data access. The clean vibration data stream is structured data that has been cleaned and freed from temperature interference and fault noise. Its length is typically 120,000 (corresponding to 10,000 sampling points for each of the 12 vibration channels), and it is specifically used to characterize the mechanical motion state of escalators.

[0040] For example, process the sensor status confirmation report of escalator ES_2F_03. If the report generated in the connecting step shows that the handrail bearing seat sensor with SensorID 7 has an "offset fault", the logical judgment result is "no", and data extraction is not performed to ensure that fault data does not enter the diagnostic model. If the status flags of all 12 sensors are confirmed as "normal" in the report of the next acquisition cycle, read the Data array in the corresponding original sensor time series. According to the index table, remove the 10 temperature data points at positions 1, 10002, 20003, 30004, 40005, 50006, 60007, 70008, 100009, and 110010; at the same time, extract the vibration data of 12 intervals, including positions 2 to 10001 (vibration at the motor load end), positions 10003 to 20002 (vibration at the motor non-load end), ..., positions 80009 to 90008 (vibration on the left side of the step rail). These 12 data segments, each containing 10,000 values, are merged sequentially to generate a pure vibration data stream containing 120,000 floating-point numbers.

[0041] In one embodiment of the present invention, step S5 includes the following steps: Based on the motor speed and bearing structural parameters of the escalator, the theoretical rotational frequencies of key moving components in the escalator transmission system are calculated. These theoretical rotational frequencies include at least the inner ring rotational frequency, cage rotational frequency, rolling element rotational frequency, cage rotational frequency relative to the inner ring, rolling element frequency over the inner ring surface, and rolling element frequency over the outer ring surface. The data sampling frequency is divided by each calculated theoretical rotational frequency, and the resulting dimensionless integer values ​​are the characteristic scale parameters. For each set of calculated characteristic scale parameters, the pure vibration data stream is divided into several continuous data segments according to the length defined by the parameter. The maximum value is extracted within each data segment to generate a compressed feature sequence at that scale. The compressed feature sequences generated at all scales are concatenated end-to-end to construct a multi-scale feature fusion vector.

[0042] Specifically, technicians first obtain the mechanical structural parameters of the escalator, including the rated speed of the motor and the bearing model parameters used in key components, such as the motor load end. The bearing model parameters include the rolling element diameter, pitch diameter, pressure angle, and number of rolling elements. Based on these physical parameters, the theoretical rotational frequencies of key motion components in the escalator drive system are calculated, including six characteristic frequencies: inner ring rotation, cage rotation, rolling element rotation, cage rotation relative to the inner ring, rolling element rolling over the inner ring surface, and rolling element rolling over the outer ring surface. Next, the sampling frequency during data acquisition is introduced and divided by the calculated six characteristic frequencies to obtain six corresponding dimensionless integer values. These values ​​are defined as feature scales, representing the sampling window length required to capture the corresponding physical motion cycle in the digital signal. Subsequently, the clean vibration data stream generated in step S4 is called, and these six feature scales are applied sequentially for segmented downsampling processing. For each feature scale, the vibration data stream is divided into several continuous and non-overlapping data slices according to the length defined by that scale, and the data point with the largest value is selected within each slice (max pooling operation) to generate a compressed feature sequence at that scale. Finally, the compressed feature sequences generated at the six scales are concatenated end to end according to the order of feature frequency calculation to construct a one-dimensional numerical array, namely the multi-scale feature fusion vector.

[0043] This step involves physical frequency derivation, sampling window calculation, and signal pooling processing, as shown in the following formulas: Calculation of characteristic frequencies of moving parts: Inner ring rotation frequency (shaft frequency): Cage rotation frequency: Rolling element rotation frequency: Cage frequency over inner circle: Frequency of rolling elements passing the inner ring: Frequency of rolling elements passing the outer ring: Feature scale calculation: Segmented max pooling and fusion: in, Motor speed (unit: r / min); to The calculated theoretical rotational frequencies (in Hz) of each component; The diameter of the rolling element is (mm). The pitch diameter is in mm. Pressure angle (degrees); Number of rolling elements (pieces); Sampling frequency (Hz); For the first The segment length parameter (number of samples, rounded up) corresponding to each feature; The input is a pure vibration data stream; For the first The first scale The maximum value of each data segment; This is the generated multi-scale feature fusion vector.

[0044] Among them, the pure vibration data stream is a one-dimensional floating-point array containing only vibration signals, output from the connecting step. The multi-scale feature fusion vector is a high-dimensional numerical vector, whose physical meaning lies in fusing the time-domain extreme value features of the escalator under the motion cycles of different mechanical components, realizing the feature mapping of data from the pure time domain to the mechanical structure domain.

[0045] For example, for escalator ES_2F_03, its motor speed is read. =1000 r / min, bearing model 6314 (rolling element diameter) , pitch diameter Pressure angle Number of rolling bodies (All are known standard parameters).

[0046] Calculate the characteristic frequencies: =16.67 Hz, =6.41 Hz, =34.16 Hz, =10.26 Hz, =82.08 Hz, =51.28 Hz.

[0047] Combined with sampling frequency =1000 Hz, a set of characteristic scale parameters were calculated: =60, =156, =29, =97, =12, =20 (all values ​​are rounded down).

[0048] Processing pure vibration data streams: use The data stream is segmented, and the maximum value is taken from every 60 points to generate the first sequence. use The data stream is segmented, and the maximum value is taken from every 156 points to generate a second set of sequences; this process is repeated to complete 6 sets of processing at different scales.

[0049] Finally, these 6 sets of sequences are concatenated end to end to generate a multi-scale feature fusion vector of the escalator at the current moment, which is used for subsequent health status determination.

[0050] In one embodiment of the present invention, step S6 includes the following steps: The position encoding in step S6 is calculated using sine and cosine functions and is used to assign temporal position information to the data points in the multi-scale feature fusion vector. The escalator status diagnosis model is built on the Transformer architecture and includes a multi-head self-attention mechanism module, which is used to capture the long-distance dependencies within the multi-scale feature fusion vector and output the classification result of whether the escalator is currently in a normal state or a specific mechanical fault state as an initial judgment of the escalator's health status.

[0051] Specifically, technicians first receive the multi-scale feature fusion vector generated in step S5. This vector needs to be supplemented with sequence position information before being input into the model. A unique temporal index is assigned to each value in the vector using a positional encoding algorithm. The encoding value for even-dimensional positions is calculated using a sine function, and the encoding value for odd-dimensional positions is calculated using a cosine function. The calculated positional encoding vector is then added bit-by-bit to the original multi-scale feature fusion vector to generate an embedding vector with positional information. This embedding vector is then input into a pre-trained escalator status diagnosis model. Inside the model, the data is first standardized through a normalization layer, then enters a multi-head self-attention module to calculate long-distance dependencies within the data and extract the correlation weights between features. The output of the multi-head self-attention mechanism is residually concatenated (i.e., added) with the original input, and after another normalization process, it is input into a multilayer perceptron network containing hidden layers for feature transformation and abstraction. Finally, the transformed feature data is passed through a Softmax classification layer to calculate the probability value of the input data belonging to various preset escalator status categories, such as normal, bearing inner ring wear, bearing outer ring wear, and gear tooth breakage. Select the status label corresponding to the category with the highest probability to generate an initial judgment of the escalator's health status, which includes the escalator health status classification results and corresponding confidence levels.

[0052] This step involves calculating the trigonometric functions for positional encoding and performing matrix operations within the neural network. The formulas are as follows: Location code calculation: Escalator condition diagnosis model calculation process: Explanation of symbol meanings: This indicates the index of the data point within the sequence; Indicates the dimension index of the embedded vector; The total size of the embedded dimension is used as a dimensionless coefficient in the calculation; The input is a multi-scale feature fusion vector; It is a positional encoding vector, which is directly added to the input vector, and the dimensions are considered to be the same (numerical value in the feature space). This is the feature representation after adding location information; Presentation layer normalization operation; This represents a multi-head self-attention mechanism operation; This represents the computation of a multilayer perceptron network; and The output features of the intermediate layer of the model; This is the state classification probability distribution vector output by the model.

[0053] It should be noted that positional encoding is a mathematical method that utilizes the periodicity of trigonometric functions to compensate for the temporal sequence lost by neural networks when processing parallel data. The escalator condition diagnosis model is a deep learning network built on the encoder structure of the Transformer architecture. Its core lies in using a multi-head self-attention mechanism to capture global dependencies in multi-scale feature data. Training data for the escalator condition diagnosis model can be obtained through one or more of the following methods: setting typical faults (such as bearing wear, broken gear teeth, and worn steps and chains) on experimental escalators and collecting data; using publicly available rotating machinery fault datasets for transfer learning or feature adaptation; and generating fault data based on escalator dynamics model simulation. During training, the data needs to be divided in the same proportion as the sensor condition diagnosis model (e.g., 7:1:2). The initial assessment of the escalator's health status includes data records of the classification results of the escalator's mechanical system operating status at a specific moment.

[0054] For example, for escalator ES_2F_03, the length of the processed multi-scale feature fusion vector is 365. (Setting the embedding dimension) =512.

[0055] Calculate the positional encoding: For the first data point in the vector (pos=0), the encoded value of the 0th dimension (j=0) is... =0; For the second data point (pos=1), the encoded value of the 0th dimension is 0.8415. These values ​​are then superimposed onto the original vector.

[0056] Model derivation: The superimposed vectors are input into the model. After internal calculation, the output layer Softmax results show that the probability of the "normal" state is 0.02, the probability of the "wear on the outer ring of the bearing at the motor load end" state is 0.97, and the probability of the "wear on the gearbox gear" state is 0.01.

[0057] Result generation: Based on the maximum probability value, the current status is determined to be "wear on the outer ring of the bearing at the motor load end", and an initial health status judgment of the escalator is generated, including: EscalatorID is ES_2F_03, FaultDescription is "wear on the outer ring of the bearing", FaultPosition is "motor load end", and Timestamp is 1744680313000.

[0058] In one embodiment of the present invention, step S7 includes the following steps: Determine if there are fault markers in the sensor status confirmation report; if there are fault markers, extract the sensor fault information from the report as the comprehensive diagnostic conclusion of the equipment, and stop the assessment of the escalator status; if there are no fault markers, extract the escalator fault information or normal status information from the initial assessment of the escalator health status as the comprehensive diagnostic conclusion of the equipment.

[0059] Specifically, technicians or the system first retrieve and parse the sensor status confirmation report generated in step S3. They check the status bits in the report and execute logical branch operations: if any sensor has an abnormal status marker in the report, the conclusion of this diagnostic cycle is determined to be a sensor system fault. In this case, the output of the escalator status diagnostic model is not called, or it is confirmed that the model has not been executed; instead, the faulty sensor number, location, and fault description information in the report are directly extracted. If the report shows that all sensors are normal, the escalator health status preliminary judgment generated in step S6 is retrieved, and the escalator fault location and fault type description are extracted. Subsequently, the extracted core fault information is encapsulated with the metadata of this diagnostic task. The metadata includes the escalator's unique device number and the data acquisition timestamp. The encapsulated object is defined as the comprehensive equipment diagnostic conclusion. Finally, a connection is established with the database, such as the MongoDB database mentioned in the embodiment, and the corresponding storage set is selected according to the type of conclusion. If it is a sensor fault, it is stored in the sensor status set according to the preset sensor diagnostic result data organization format; if it is an escalator fault or normal, it is stored in the escalator status set according to the preset escalator status diagnostic result data organization format, completing the archiving in document form for querying and calling by the front-end maintenance interface.

[0060] The comprehensive equipment diagnostic conclusion is the final judgment result output within a complete monitoring cycle. It is mutually exclusive; that is, in a single diagnosis, the reliability of the sensor itself is reported first, and mechanical health problems are reported only when the reliability is reliable. The database adopts a NoSQL database structure, storing structured documents in key-value pairs.

[0061] For example, for escalator ES_2F_03, the timestamp is 1744680313000.

[0062] Scenario 1: The sensor status confirmation report is read and confirmed to be normal. The initial health status assessment of the escalator is then read, showing "Wearing on the outer ring of the bearing at the motor load end." A comprehensive equipment diagnostic conclusion is generated, with the following key-value pairs: EscalatorID "ES_2F_03", Timestamp "1744680313000", FaultPosition "Motor load end", and FaultDescription "Wearing outer ring". This record is written to the escalator diagnostic results collection in the database.

[0063] Scenario 2 (Logical Branch Comparison): If the report in step S3 shows a fault in sensor 7 (as shown in the example connecting to step S3), a comprehensive equipment diagnostic conclusion is generated, including: EscalatorID as "ES_2F_03", Timestamp as "1744680313000", SensorID as "7", SensorPosition as "Handrail bearing seat (left side)", and FaultDescription as "Offset fault". This record is written to the sensor diagnostic result set in the database, and this conclusion does not contain escalator mechanical status information, achieving precise isolation of the fault source.

[0064] See appendix Figure 2 The present invention also proposes an escalator fault detection system with data self-verification, comprising the following modules: The data acquisition and sequence generation module is used to acquire sensor data installed on key transmission components of escalators and convert the continuously acquired temperature and vibration analog signals within a specified time period into a raw sensor time series containing the readings and acquisition timestamps of each sensor. The composite input construction module is used to receive the raw sensor time series, perform differential operations on it to extract the dynamic change features of the data, and merge it with the raw data to generate sensor diagnostic composite data. The sensor status diagnosis module is used to input the sensor diagnosis composite data into a pre-trained sensor status diagnosis model. The model determines whether there are any abnormalities in the data caused by sensor malfunctions and generates a sensor status confirmation report. The data cleaning and extraction module is used to parse the sensor status confirmation report. When it is confirmed that all sensors are in normal working condition, the vibration data part of the original sensor time series is extracted to generate a clean vibration data stream. The multi-scale feature fusion module calculates multiple sets of feature scale parameters based on the inherent rotation frequency and data sampling frequency of each moving part of the escalator, and uses these feature scale parameters to segment and compress the pure vibration data stream to generate a multi-scale feature fusion vector. The escalator health diagnosis module is used to add position encoding to the multi-scale feature fusion vector and input the result into the pre-trained escalator status diagnosis model to generate an initial judgment of the escalator health status. The comprehensive diagnostic output module integrates the sensor status confirmation report with the preliminary assessment of the escalator's health status to generate the final comprehensive equipment diagnostic conclusion.

[0065] It should be noted that the formulas described above, through the principle of dimensional consistency and mathematical standardization methods (such as normalization, dimensionless parameter conversion, or unit system unification), can translate physical quantities with different properties into unitless standard values ​​or superimposed parameters of the same dimension. This eliminates the interference of different dimensions on the computational logic, allowing the formulas to retain the original data distribution characteristics while possessing mathematical rationality and adaptability to objective laws. The descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the invention.

[0066] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0067] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for detecting faults in escalators with self-verification of data, characterized in that, Includes the following steps: S1. Collect sensor data installed on key transmission components of escalators, and convert the continuously collected temperature and vibration simulation signals within a specified time period into a raw sensor time series containing the readings and timestamps of each sensor. S2. Receive the raw sensor time series, perform differential operations on it to extract the dynamic change characteristics of the data, and merge it with the raw data to generate sensor diagnostic composite data; S3. Input the sensor diagnostic composite data into the pre-trained sensor status diagnostic model. The model will determine whether there are any abnormalities in the data caused by sensor malfunctions and generate a sensor status confirmation report. S4. Analyze the sensor status confirmation report. When it is confirmed that all sensors are in normal working condition, extract the vibration data portion from the original sensor time series and generate a clean vibration data stream. S5. Based on the inherent rotation frequency and data sampling frequency of each moving part of the escalator, calculate multiple sets of characteristic scale parameters, and use these characteristic scale parameters to segment and compress the pure vibration data stream to generate a multi-scale feature fusion vector. S6. Add position encoding to the multi-scale feature fusion vector and input the result into the pre-trained escalator status diagnosis model to generate an initial judgment of the escalator health status. S7. Integrate the sensor status confirmation report with the preliminary assessment of the escalator's health status to generate the final comprehensive equipment diagnostic conclusion.

2. The escalator fault detection method with data self-verification according to claim 1, characterized in that, Generating sensor diagnostic composite data includes the following steps: Set a relative change scale to define the time interval for calculating dynamic changes in data; Subtract the data point at the current moment in the original sensor time series from the data point that is T moments later on the time axis to obtain the relative change sequence; The relative change sequence is aligned with the original sensor time series according to the timestamps of the data points, and the values ​​at the corresponding positions are added point by point. The merged sequence is then output as the sensor diagnostic composite data.

3. The escalator fault detection method with data self-verification according to claim 1, characterized in that, The sensor condition diagnosis model in step S3 is a classification model built on a neural network, which can identify at least the following sensor fault types: bias fault, drift fault, spike fault, accuracy degradation fault, and jamming fault. The discrimination process includes: dividing the sensor diagnostic composite data into multiple data segments, inputting them into the model to calculate the probability distribution of each fault type; if a fault is identified, marking the faulty sensor number, location, and fault type in the sensor status confirmation report; if no fault is identified, marking the sensor status as normal.

4. The escalator fault detection method with data self-verification according to claim 1, characterized in that, Generating a clean vibration data stream includes the following steps: Check the status markers in the sensor status confirmation report and only perform data extraction if all markers are normal. Based on the preset data storage structure mapping table, the reading indexes corresponding to all temperature sensors are identified and removed from the original sensor time series, while vibration signal segments corresponding to all vibration sensors are retained and extracted. The extracted vibration signal segments are spliced ​​together according to the physical installation order of the sensors on the escalator to form a one-dimensional pure vibration data stream.

5. The escalator fault detection method with data self-verification according to claim 1, characterized in that, Calculate multiple sets of feature scale parameters, including the following steps: Based on the motor speed and bearing structure parameters of the escalator, calculate the theoretical rotational frequency of key moving parts in the escalator transmission system. The theoretical rotational frequency includes at least the inner ring rotational frequency, the cage rotational frequency, the rolling element rotational frequency, the cage rotational frequency relative to the inner ring, the rolling element rolling over the inner ring surface, and the rolling element rolling over the outer ring surface. Divide the data sampling frequency by the calculated theoretical rotation frequency to obtain the dimensionless integer value, which is the characteristic scale parameter.

6. The escalator fault detection method with data self-verification according to claim 5, characterized in that, Generating a multi-scale feature fusion vector includes the following steps: For each set of calculated characteristic scale parameters, the pure vibration data stream is divided into several continuous data segments according to the length defined by the parameter; Extract the maximum value within each data segment to generate a compressed feature sequence at that scale; By concatenating the compressed feature sequences generated at all scales, a multi-scale feature fusion vector is constructed.

7. The escalator fault detection method with data self-verification according to claim 1, characterized in that, The position encoding in step S6 is calculated using sine and cosine functions and is used to assign temporal position information to the data points in the multi-scale feature fusion vector. The escalator status diagnosis model is built on the Transformer architecture and includes a multi-head self-attention mechanism module, which is used to capture the long-distance dependencies within the multi-scale feature fusion vector and output the classification result of whether the escalator is currently in a normal state or a specific mechanical fault state as a preliminary judgment of the escalator's health status.

8. The escalator fault detection method with data self-verification according to claim 1, characterized in that, The process of generating a comprehensive diagnostic conclusion for the equipment includes the following steps: Determine if a fault marker exists in the sensor status confirmation report; If a fault marker exists, the sensor fault information in the report is extracted as the comprehensive diagnostic conclusion for the equipment, and the judgment of the escalator status is suspended. If no fault marker is found, extract the escalator fault information or normal status information from the initial health status assessment of the escalator as the comprehensive diagnostic conclusion for the equipment.

9. The escalator fault detection method with data self-verification according to claim 1, characterized in that, The locations of the sensors collected in step S1 include at least the following: the motor load end housing, the motor non-load end housing, the gearbox input end housing, the gearbox output end housing, the main drive bearing housing, the handrail belt bearing housing, the step rail, and the tension bracket bearing housing of the escalator. Arrange temperature and vibration combined sensors or single-channel vibration sensors at the location, and synchronously collect and generate the original sensor time series at a preset sampling frequency.

10. An escalator fault detection system with data self-verification, characterized in that, Includes the following modules: The data acquisition and sequence generation module is used to acquire sensor data installed on key transmission components of escalators and convert the continuously acquired temperature and vibration analog signals within a specified time period into a raw sensor time series containing the readings and acquisition timestamps of each sensor. The composite input construction module is used to receive the raw sensor time series, perform differential operations on it to extract the dynamic change features of the data, and merge it with the raw data to generate sensor diagnostic composite data. The sensor status diagnosis module is used to input the sensor diagnosis composite data into a pre-trained sensor status diagnosis model. The model determines whether there are any abnormalities in the data caused by sensor malfunctions and generates a sensor status confirmation report. The data cleaning and extraction module is used to parse the sensor status confirmation report. When it is confirmed that all sensors are in normal working condition, the vibration data part of the original sensor time series is extracted to generate a clean vibration data stream. The multi-scale feature fusion module calculates multiple sets of feature scale parameters based on the inherent rotation frequency and data sampling frequency of each moving part of the escalator, and uses these feature scale parameters to segment and compress the pure vibration data stream to generate a multi-scale feature fusion vector. The escalator health diagnosis module is used to add position encoding to the multi-scale feature fusion vector and input the result into the pre-trained escalator status diagnosis model to generate the initial judgment result of the escalator health status. The comprehensive diagnostic output module integrates the sensor status confirmation report with the preliminary assessment of the escalator's health status to generate the final comprehensive equipment diagnostic conclusion.

Images