A data-driven elevator trapping fault on-demand maintenance method

Abstract

The application provides a data-driven elevator trapping fault on-demand maintenance method, and belongs to the technical field of elevator state monitoring and on-demand maintenance; the method first establishes an elevator database for storing original data, cleans and processes the original data in the elevator database, generates multi-dimensional original time series, adopts an empirical mode decomposition algorithm (EMD) to decompose the original time series, and generates a training set and a test set; then, an LSTM prediction model of machine learning is constructed, the LSTM prediction model is trained and tested by using the decomposed time series data, the trained LSTM prediction model is called to predict elevator faults, and corresponding maintenance strategies are formulated according to the prediction results; the application uses a large amount of elevator data as a driving force, learns and predicts trends by using machine learning, performs state monitoring and on-demand maintenance of the elevator according to the prediction results, and effectively improves the efficiency of elevator operation and maintenance and the reliability of elevator operation.

Description

Technical Field

[0001] This invention relates to the field of elevator condition monitoring and on-demand maintenance, and in particular to a data-driven on-demand maintenance method for elevator entrapment faults. Background Technology

[0002] With the rapid development of my country's elevator industry, the country leads the world in both the number of elevators in use and the number of elevators manufactured. Elevators have become an indispensable means of transportation, and elevator safety has become a major concern for all sectors of society. Various malfunctions that occur during normal elevator use can lead to people being trapped, posing a significant threat to people's lives and property.

[0003] With the intelligent development of the elevator industry, artificial intelligence, big data, and the Internet of Things are gradually being integrated into elevators, empowering them from multiple perspectives and improving their safety and reliability. While elevator maintenance has shifted from scheduled maintenance to on-demand maintenance, scheduled maintenance still dominates in practice. Traditional scheduled maintenance can ensure safe elevator operation to a certain extent, but it still has the following shortcomings: 1. The scheduled maintenance model is rigid; applying the same maintenance items to all elevators with different conditions every 15 days is not scientific enough; 2. As the volume of elevator maintenance continues to expand, issues such as non-standardized maintenance and untimely maintenance may arise, thus affecting the safe and stable operation of elevators. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a data-driven method for on-demand maintenance of elevators entrapment faults. This method can realize the transformation from preventive periodic maintenance to on-demand maintenance, improving maintenance efficiency while also enhancing the reliability and safety of elevator operation.

[0005] To achieve the above objectives, this invention provides a data-driven, on-demand maintenance method for elevator entrapment faults, comprising:

[0006] S1: Establish an elevator database to link with the enterprise's IoT database and relevant databases of government regulatory departments, store elevator-related data and update it in real time;

[0007] S2: Collect raw elevator data and upload it to the elevator database;

[0008] S3: Retrieve the raw data from the elevator database, clean and process it to generate the raw time series, and use the empirical mode decomposition algorithm to decompose the raw time series. After decomposition, divide the data into training set and test set.

[0009] S4: Construct an LSTM prediction model, set the parameters for inputting the LSTM prediction model, train the LSTM prediction model using the decomposed training set, and test the LSTM prediction model using the decomposed test set to obtain the accuracy of the prediction results on the test set.

[0010] S5: Optimize the LSTM prediction model by using the cross-entropy loss function and Adam optimizer, and evaluate the performance of the LSTM prediction model by combining the accuracy of the prediction results on the test set. Adjust the structure or training process of the LSTM prediction model based on the evaluation results.

[0011] S6: Use the test set to perform weight analysis on the input parameters of the LSTM prediction model, obtain the weight corresponding to each parameter, adjust the input parameters according to the parameter weights, and repeat the process of steps S3 to S6.

[0012] S7: Input the test set into the prediction program, call the LSTM prediction model in step S4 to predict entrapment, search for elevator fault information in the elevator database and record it based on the entrapment prediction results, and generate an early warning list.

[0013] S8: Verify the generated warning list;

[0014] S9: Develop maintenance strategies based on the early warning list.

[0015] Preferably, the method for processing the raw data in step S3 includes:

[0016] The missing information was filled by consulting data sources, taking the mean, setting zero values, and using interpolation.

[0017] Establish unified standards to standardize information formats;

[0018] Abnormal information is handled by querying relevant data to fill in or set it to zero;

[0019] Text format is converted into discrete encoding through enumeration encoding and one-hot encoding;

[0020] The continuous data is normalized using the following formula:

[0021]

[0022] Where x′ is the processed feature value, x min Let x' be the minimum value of the eigenvalue. max Let x' be the maximum value of the eigenvalue, where x is the eigenvalue before processing.

[0023] Preferably, the steps of the empirical mode decomposition algorithm to decompose the original time series x(t) include:

[0024] S3.1: Solve for all maxima and minima in the sequence, and construct the upper envelope u(t) and lower envelope l(t) based on the maxima and minima;

[0025] S3.2: Calculate the mean m(t) of the upper and lower envelopes. The formula for calculation is:

[0026]

[0027] The mean m(t) of the upper and lower envelopes is removed from the original time series x(t) to obtain the first decomposed signal h1(t), which is given by the following formula:

[0028] h1(t) = x(t) - m(t),

[0029] S3.3: Determine whether the first decomposed signal h1(t) meets the IMF condition; if it does, then the first decomposed signal h1(t) is the first IMF of the original time series x(t), denoted as c1; if it does not meet the condition, then replace the original time series x(t) to be decomposed with the first decomposed signal h1(t); repeat the above steps until there are no terms that meet the IMF condition, and finally obtain n intrinsic variables h1(t), h2(t), ..., h n (t) and a residual variable r n (t), which is:

[0030]

[0031] Wherein, IMF stands for intrinsic mode function.

[0032] Preferably, the cross-entropy loss function Loss is used to optimize the parameters of the STM prediction model, and its optimization formula is as follows:

[0033]

[0034] Where n is the total number of categories, y i Let t be the predicted probability of the i-th category. i It is the i-th element of the real label, indicating whether the sample belongs to the i-th category.

[0035] Preferably, the formula for calculating the accuracy of the test set prediction results is as follows:

[0036]

[0037] Wherein, TP is the number of elevators correctly classified as faulty, FP is the number of elevators incorrectly classified as non-faulty, FN is the number of elevators incorrectly classified as faulty, and TN is the number of elevators correctly classified as non-faulty.

[0038] Preferably, the warning list includes: elevator code, elevator location, maintenance unit, failure probability, and warning level.

[0039] Preferably, the process of verifying the generated warning list is as follows:

[0040] Randomly select elevators whose prediction result is red in the next period from the prediction list, and divide the selected elevators into response elevators and verification elevators with the same number;

[0041] On the same day, maintenance personnel were dispatched to inspect and maintain all responding elevators, record the elevator's operating status and maintenance work, and verify that the elevator used as a control group did not undergo maintenance.

[0042] Preferably, when formulating maintenance strategies, corresponding maintenance measures are taken for elevators based on the response levels formed by the number and distribution of warning levels within a predetermined period in the warning list.

[0043] Preferably, the response level has three levels. If the response level is level one, maintenance work must be carried out immediately in the same week; if the response level is level two, the relevant elevators must be monitored in real time through the Internet of Things; if the response level is level three, maintenance is carried out according to the warning situation of the next cycle.

[0044] Preferably, when the response level is Level 1, the maintenance work includes:

[0045] Check the historical fault data and causes of the elevator entrapment. If the elevator has no history of faults, skip this step.

[0046] Check for any abnormalities in the relevant mechanical parts;

[0047] Check electrical components and safety circuits for any abnormalities.

[0048] The present invention provides a data-driven on-demand maintenance method for elevator entrapment faults, the advantages of which are:

[0049] Based on raw elevator data—including elevator IoT data, basic elevator information data, maintenance data, inspection data, and elevator operating environment data—a multi-dimensional raw time series with time, historical data, and environmental factors is generated after data cleaning. This raw time series is then decomposed into multiple sequence segments using the EMD method. These segments are further divided into test and validation sets to provide data support for the machine learning prediction model, ensuring the accuracy of the prediction results. This invention employs the LSTM prediction model from machine learning, which effectively solves the long-distance dependency problem. The LSTM prediction model is trained and tested using validation and test sets, and the tested LSTM prediction model is optimized, effectively improving the accuracy of the prediction results. The trained LSTM prediction model is then used to predict elevator fault trends, and corresponding maintenance strategies are formulated based on the prediction results, effectively improving the efficiency of elevator operation and maintenance and the reliability of elevator operation. Attached Figure Description

[0050] Figure 1 The flowchart illustrates a data-driven, on-demand maintenance method for elevator entrapment faults provided by this invention. Detailed Implementation

[0051] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0052] like Figure 1 As shown, this invention provides a data-driven on-demand maintenance method for elevator entrapment faults, which includes:

[0053] S1: Establish an elevator database to link with the enterprise's IoT database and relevant databases of government regulatory departments, store relevant elevator data and update it in real time;

[0054] S2: Collect raw elevator data and upload it to the elevator database;

[0055] S3: Retrieve the raw data from the elevator database, clean and process it to generate the raw time series, and use the empirical mode decomposition algorithm to decompose the raw time series. After decomposition, divide the data into training set and test set.

[0056] S4: Construct an LSTM prediction model, set the parameters for inputting the LSTM prediction model, train the LSTM prediction model using the decomposed training set, and test the LSTM prediction model using the decomposed test set to obtain the accuracy of the prediction results on the test set.

[0057] S5: Optimize the LSTM prediction model by using the cross-entropy loss function and Adam optimizer, and evaluate the performance of the LSTM prediction model by combining the accuracy of the prediction results on the test set. Adjust the structure or training process of the LSTM prediction model based on the evaluation results.

[0058] S6: Use the test set to perform weight analysis on the input parameters of the LSTM prediction model, obtain the weight corresponding to each parameter, adjust the input parameters according to the parameter weights, and repeat the process of steps S3 to S6.

[0059] S7: Input the test set into the prediction program, call the LSTM prediction model in step S4 to predict entrapment, search for elevator fault information in the elevator database and record it based on the entrapment prediction results, and generate an early warning list.

[0060] S8: Verify the generated warning list;

[0061] S9: Develop maintenance strategies based on the early warning list.

[0062] Specifically, this invention establishes an elevator database that links enterprise IoT databases and relevant data from government regulatory departments, storing raw elevator data. This opens up information channels between enterprises and government regulatory departments, synchronizing information through real-time updates of the raw data. The raw data in the elevator database is retrieved, cleaned, and processed, removing missing and abnormal data, and standardizing all data formats to generate raw time series with a unified format, providing a data foundation for machine learning model training. Based on the multi-dimensional raw time series data, the Empirical Mode Decomposition (EMD) algorithm is used to decompose it into multiple sequence segments, providing data support for the machine learning prediction model and ensuring the accuracy of the prediction results. The LSTM prediction model from machine learning is used as the prediction model, effectively solving the long-distance dependency problem. The LSTM prediction model is trained and tested using the decomposed raw time series data, ultimately outputting a predictive warning list. Based on the warning list, corresponding elevator maintenance strategies are formulated. Maintenance personnel perform elevator maintenance as needed according to the maintenance strategies, improving the efficiency of predictive maintenance and enhancing the safety and reliability of elevator operation to a certain extent.

[0063] In this embodiment, raw elevator data is collected via the Internet of Things (IoT). This raw elevator data includes: basic equipment data, maintenance data, historical fault and entrapment data, historical inspection and testing data, and operating environment data. By using the raw elevator data as the basis for prediction, an accurate and reliable maintenance strategy can be generated.

[0064] In this embodiment, during step S3, the raw data is cleaned, including: elevator registration code, equipment model, city / district, street, location name, user unit, maintenance unit, manufacturer, manufacturing date, installation unit, modification unit, modification date, rated speed, rated load, and floors; maintenance unit, maintenance personnel information, maintenance type information, maintenance non-compliance items, and maintenance time information; fault occurrence time, number of rescuers, arrival time, rescue time, rescue unit, and fault cause; inspection unit and inspection time; and temperature and humidity information of the elevator operating environment. Seasonal and holiday information is also added. This comprehensive data cleaning ensures the accuracy of the final output maintenance requirements.

[0065] The methods for processing the raw data in step S3 include:

[0066] The missing information was filled by consulting data sources, taking the mean, setting zero values, and using interpolation.

[0067] Establish unified standards to standardize information formats; for example, for elevator location information, uniformly remove building number, building, and unit, and separate the information before and after with ' / ', etc.

[0068] Abnormal information can be handled by filling in or setting zeros in relevant data; for example, removing parentheses and characters within parentheses from the data, and handling prefix garbled characters by setting zeros.

[0069] Text format is converted into discrete encoding through enumeration encoding and one-hot encoding;

[0070] The enumeration encoding process converts text format to discretization encoding as follows: Each unique value in the current field is assigned a unique code. If the current field contains n unique values, then the current field's values ​​are mapped to [1, n] as the new value for the current field. The one-hot encoding process converts each unique value in the current field into a dedicated field with only two values: 0 and 1. If the current field contains n unique values, then the current field is expanded into n independent fields, with only one value among the n fields being 1.

[0071] Continuous data is normalized using the following formula:

[0072]

[0073] Where x′ is the processed feature value, x min Let x' be the minimum value of the eigenvalue. max Let x' be the maximum value of the eigenvalue, where x is the eigenvalue before processing.

[0074] It should be noted that there is no specific order required for the above methods of processing the raw data. The processing order can be combined according to actual needs.

[0075] After processing the raw data, a multidimensional raw time series is generated, incorporating dimensions such as time, historical records, and environmental factors. For this raw time series data, the Empirical Mode Decomposition (EMD) algorithm is used to decompose the raw time series x(t), and training and testing sets are created. The advantages of using EMD lie in its adaptability and intuitiveness; it can decompose complex signals into a series of physically meaningful intrinsic mode functions (IMFs) without requiring predefined basis functions or models. Furthermore, it can reveal the intrinsic characteristics of the sequence, contributing to a better understanding of the composition and dynamic properties of the time series. The decomposition process includes:

[0076] S3.1: Solve for all maxima and minima in the sequence, and construct the upper envelope u(t) and lower envelope l(t) based on the maxima and minima;

[0077] S3.2: Calculate the mean m(t) of the upper and lower envelopes. The formula for calculation is:

[0078]

[0079] The mean m(t) of the upper and lower envelopes is removed from the original time series x(t) to obtain the first decomposed signal h1(t), which is given by the following formula:

[0080] h1(t) = x(t) - m(t)

[0081] S3.3: Determine whether the first decomposed signal h1(t) meets the IMF conditions, specifically: Condition 1: The number of local extrema and zero-crossings must be equal or differ by at most one over the entire time range. Condition 2: At any given time point, the average of the envelopes of the local maxima (upper envelope) and the envelopes of the local minima (lower envelope) must be zero.

[0082] If the conditions are met, the first decomposed signal h1(t) is the first IMF of the original time series x(t), denoted as c1; if not, the first decomposed signal h1(t) replaces the original time series x(t) to be decomposed; the above steps are repeated until no term meets the IMF condition, and finally n intrinsic variables h1(t), h2(t), ..., hn (t) and a residual variable r n (t), which is:

[0083]

[0084] Wherein, IMF stands for intrinsic mode function.

[0085] In this invention, the machine learning model used is an LSTM prediction model, i.e., a Long Short-Term Memory network, as the prediction model for the early warning strategy. The purpose of choosing a Long Short-Term Memory network is that traditional RNNs lack the ability to acquire other feature information, and the residuals returned by traditional recurrent neural networks decrease exponentially with increasing runtime, leading to slow weight updates and a loss of the ability to acquire long-distance dependencies. Long Short-Term Memory (LSTM) is a special type of RNN with certain improvements, primarily designed to solve the gradient vanishing and gradient exploding problems during long sequence training. LSTM improves the model structure by adding three gating units, which is equivalent to increasing the number of unit states. Therefore, by using forget gates and remember gates to control residual information, LSTM effectively solves the long-distance dependency problem.

[0086] During LSTM prediction model training, based on the PyTorch deep learning framework, the decomposed training set is input into the LSTM prediction model for training. The decomposed test set is then used to test the LSTM prediction model, obtaining the accuracy of the prediction results on the test set. Essentially, by testing the LSTM prediction model with the test set, the corresponding accuracy can be obtained. The formula for calculating the accuracy of the prediction results on the test set is:

[0087]

[0088] Wherein, TP is the number of elevators correctly classified as faulty, FP is the number of elevators incorrectly classified as non-faulty, FN is the number of elevators incorrectly classified as faulty, and TN is the number of elevators correctly classified as non-faulty.

[0089] In this embodiment, after the LSTM prediction model has completed the training and testing phases, it is optimized. This invention uses the cross-entropy loss function, the Adam optimizer, and the accuracy of the test set prediction results to optimize the model. The cross-entropy loss function Loss is used to optimize the parameters of the STM prediction model, and its optimization formula is as follows:

[0090]

[0091] Where n is the total number of categories, y i Let t be the predicted probability of the i-th category.i It is the i-th element of the true label, indicating whether the sample belongs to the i-th category. The performance of the LSTM prediction model is evaluated based on the optimization results of the cross-entropy loss function, the accuracy of the Adam optimizer and the prediction results on the test set. The structure or training process of the LSTM prediction model is adjusted according to the evaluation results to ensure the accuracy of the trained LSTM prediction model.

[0092] In optimizing the LSTM prediction model, a test set is used to perform weight analysis on the model input parameters to obtain the weight corresponding to each parameter. The input parameters are then adjusted based on their weights. Currently, as many as 150 input parameters are used. Parameters with weights greater than 0.5% are selected, while those with weights less than 0.5% are removed. The sum of the remaining input parameter weights reaches 87.454%. Overall, the larger the weight percentage of the input parameters, the greater their impact on the probability of trapping a person.

[0093] In this embodiment, after the LSTM prediction model is trained, the prediction sequence is input into the prediction program, which then calls the trained LSTM prediction model for accurate prediction. The prediction period is set to three weeks, with each week as a time unit. The model will predict possible elevator malfunctions over the next three weeks. After the prediction is completed, the program will generate a detailed elevator warning list. This list not only shows the predictions for the first, second, and third weeks, but also includes the elevator code, the elevator's specific location, the maintenance unit responsible, the predicted probability of malfunction, and the corresponding warning level, as shown in Table 1.

[0094]

[0095] Table 1

[0096] After obtaining the warning list, a verification experiment is conducted on the generated warning list content to verify the reliability of the warning list results. The verification process includes:

[0097] Randomly select elevators whose prediction result is red in the next period from the prediction list, and divide the selected elevators into equal numbers of response elevators and verification elevators; for example, randomly select 50 elevators whose prediction result is red in the next week, and divide the 50 elevators into 25 response elevators and 25 verification elevators.

[0098] On the same day, maintenance personnel were dispatched to inspect and maintain all responding elevators, recording their operating status and maintenance procedures. This verified that elevators used as a control group did not undergo maintenance. Throughout this process, all elevators were within their maintenance schedules.

[0099] The experimental results showed that for responsive elevators, the majority of elevators found to have problems during maintenance inspections; for verification elevators, the majority of elevators malfunctioned within a week. These results demonstrate the reliability of the early warning list and the necessity of maintenance.

[0100] In this embodiment, a corresponding maintenance strategy is formulated based on the verified early warning list, and then maintenance personnel are dispatched to inspect and maintain the elevator. When formulating the maintenance strategy, corresponding maintenance measures are implemented for the elevator based on the response levels formed by the number and distribution of early warning levels within a predetermined period in the early warning list. The early warning levels include red, yellow, and green codes, and there are three response levels. If the number of red codes is high within the next three weeks, the response level is determined to be Level 1; if the number of yellow codes is high, the response level is determined to be Level 2; and if the number of green codes is high, the response level is determined to be Level 3. An example diagram of the response levels is shown in Table 2.

[0101]

[0102] Table 2

[0103] If the response level is Level 1, maintenance work must be carried out immediately within the week, and the maintenance work includes:

[0104] First, check the historical fault data and causes of the elevator entrapment, and prioritize checking these causes. If the elevator has no history of faults, skip this step.

[0105] Secondly, check whether the relevant mechanical parts are abnormal; check for severe wear, jamming, etc., such as the elevator door system, the elevator door slider, sill and door lock need to be cleaned to ensure that the door opens and closes smoothly.

[0106] Finally, check for any abnormalities in the electrical components and safety circuits. For example, misalignment of the hall door lock hook and lock contacts, corrosion of the car door lock contacts, elevator overspeed causing the safety brake and speed governor safety circuits to activate, and poor contact of the upper and lower limit switches can all lead to elevator entrapment.

[0107] If the response level is level 2, the relevant elevators need to be monitored in real time via the Internet of Things.

[0108] If the response level is level three, the elevator needs to be monitored and maintenance should be carried out based on the warning situation in the next cycle.

[0109] Based on the established maintenance strategy, maintenance personnel are intelligently dispatched to promptly inspect and maintain elevators, reducing elevator entrapment accidents and improving elevator reliability and safety.

[0110] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. A data-driven, on-demand maintenance method for elevator entrapment faults, characterized in that, include: S1: Establish an elevator database to link with the enterprise's IoT database and relevant databases of government regulatory departments, store relevant elevator data and update it in real time; S2: Collect raw elevator data and upload it to the elevator database; S3: Retrieve the raw data from the elevator database, clean and process it to generate the raw time series, and use the empirical mode decomposition algorithm to decompose the raw time series. After decomposition, divide the data into training set and test set. S4: Construct an LSTM prediction model, set the parameters for inputting the LSTM prediction model, train the LSTM prediction model using the decomposed training set, and test the LSTM prediction model using the decomposed test set to obtain the accuracy of the prediction results on the test set. S5: Optimize the LSTM prediction model by using the cross-entropy loss function and Adam optimizer, and evaluate the performance of the LSTM prediction model by combining the accuracy of the prediction results on the test set. Adjust the structure or training process of the LSTM prediction model based on the evaluation results. S6: Use the test set to perform weight analysis on the input parameters of the LSTM prediction model, obtain the weight corresponding to each parameter, adjust the input parameters according to the parameter weights, and repeat the process of steps S3 to S6. S7: Input the test set into the prediction program, call the LSTM prediction model in step S4 to predict entrapment, search for elevator fault information in the elevator database and record it based on the entrapment prediction results, and generate an early warning list. S8: Verify the generated warning list; S9: Develop maintenance strategies based on the early warning list.

2. The data-driven on-demand maintenance method for elevator entrapment faults according to claim 1, characterized in that, The methods for processing the raw data in step S3 include: The missing information was filled by consulting data sources, taking the mean, setting zero values, and using interpolation. Establish unified standards to standardize information formats; Abnormal information is handled by querying relevant data to fill in or set it to zero; Text format is converted into discrete encoding through enumeration encoding and one-hot encoding; The continuous data is normalized using the following formula: Where x′ is the processed feature value, x m i n Let x' be the minimum value of the eigenvalue. max Let x' be the maximum value of the eigenvalue, where x is the eigenvalue before processing.

3. The data-driven on-demand maintenance method for elevator entrapment faults according to claim 1, characterized in that, The steps of the empirical mode decomposition algorithm to decompose the original time series x(t) include: S3.1: Solve for all maxima and minima in the sequence, and construct the upper envelope u(t) and lower envelope l(t) based on the maxima and minima; S3.2: Calculate the mean m(t) of the upper and lower envelopes. The formula for calculation is: The mean m(t) of the upper and lower envelopes is removed from the original time series x(t) to obtain the first decomposed signal h1(t), which is given by the following formula: h1(t) = x(t) - m(t), S3.3: Determine whether the first decomposed signal h1(t) meets the IMF condition; if it does, then the first decomposed signal h1(t) is the first IMF of the original time series x(t), denoted as c1; if it does not meet the condition, then replace the original time series x(t) to be decomposed with the first decomposed signal h1(t); repeat the above steps until there are no terms that meet the IMF condition, and finally obtain n intrinsic variables h1(t), h2(t), ..., h n (t) and a residual variable r n (t), which is: Wherein, IMF stands for intrinsic mode function.

4. The data-driven on-demand maintenance method for elevator entrapment faults according to claim 1, characterized in that, The cross-entropy loss function Loss is used to optimize the parameters of the STM prediction model, and its optimization formula is as follows: Where n is the total number of categories, y i Let t be the predicted probability of the i-th category. i It is the i-th element of the real label, indicating whether the sample belongs to the i-th category.

5. The data-driven on-demand maintenance method for elevator entrapment faults according to claim 1, characterized in that, The formula for calculating the accuracy of the prediction results on the test set is as follows: Wherein, TP is the number of elevators correctly classified as faulty, FP is the number of elevators incorrectly classified as non-faulty, FN is the number of elevators incorrectly classified as faulty, and TN is the number of elevators correctly classified as non-faulty.

6. The data-driven on-demand maintenance method for elevator entrapment faults according to claim 1, characterized in that, The warning list includes: elevator code, elevator location, maintenance unit, probability of failure, and warning level.

7. The data-driven on-demand maintenance method for elevator entrapment faults according to claim 1, characterized in that, The process of verifying the generated warning list is as follows: Randomly select elevators whose prediction result is red in the next period from the prediction list, and divide the selected elevators into response elevators and verification elevators with the same number; On the same day, maintenance personnel were dispatched to inspect and maintain all responding elevators, record the elevator's operating status and maintenance work, and verify that the elevator used as a control group did not undergo maintenance.

8. The data-driven on-demand maintenance method for elevator entrapment faults according to claim 1, characterized in that, When formulating maintenance strategies, corresponding maintenance measures are taken for elevators based on the response levels formed by the number and distribution of warning levels within the predetermined period in the warning list.

9. The data-driven on-demand maintenance method for elevator entrapment faults according to claim 8, characterized in that, The response level has three levels. If the response level is level one, maintenance work must be carried out immediately within the week. If the response level is level two, the relevant elevators must be monitored in real time through the Internet of Things. If the response level is level three, maintenance will be carried out based on the warning situation in the next cycle.

10. The data-driven on-demand maintenance method for elevator entrapment faults according to claim 9, characterized in that, When the response level is Level 1, the maintenance work includes: Check the historical fault data and causes of the elevator entrapment. If the elevator has no history of faults, skip this step. Check for any abnormalities in the relevant mechanical parts; Check electrical components and safety circuits for any abnormalities.

Images