Elevator fault judgment and rescue decision-making system based on autonomous training and real-time diagnosis

Abstract

The invention discloses an elevator fault judgment and rescue decision-making system based on autonomous training and real-time diagnosis, and relates to the technical field of elevator safety monitoring and intelligent rescue, and the elevator fault judgment and rescue decision-making system comprises the following specific steps: step 1, multivariate data fusion; step 2, feature extraction; step 3, a deep training mechanism; 4, performing dynamic risk assessment; according to the method, control signals of an elevator control system, detection signals of a vibration sensor and door motor signals are collected and stored in a historical fault database in an associated mode, integration of multi-path signals can be achieved, and the reliability of follow-up model training and real-time diagnosis is guaranteed; a time sequence convolutional network and a graph neural network are combined for modeling, and a topological relation between sensor time sequence data and a system can be constructed, so that real-time fault prediction can be carried out according to sensor signals; the emergency degree of the fault can be judged through a multi-objective optimization algorithm, and then a corresponding rescue scheme is generated in cooperation with fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of elevator safety monitoring and intelligent rescue, in particular to an elevator fault determination and rescue decision system based on autonomous training and real-time diagnosis. BACKGROUND

[0002] Traditional elevator fault detection technology adopts a monitoring system based on threshold alarm or a manual diagnosis method of regular maintenance. The monitoring system based on threshold alarm uses a single sensor (such as an accelerometer) to analyze isolated signals and triggers an alarm through a pre-set fixed threshold (such as vibration acceleration > 0.5g). There is a high false alarm rate (industry average of 35%), which cannot distinguish between normal start-stop and real faults, and can only detect explicit faults (such as steel wire rope breakage), but cannot predict potential risks (such as progressive failure caused by bearing wear). The manual diagnosis method of regular maintenance relies on routine checks every 15 days, which determines the state through artificial auscultation and visual inspection, and has the problem that complex fault modes (such as occasional stop caused by frequency converter harmonic interference) are difficult to reproduce by artificial experience, and 90% of sudden failures occur within the interval between two maintenance periods (China Special Equipment Inspection Research Institute 2023 data).

[0003] In view of the above, an elevator fault determination and rescue decision system based on autonomous training and real-time diagnosis is designed. SUMMARY

[0004] In order to overcome the above-mentioned shortcomings, an elevator fault determination and rescue decision system based on autonomous training and real-time diagnosis is provided.

[0005] The present application realizes the above-mentioned purpose through the following technical solutions:

[0006] An elevator fault determination and rescue decision system based on autonomous training and real-time diagnosis comprises the following specific steps:

[0007] Step one: multi-element data fusion, collecting control signals of the elevator control system, detection signals of the vibration sensor and door machine signals, and storing them in association with the historical fault database;

[0008] Step two: feature extraction, feature extraction is performed on the multi-source data collected in step one, and finally a feature vector is output;

[0009] Step three: deep training mechanism, time series convolution network and graph neural network are jointly modeled to process sensor time series data and system topology relationship, and corresponding prediction results are output according to the feature vector;

[0010] Step four: dynamic risk assessment, according to the prediction results of step three, corresponding risk assessment values, corresponding risk value intervals and corresponding response actions are made;

[0011] Step 5: Fault diagnosis and rescue decision-making. Fault diagnosis is performed based on the prediction results of step 3, the dynamic risk assessment value of step 4, and the historical fault database, and a rescue decision plan is generated.

[0012] Preferably, the specific steps of step 1 are as follows:

[0013] S1, data acquisition, collecting elevator control system, vibration sensor and door machine signals;

[0014] S2: Time synchronization: The elevator control system's control signal and the vibration sensor's detection signal are synchronized. The elevator control signal (50kHz) is 50 times faster than the vibration signal (1kHz). Time synchronization is performed based on the signal with the highest frequency. The 1kHz vibration signal is interpolated 50 times.

[0015] S3, spatial alignment, synchronizes the dimensions of the door machine signal and the vibration signal. The dimensions of the door machine state (discrete event) and the vibration signal (continuous signal) do not match. Therefore, for the discrete door machine state, trigger it by edge and fill it into the time series.

[0016] S4. Data storage: data is divided into real-time data and historical data. Real-time data is used for subsequent real-time diagnosis and fault judgment. Historical data, that is, real-time data, is stored in the historical fault database after use and is used for model training.

[0017] Preferably, in step S2, the alignment cost function used for time synchronization is as follows:

[0018]

[0019] Where φ is the alignment path function; P(φ) is the path smoothing term, which is used to constrain the mapping φ to be as smooth as possible to avoid drastic jumps or irregular changes; C(φ) is the causal constraint term, which is used to impose order constraints on the mapping φ, C(φ) = max(0, t 原因 -t 结果 ) ensures that the cause event is no later than the result; λ1 and λ2 are regularization parameters used to balance the importance of different regularization terms;

[0020] N represents the total number of data points, i represents the i-th data point, i=1…N;

[0021] x i Represented as the i-th data point in the source data, y φ(i) It is represented as the i-th data point in the target data after the mapping φ, where φ is a mapping function from the source data index to the target data index, that is, φ(i) gives the value of x iIndex of the corresponding target data point;

[0022] ||x i -y φ(i) ||2 represents x i and y φ(i) The Euclidean distance between the two data points.

[0023] As preferred, the specific steps of step two are as follows:

[0024] S5, data preprocessing, signal denoising and outlier rejection are performed on the multi-source data collected in step one;

[0025] S6, feature generation, analyze the type of data signal, extract features according to the signal type, for time series signal features, adopt the method of joint features in time and frequency domain, for discrete time features, adopt the method of time interval statistics and state transition matrix, for spatial topology features, adopt the method of graph node embedding;

[0026] S7, data post-processing, the data after feature extraction is standardized and then dimensionally compressed;

[0027] S8, output feature vector.

[0028] As preferred, the specific working steps of step three are as follows:

[0029] S9, batch data is obtained from the data in step two;

[0030] S10, time series processing is performed through a time series convolution network, local signs are extracted through a convolution network, and long dependencies are captured by a dilated convolution;

[0031] S11, topological analysis is performed through a graph neural network, a lift component relationship graph is constructed, nodes are components, and edges are connection relationships, and then a graph attention mechanism is used for weighted aggregation.

[0032] S12, output prediction and calculate accuracy, according to the feature, output prediction results are obtained, and loss calculation is performed by calculating a loss function, the prediction results are counted into the verification set to calculate the accuracy.

[0033] As preferred, the specific working steps of step five are as follows:

[0034] S13, reference signal input, there are three kinds of reference signals, which are the prediction results of step three, the dynamic risk assessment value of step four and the historical fault database;

[0035] S14, fault judgment, according to the prediction results of step three and the historical fault database, the corresponding fault is matched out;

[0036] S15, rescue scheme generation, according to the dynamic risk assessment value of step four, the fault risk index is judged, and the multi-element data acquisition of step one is judged, the emergency degree of the fault is judged through the multi-objective optimization algorithm, and the corresponding rescue scheme is generated according to the judged fault;

[0037] S16, rescue scheme output, sends the rescue scheme to the background, and executes quickly.

[0038] As preferred, the prediction result of step three includes a device topology map.

[0039] As preferred, the calculation formula of the multi-objective optimization algorithm is as follows:

[0040]

[0041] Wherein, P is a comprehensive evaluation value of the fault, x1 is a response time fault evaluation value, x2 is a device damage cost evaluation value, x3 is a personnel safety coefficient evaluation value, alpha1 is a response time fault adjustment coefficient, alpha2 is a device damage cost adjustment coefficient, and alpha3 is a personnel safety coefficient adjustment coefficient.

[0042] The beneficial effects of the present application are that the elevator fault determination and rescue decision system based on autonomous training and real-time diagnosis has:

[0043] 1, multi-element data fusion, the control signal of the elevator control system, the detection signal of the vibration sensor and the door machine signal are collected and stored in association with the historical fault database, which can realize the integration of multi-path signals and ensure the reliability of subsequent model training and real-time diagnosis;

[0044] 2, time sequence convolution network and graph neural network are used for joint modeling, which can construct sensor time sequence data and system topology relationship, so that real-time fault prediction can be carried out according to sensor signals;

[0045] 3, the emergency degree of the fault can be judged through the multi-objective optimization algorithm, and then the corresponding rescue scheme is generated according to the fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0046] The present application will be described by examples and with reference to the accompanying drawings, in which:

[0047] Figure 1 is a structural schematic diagram of the present application;

[0048] Figure 2 is an elevator component relationship diagram of the present application. DETAILED DESCRIPTION

[0049] The application will be described in further detail below with reference to the drawings. These drawings are simplified schematic diagrams which only show the basic structure of the application in a schematic manner and therefore only show the components which are relevant to the application.

[0050] As Figure 1 shown, an elevator fault determination and rescue decision system based on autonomous training and real-time diagnosis includes the following specific steps:

[0051] Step one: multi-element data fusion, collect the control signals of the elevator control system, the detection signals of the vibration sensor and the door machine signals, and store them in association with the historical fault database;

[0052] Step two: feature extraction, feature extraction is performed on the multi-source data collected in step one, and the feature vector is finally output;

[0053] Step three: deep training mechanism, time convolution network and graph neural network are used to jointly model the sensor time series data and system topology relationship, and the corresponding prediction result is output according to the feature vector;

[0054] Step four: dynamic risk assessment, according to the prediction result of step three, the corresponding risk assessment value is made, the corresponding risk value interval is made, and the corresponding response action is made;

[0055] Step five: fault determination and rescue decision, according to the prediction result of step three, the dynamic risk assessment value of step four and the historical fault database, the fault determination is made, and the rescue decision scheme is generated.

[0056] As a specific embodiment, the specific steps of step one are as follows:

[0057] S1, data acquisition, collecting elevator control system, vibration sensor and door machine signal;

[0058] S2, time synchronization, time synchronization is performed on the control signals of the elevator control system and the detection signals of the vibration sensor, the elevator control signal (50kHz) is 50 times faster than the vibration signal (1kHz), and the time synchronization is performed according to the signal with the highest frequency. The 1kHz vibration signal is processed by 50 times interpolation.

[0059] S3, space alignment, the door machine signal and the vibration signal are dimensionally synchronized, the door machine state (discrete event) and the vibration signal (continuous signal) are not matched in dimension; therefore, for the discrete door machine state, the edge is triggered and filled into the time sequence.

[0060] S4, data storage, the data is divided into real-time data and historical data, the real-time data is used for subsequent real-time diagnosis and fault determination, and the historical data is classified into the historical fault database after the real-time data is used, which is used for model training.

[0061] As a specific embodiment, the alignment cost function used in time synchronization in step S2 is as follows:

[0062]

[0063] Wherein, φ represents the alignment path function; P(φ) represents the path smoothing term, which is used to constrain the mapping φ to be as smooth as possible, avoiding sharp jumps or irregular changes; C(φ) represents the causal constraint term, which is used to impose a sequential constraint on the mapping φ, C(φ) = max(0, t 原因 -t 结果 ) ensures that the cause event is not later than the result; λ1, λ2 represent the regularization parameters, which are used to balance the importance of different regularization terms;

[0064] N represents the total number of data points, i represents the i-th data point, i = 1…N;

[0065] x i represents the i-th data point in the source data, y φ(i) represents the i-th data point in the target data after mapping φ, where φ is a mapping function from the source data index to the target data index, i.e. φ(i) gives the index of the target data point corresponding to x i ;

[0066] ||x i -y φ(i) ||2represents the Euclidean distance between the two data points x i and y φ(i) .

[0067] As a specific embodiment, the specific steps of step two are as follows:

[0068] S5, data preprocessing, signal denoising and outlier rejection are performed on the multi-source data collected in step one;

[0069] S6, feature generation, analyze the signal type of the data, extract features according to the signal type, for time series signal features, use the method of joint time-frequency domain features, for discrete time features, use the method of time interval statistics and state transition matrix, for spatial topology features, use the method of graph node embedding;

[0070] S7, data post-processing, the data after feature extraction is standardized, and then dimension compression is performed;

[0071] S8, output feature vector.

[0072] As a specific embodiment, the specific working steps of step three are as follows:

[0073] S9, obtain batch data from the data in step two;

[0074] S10, perform time series processing through a time series convolution network, extract local signs through a convolution network, and capture long dependencies through a hollow convolution;

[0075] S11, perform topology analysis through a graph neural network, construct an elevator component relationship graph, wherein nodes are components and edges are connection relationships, and then perform graph attention mechanism weighted aggregation.

[0076] S12, output a prediction and calculate an accuracy, output a prediction result according to a feature, and simultaneously perform loss calculation through a loss function, and the prediction result is counted into a verification set to perform accuracy statistics.

[0077] According to Figure 2 As shown in FIG. 1, each node represents a component, and the components include a control cabinet, a safety circuit, a speed reducer, a safety clamp, a traction machine, a guide wheel, a car, and a door machine system.

[0078] According to the node monitoring data, the node faults are divided into the following definitions:

[0079] 1. Ghost node: a component that theoretically exists but has no sensor, and a fault example is hidden cable wear;

[0080] 2. Environmental node: temperature and humidity / environmental problems, and a fault example is a lightning strike causing system abnormalities;

[0081] 3. Compound node: a multi-component coupling part, and a fault example is a comprehensive failure of the guide rail- guide shoe- lubrication system;

[0082] 4. Time series node: dynamic process characteristics, and a fault example is abnormal vibration in the acceleration section;

[0083] 5. Power surge node: simulates the influence of power grid fluctuations on the control system;

[0084] 6. Malicious operation node: records abnormal use patterns (such as long-term door blocking)

[0085] 7. Biological contamination node: considers abnormality caused by small animal intrusion.

[0086] The above fault recognition methods are:

[0087] 1. Basic fault recognition: single-component explicit fault;

[0088] 2. Coupling fault diagnosis: multi-component chain fault;

[0089] 3. Extreme scenario simulation: <5% probability of abnormal situation.

[0090] As a specific embodiment, the specific working steps of step five are as follows:

[0091] S13, reference signal input, the reference signal has three kinds, which are the prediction result of step three, the dynamic risk assessment value of step four and the historical failure database;

[0092] S14, fault judgment, according to the prediction result of step three and the historical failure database, the corresponding fault is matched out;

[0093] S15, rescue scheme generation, according to the dynamic risk assessment value of step four, the fault risk index is judged, and the multi-element data collection of step one is judged, the emergency degree of the fault is judged by the multi-objective optimization algorithm, and the corresponding rescue scheme is generated by matching the judged fault;

[0094] S16, rescue scheme output, send the rescue scheme to the background for quick execution.

[0095] The fault diagnosis level is classified according to a five-level classification system (L0-L4):

[0096] Rank Disposal time requirement Typical failure example L0 Immediate stop ladder (response < 15 s) Safety gear misoperation, falling risk L1 Emergency repair (<4 hours) Traction machine overheating, door lock failure L2 Planned maintenance<br (<24 hours) Guide shoe abnormal wear L3 Observational monitoring (<72 hours) Slight vibration exceeds the standard L4 Routine maintenance Sensor drift

[0097] As a specific embodiment, the prediction result of step three includes the device topology map.

[0098] As a specific embodiment, the calculation formula of the multi-objective optimization algorithm is as follows:

[0099]

[0100] Wherein, P is the fault comprehensive evaluation value, x1 is the response time fault evaluation value, x2 is the equipment damage cost evaluation value, x3 is the personnel safety coefficient evaluation value, alpha1 is the response time fault adjustment coefficient, alpha2 is the equipment damage cost adjustment coefficient, and alpha3 is the personnel safety coefficient adjustment coefficient.

[0101] Wherein, the variable evaluation value corresponds to the following table

[0102] Core decision variable table:

[0103]

[0104] The fault comprehensive evaluation value is calculated by the multi-objective optimization algorithm, the emergency degree of the fault is judged, and the corresponding rescue scheme is generated by matching the judged fault.

[0105] According to the above description, relevant personnel can make various changes and modifications without deviating from the technical idea of the present application. The technical scope of the present application is not limited to the contents of the specification, and must be determined according to the scope of the claims.

Claims

1. An elevator fault determination and rescue decision-making system based on autonomous training and real-time diagnosis, characterized by: The specific steps include: Step 1: Multi-dimensional data fusion: collect the control signal of the elevator control system, the detection signal of the vibration sensor, and the door machine signal, and associate and store them with the historical fault database; Step 2: Feature extraction, extract features from the multi-source data collected in step 1, and finally output feature vectors; Step 3: Deep training mechanism, using a joint modeling of time series convolutional networks and graph neural networks to process the relationship between sensor time series data and system topology, and output corresponding prediction results based on feature vectors; Step 4: Dynamic risk assessment: Based on the prediction results of step 3, a corresponding risk assessment value and risk value range are generated, and corresponding response actions are taken; Step 5: Fault diagnosis and rescue decision-making. Fault diagnosis is performed based on the prediction results of step 3, the dynamic risk assessment value of step 4, and the historical fault database, and a rescue decision plan is generated.

2. The elevator fault determination and rescue decision-making system based on autonomous training and real-time diagnosis according to claim 1 is characterized by: The specific steps of step one are as follows: S1, data acquisition, collecting elevator control system, vibration sensor and door machine signals; S2, time synchronization, time synchronization of the control signal of the elevator control system and the detection signal of the vibration sensor; S3, spatial alignment, dimensional synchronization of door machine signals and vibration signals; S4. Data storage: data is divided into real-time data and historical data. Real-time data is used for subsequent real-time diagnosis and fault judgment. Historical data, that is, real-time data, is stored in the historical fault database after use and is used for model training.

3. The elevator fault determination and rescue decision-making system based on autonomous training and real-time diagnosis according to claim 2 is characterized by: In step S2, the alignment cost function used for time synchronization is as follows: Where φ is the alignment path function; P(φ) is the path smoothing term, which is used to constrain the mapping φ to be as smooth as possible to avoid drastic jumps or irregular changes; C(φ) is the causal constraint term, which is used to impose order constraints on the mapping φ, C(φ) = max(0, t 原因 -t 结果 ) ensures that the cause event is no later than the result; λ1 and λ2 are regularization parameters used to balance the importance of different regularization terms; N represents the total number of data points, i represents the i-th data point, i=1…N; x i Represented as the i-th data point in the source data, y φ(i) It is represented as the i-th data point in the target data after the mapping φ, where φ is a mapping function from the source data index to the target data index, that is, φ(i) gives the value of x i The index of the corresponding target data point; ||x i -y φ(i) ||2 represents x i and y φ(i) The Euclidean distance between these two data points.

4. The elevator fault determination and rescue decision-making system based on autonomous training and real-time diagnosis according to claim 1 is characterized in that: The specific steps of step 2 are as follows: S5, data preprocessing, signal denoising and outlier removal for the multi-source data collected in step 1; S6. Feature generation: Analyze the data signal type and extract features based on the signal type. For time series signal features, use the time-frequency domain joint feature method; for discrete time features, use the time interval statistics and state transition matrix method; for spatial topology features, use the graph node embedding method; S7, data post-processing, after feature extraction, the data is normalized and then dimensionally compressed; S8. Output feature vector.

5. The elevator fault determination and rescue decision-making system based on autonomous training and real-time diagnosis according to claim 1 is characterized by: The specific working steps of step three are as follows: S9, obtaining batch data from the data in step 2; S10, perform temporal processing through the temporal convolutional network, extract local features through the convolutional network, and then capture long-term dependencies through the dilated convolution; S11. Perform topological analysis through graph neural networks to construct an elevator component relationship graph, where nodes represent components and edges represent connections. This is followed by weighted aggregation using a graph attention mechanism. S12. Output the prediction and calculate the accuracy. Output the prediction result according to the feature correspondence. At the same time, calculate the loss by calculating the loss function. The prediction result statistics are entered into the validation set for accuracy statistics.

6. The elevator fault determination and rescue decision-making system based on autonomous training and real-time diagnosis according to claim 1 is characterized by: The specific working steps of step five are as follows: S13, reference signal input, there are three types of reference signals, namely the prediction result of step 3, the dynamic risk assessment value of step 4, and the historical fault database; S14, fault judgment, matching the corresponding fault according to the prediction result of step 3 and the historical fault database; S15. Generate a rescue plan. Based on the dynamic risk assessment value in step 4 and the multi-dimensional data collected in step 1, determine the urgency of the fault using a multi-objective optimization algorithm. Then, based on the determined fault, generate a corresponding rescue plan. S16. Output the rescue plan and send it to the backend for quick execution.

7. The elevator fault determination and rescue decision-making system based on autonomous training and real-time diagnosis according to claim 6 is characterized in that: The prediction result of step three includes a device topology map.

8. The elevator fault determination and rescue decision-making system based on autonomous training and real-time diagnosis according to claim 6 is characterized in that: The calculation formula of the multi-objective optimization algorithm is as follows: Where P is the comprehensive fault assessment value, x1 is the response time fault assessment value, x2 is the equipment damage cost assessment value, x3 is the personnel safety factor assessment value, α1 is the response time fault adjustment coefficient, α2 is the equipment damage cost adjustment coefficient, and α3 is the personnel safety factor adjustment coefficient.

Images