Elevator fault prediction method based on combination of proportional risk model and machine learning

Abstract

The invention discloses an elevator fault prediction method based on combination of a proportional risk model and machine learning. The elevator fault prediction method is used for predicting the remaining life of an elevator. The hybrid method comprises the following steps: firstly, analyzing a relationship between life and a covariable by using a proportional risk model to obtain an accumulative risk function, and calculating to obtain a life estimated value; then modeling is carried out on the life estimation value and the fault interval time sequence based on a machine learning algorithm, the remaining life is obtained through prediction, and the method comprises the following steps that a data set is imported, and the data set comprises elevator parameter data and elevator maintenance data; extracting a covariable influencing the life of the elevator, a deletion state and the life of the elevator from the elevator parameter data and the elevator maintenance data; a risk function estimated life is calculated, a sub-model 1-COx proportional risk model (WeibullAFTFitter Weibull accelerated failure model) is used, a covariable which affects the life of the elevator and is extracted from elevator parameter data and elevator maintenance data, a deletion state and the life of the elevator are used as input, and the risk function estimated life is output; extracting a historical maintenance time sequence from the elevator maintenance data; and predicting the remaining life of the elevator, and outputting the predicted remaining life of the elevator by using the output of the sub-model 1 and the historical maintenance time sequence extracted from the elevator maintenance data as input by using a sub-model 2-machine learning model (support vector regression model).

Description

Technical Field

[0001] The present invention is based on elevator fault prediction combining a proportional risk model and machine learning, and belongs to the technical field of elevator fault prediction. Background Art

[0002] With the rapid development of my country's economy, the number of high-rise buildings is increasing, and the demand for elevators is increasing day by day. The "Notice on the Safety Status of Special Equipment in China in 2018" issued by the State Administration for Market Regulation shows that by the end of 2018, the number of elevators in my country had reached 6.2783 million units, and this number will continue to grow. As a special equipment closely related to people's lives, the safety of elevators directly affects the life and property safety of users, so elevator fault prediction is particularly important. Fault prediction is essentially the prediction of the remaining life of the equipment. The remaining life, also known as the remaining useful life (RUL), refers to the length of time from the current time to the end of the service life of the equipment. The remaining life prediction methods can be divided into three categories: the method based on the physics-of-failure (PoF), the data-driven (DD) method and the hybrid method (HM). The method based on the fault mechanism refers to the use of the physical degradation principle of the equipment to build a remaining life degradation model. This type of method is usually used in the remaining life prediction of simple mechanical structures, such as bearings, rotors, etc., but is less used in the life prediction of complex electromechanical equipment. The data-driven method refers to the application of statistical or machine learning models to predict the remaining life of equipment. This type of method is widely used. The hybrid method refers to the comprehensive use of two or more algorithms to build a remaining life prediction model for equipment in steps. It usually includes two steps: the first step is to use the feature extraction algorithm to generate relevant features of equipment life degradation; the second step is to use the prediction algorithm to build a remaining life prediction model for equipment. The hybrid method is mostly used for equipment life prediction based on state-type data. The hybrid method can select appropriate feature extraction algorithms and life prediction algorithms according to the data type and remaining life characteristics of the equipment. It is superior to the fault mechanism-based method and data-driven method in terms of method flexibility and life prediction accuracy.

[0003] Elevator fault prediction is to use the elevator's historical fault records and elevator equipment parameter information to establish an elevator life prediction model and predict the remaining life of the elevator. According to the remaining life, inspection and maintenance can be arranged in advance to eliminate the safety hazards of the elevator. The elevator's historical fault records are event-type data, which is very different from state-type data. The amount of data, data types and information contained in event-type data are far less than state-type data, which increases the difficulty of life prediction. The elevator equipment parameter information includes data such as the elevator age, load, and running speed of the elevator, which affects the life degradation characteristics of the elevator. However, the fault feature extraction algorithms used in the existing hybrid methods are mostly designed for state-type data, and cannot effectively extract life degradation features from event-type data. At the same time, these methods cannot apply equipment parameter data to the life prediction model, so the existing hybrid methods are not suitable for the remaining life prediction of elevators. Summary of the invention

[0004] The present invention discloses an elevator fault prediction method based on a combination of a proportional risk model and machine learning, which is used to predict the remaining life of an elevator. The hybrid method first uses a proportional risk model to analyze the relationship between life and covariates, obtains a cumulative risk function, and calculates a life estimate; then, based on a machine learning algorithm, the life estimate and the failure interval time series are modeled to predict the remaining life. In order to achieve the above purpose, the present invention is implemented by the following technical solutions, and the implementation steps are as follows:

[0005] Step 1: Import elevator parameter data and elevator maintenance data as data sets.

[0006] Step 2: Extract the covariates that affect the elevator lifespan as well as the censoring status and elevator lifespan from the elevator parameter data and elevator maintenance data.

[0007] Step 3: Calculate the estimated life of the risk function. Use the sub-model 1-Cox proportional hazard model (Weibull accelerated failure model), take the covariates affecting the elevator life extracted from the elevator parameter data and elevator maintenance data, as well as the missing status and elevator life as input, and output the estimated life of the risk function.

[0008] Step 4: Extract historical maintenance time series from elevator maintenance data.

[0009] Step 5: Predict the remaining life of the elevator. Use sub-model 2 - machine learning model (support vector regression model), take the output of sub-model 1 and the historical maintenance time series extracted from the elevator maintenance data as input, and output the predicted remaining life of the elevator.

[0010] 1. In the above step 1, elevator parameter data and elevator maintenance data are imported and data preprocessing is performed. The original elevator maintenance data is maintenance information recorded in chronological order and cannot be used directly. For all maintenance data, they are classified according to the equipment number, and the maintenance data with the same number are classified into one category. Then, they are arranged in chronological order to obtain the historical maintenance records of all elevators. Since the original data is manually entered, there are a large number of duplicate records and some record content is missing, so it is necessary to remove outliers. First, remove the duplicate records of a single day in the records. Such records are generally caused by repeated repairs in a short period of time due to the fact that the first repair was not completely completed, or by repeated records by maintenance personnel; second, remove the records of elevators with a total number of repairs less than 5 times. Such elevators are generally elevators with a short service life, good quality, and low frequency of use, or some maintenance records are lost or omitted. In the original data, there are a large number of elevators with only one or two maintenance records, and it is impossible to extract fault records in the past period of time. The maintenance records of such elevators cannot be used for modeling; finally, use the box plot method to remove outliers, that is, records that exceed 1.5 times the interquartile range of the upper quartile or are lower than 1.5 times the interquartile range of the lower quartile. In the original data, there are more data that exceed 1.5 times the interquartile range of the upper quartile, and most of these data come from elevators with low frequency of use.

[0011] 2. In the above step 2, the data required for training the proportional hazard model are covariates and life expectancy. Among them, the covariate X consists of two parts: the first part is the elevator parameter covariate X1=(X 11 ,X 12 ,...,X 1m ), such covariates are related to the equipment parameters and working environment of the elevator. The corresponding regression coefficients represent the influence of different elevator star signs and operating environments on the elevator risk function, including elevator use, elevator design maximum operating speed, elevator design maximum load capacity, elevator age, contract type, region and city level; the second part is the maintenance information covariate X2=(X 21 ,X 22 ,...,X 2n), this type of covariate is related to the fault and maintenance information of the elevator. The corresponding regression coefficient represents the degree of influence of different fault types on the elevator life risk function, including faulty components, fault causes, maintenance operations and cumulative number of faults. The two types of covariates are combined to form the covariate X = (X1, X2). In survival analysis, the life of the equipment can be divided into two types according to the missing state E: one is the life data of the next fault that has occurred during the observation period, and its missing state E = 0; the other is the life data of the next fault that has not occurred during the observation period, and its missing state E = 1. Both types of life data can be obtained from the elevator maintenance data: for the case of missing state E = 0, the life T is the length of time between this fault and the next fault; for the case of missing state E = 1, the life T is the length of time between this fault and the end of the observation.

[0012] 3. In step 3 above, the proportional hazard model is also called the Cox model. It was proposed by Cox in 1972 and was first used in survival analysis in the medical field. It has now been widely used in reliability, economics, sociology and other fields. The basic form of the proportional hazard model is:

[0013]

[0014] where h(t, X) is the risk function, also known as the instantaneous failure rate; X = [X1, X2, ..., X m ] is a covariate; β = [β1, β2, ..., β m ] is the regression coefficient, h0(t) is the benchmark risk function, and its expression is:

[0015]

[0016] By integrating the risk function, we can get the cumulative risk function:

[0017]

[0018] The proportional hazard model can use both the censored state E and the life span T, which is specifically reflected in the likelihood function of the model, and its expression is:

[0019]

[0020] Among them, I is the fault data with censoring state E = 0, and set J is the censoring data set with censoring state E = 1. The estimated value of the parameter can be obtained by using maximum likelihood estimation Using the covariate data X, the censored state data E, and the life data T, a proportional hazard model is constructed, and the cumulative hazard function H(t, X) is obtained. i , the corresponding covariate X iAnd the cumulative risk function H(t, X) can be calculated to get the cumulative risk function value H of the equipment when the next failure occurs i =H(T i , X i ). For all fault records, the cumulative risk average value of the next fault occurrence is calculated according to the above method. This average value is set as the cumulative risk threshold of the equipment. i , cumulative risk threshold And the cumulative risk function H(t,X), construct the equation The solution to this equation That is, the risk function estimates the life of the equipment after the failure.

[0021] 4. In step 4 above, the historical maintenance time series is extracted from the elevator maintenance data. The elevator maintenance data records the relevant information of a maintenance, such as the number of the faulty elevator, the time of maintenance and the cause of the fault. The fault time information is recorded in the form of a time point, and the time format is "year / month / day". From the elevator maintenance data, all the historical maintenance time series of a certain elevator {d0, d1, ..., d n}, where d i represents the time of occurrence of the i-th fault. For a certain fault with a time of d i The time between the next failure and the next failure can be regarded as the remaining life after this failure, that is, y i =d i+1 -d i , we can calculate the time difference between adjacent faults to obtain the historical fault interval time series {y1, y2, ..., y n}, this sequence also represents the historical remaining life sequence of the elevator. The historical fault information of the elevator can reflect the health of the elevator to a certain extent. If an elevator has frequent faults in the past, the interval between its next fault will be relatively short, and vice versa. For each fault, its fault information in the past period of time can reflect the health of the equipment to a certain extent. Therefore, for each fault, the historical fault interval time sequence with a time span of T is extracted forward to correspond to it, that is, for the i-th fault interval time y i , extract the fault interval time series within the time span T forward Where: k i represents the number of historical failures within the time span T, y iThis is the output of the machine learning model. Then, the distribution features are extracted from the fault interval time series within the time span T. The extracted features are common features in time domain analysis and can better describe the distribution characteristics of a time interval series, including the last fault interval (previous), the cumulative number of faults (count), the minimum fault interval (min), the lower quartile (Q1), the median (median), the mean (mean), the upper quartile (Q3), the maximum (max) and the standard deviation (std). These features constitute the second part of the input X of the support vector regression model. Ti =(previous i , count i , min i , Q1, …).

[0022] 5. In step 5 above, the output of step 3 is used as the first input of the machine learning model, i.e., the risk function estimated life of the equipment after the failure The output of step 4 is fed into the second part of the machine learning model X Ti =(previous i , count i , min i , Q1, …), the combination of the two features is the input of the model This gives the model input and output The above operation is performed for the historical maintenance records of all elevators to obtain the modeling data set DataSet = {(X1, y1), (X2, y2), ..., (X N ,y N )}, where X i is the characteristic of a certain fault, y i is the remaining life corresponding to the fault. The “holdout method” is used to randomly split the data set into two mutually exclusive sets: a training set S and a test set T, where: the training set is used to train the model, accounting for 70%; the test set is used to evaluate the performance of the model, accounting for 30%. When training the regression model, the S-Fold Cross Validation method is used to split the training set for learning the parameters in the model. The regression model in sub-model 2 is constructed using four machine learning algorithms: support vector regression, multilayer perceptron, random forest, and decision tree regression. The mean absolute error (MAE), mean squared error (MSE), and model score (Score) are used to evaluate the prediction effect of the model. The model score is obtained based on the prediction error of the model. For the prediction error of each data in the test set, the score S of the data is obtained. i, and then average the scores of all data to get the score of the test set. The lower the score, the smaller the prediction error. If the prediction is ahead or behind by the same number of days, the score of the prediction ahead is lower, that is, it tends to predict ahead of time. The smaller the values ​​of these three indicators, the smaller the average prediction error and the higher the accuracy of the model. The calculation method is as follows:

[0023]

[0024]

[0025] Use the grid search algorithm (Grid Search) for parameter tuning. The grid search algorithm can perform parameter tuning according to the loss function within the set search range to find the parameter value with the minimum loss function. Here, the loss function is defined as the Score above. At the same time, since the mean absolute error and mean square error have no bias for early and delayed predictions, and the model score tends to predict in advance, it is more conducive to arranging maintenance in advance and eliminating potential hidden dangers in fault prediction. Therefore, in the comparison of the four different models of the hybrid model, the comparison of model scores will be the main focus, supplemented by the comparison of the mean absolute error and mean square error. Considering the three comparison results comprehensively, it is believed that the support vector regression model in the hybrid model has the best prediction effect. The support vector regression model formula is as follows:

[0026]

[0027] where α j and is the Lagrange multiplier used to solve optimization problems, The sum function is used to map the input features to a higher dimensional space. Here, the rbf kernel function is determined by grid search: radial basis kernel function (Gaussian kernel), which is used for nonlinear problems. b is the bias term, f(X i ) is the output of the method, i.e., the remaining life of the elevator.

[0028] The beneficial effects of the present invention are:

[0029] (1) A hybrid method is proposed for elevator fault prediction, which can more accurately predict the remaining life of the elevator.

[0030] (2) Combining the proportional risk model and the machine learning model enriches the perspective of elevator failure prediction;

[0031] (3) Combining two different types of data, elevator parameter information and elevator maintenance records, to predict faults has a better prediction effect; BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1This is a flow chart of the elevator failure prediction method based on the combination of proportional risk model and machine learning; DETAILED DESCRIPTION

[0033] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0034] The flowchart of the low-resource text classification method combining graph neural network and prompt tuning is as follows Figure 1 As shown, the steps include:

[0035] Step 1: Import elevator parameter data and elevator maintenance data as data sets.

[0036] Step 2: Extract the covariates that affect the elevator life, as well as the missing status and elevator life from the elevator parameter data and elevator maintenance data.

[0037] Step 3: Calculate the estimated life of the risk function. Use the sub-model 1-Cox proportional hazard model (Weibull accelerated failure model), take the covariates affecting the elevator life extracted from the elevator parameter data and elevator maintenance data, as well as the missing status and elevator life as input, and output the estimated life of the risk function.

[0038] Step 4: Extract historical maintenance time series from elevator maintenance data.

[0039] Step 5: Predict the remaining life of the elevator. Use sub-model 2 - machine learning model (support vector regression model), take the output of sub-model 1 and the historical maintenance time series extracted from the elevator maintenance data as input, and output the predicted remaining life of the elevator.

[0040] 1. In the above step 1, elevator parameter data and elevator maintenance data are imported and data preprocessed. The processing sequence includes the following steps:

[0041] (1) First, remove the duplicate records of a single day in the records. This type of record is usually caused by repeated repairs within a short period of time due to the first repair not being completely completed, or by duplicate records by maintenance personnel;

[0042] (2) Secondly, remove the records of elevators with a total maintenance number of less than 5 times. Such elevators are generally of short service life, good quality, and low frequency of use, or some maintenance records are lost or missed. In the original data, there are a large number of elevators with only one or two maintenance records, and it is impossible to extract fault records in the past period of time. The maintenance records of such elevators cannot be used for modeling;

[0043] (3) Finally, the box plot method was used to remove outliers, i.e., records that were more than 1.5 times the interquartile range of the upper quartile or less than 1.5 times the interquartile range of the lower quartile. In the original data, there were more records that were more than 1.5 times the interquartile range of the upper quartile, and most of these records came from elevators with lower usage frequency.

[0044] 2. In the above step 2, the data required for training the proportional hazard model are covariates and life expectancy. Among them, the covariate X consists of two parts: the first part is the elevator parameter covariate X1=(X 11 ,X 12 ,...,X 1m ), such covariates are related to the equipment parameters and working environment of the elevator. The corresponding regression coefficients represent the influence of different elevator star signs and operating environments on the elevator risk function, including elevator use, elevator design maximum operating speed, elevator design maximum load capacity, elevator age, contract type, region and city level; the second part is the maintenance information covariate X2=(X 21 ,X 22 ,...,X 2n ), this type of covariate is related to the fault and maintenance information of the elevator. The corresponding regression coefficient represents the degree of influence of different fault types on the elevator life risk function, including faulty components, fault causes, maintenance operations and cumulative number of faults. The two types of covariates are combined to form the covariate X = (X1, X2). In survival analysis, the life of the equipment can be divided into two types according to the missing state E: one is the life data of the next fault that has occurred during the observation period, and its missing state E = 0; the other is the life data of the next fault that has not occurred during the observation period, and its missing state E = 1. Both types of life data can be obtained from the elevator maintenance data: for the case of missing state E = 0, the life T is the length of time between this fault and the next fault; for the case of missing state E = 1, the life T is the length of time between this fault and the end of the observation.

[0045] 3. In step 3 above, the proportional hazard model training calculation includes the following steps:

[0046] (1) Use the true life of the equipment after a certain failure T i , the corresponding covariate X i And the cumulative risk function H(t, X) can be calculated to get the cumulative risk function value H of the equipment when the next failure occurs i =H(T i ,X i). The proportional hazard model, also known as the Cox model, was proposed by Cox in 1972. It was first used in survival analysis in the medical field and has now been widely used in many fields such as reliability, economics, and sociology. The basic form of the proportional hazard model:

[0047]

[0048] where h(t,X) is the risk function, also known as the instantaneous failure rate; X = [X1, X2, ..., X m ] is a covariate; β = [β1, β2, ..., β m ] is the regression coefficient, h0(t) is the benchmark risk function, and its expression is:

[0049]

[0050] By integrating the risk function, we can get the cumulative risk function:

[0051]

[0052] The proportional hazard model can use both the censored state E and the life span T, which is specifically reflected in the likelihood function of the model, and its expression is:

[0053]

[0054] Among them, I is the fault data with censoring state E = 0, and set J is the censoring data set with censoring state E = 1. The estimated value of the parameter can be obtained by using maximum likelihood estimation Using covariate data X, censored status data E, and life data T, a proportional hazard model is constructed, and the cumulative hazard function H(t,X) is obtained;

[0055] (2) For all fault records, the cumulative risk average value when the next fault occurs is calculated according to the above method. And set this average value as the cumulative risk threshold of the device;

[0056] (3) Using the covariate X of a certain failure i , cumulative risk threshold And the cumulative risk function H(t,X), construct the equation The solution to this equation That is, the risk function estimates the life of the equipment after the failure.

[0057] 4. In the above step 4, extracting the historical maintenance time series from the elevator maintenance data includes the following steps:

[0058] (1) From the elevator maintenance data, we can extract all the historical maintenance time series of a certain elevator {d0, d1, ..., d n}, where d i represents the time of occurrence of the i-th fault. For a certain fault with a time of d i The time between the next failure and the next failure can be regarded as the remaining life after this failure, that is, y i =d i+1 -d i , calculate the time difference between adjacent faults to obtain the historical fault interval time series {y 1， y2,…,y n}, this sequence also represents the historical remaining life sequence of the elevator;

[0059] (2) For each failure, extract the historical failure interval time series within the time span T corresponding to it, that is, for the i-th failure interval time y i , extract the fault interval time series within the time span T forward Where: k i represents the number of historical failures within the time span T, y i This is the output of the machine learning model;

[0060] (3) Then extract distribution features from the fault interval time series within the time span T. The extracted features are commonly used features in time domain analysis and can better describe the distribution characteristics of a time interval series, including the last fault interval time (previous), cumulative number of faults (count), minimum fault interval (min), lower quartile (Q1), median (median), mean (mean), upper quartile (Q3), maximum (max) and standard deviation (std). These features constitute the second part of the support vector regression model input X Ti =(previous i , count i ,min i , Q1, …).

[0061] 5. In step 5 above, the machine learning modeling training calculation includes the following steps:

[0062] (1) The output of step 3 is used as the first input of the machine learning model, that is, the risk function of the equipment after the failure to estimate the life span

[0063] (2) Input the output of step 4 as the second part of the machine learning model into X Ti =(previous i , count i , min i ,Q1,…), the combination of the two features is the input of the model This gives the model input and output The above operation is performed for the historical maintenance records of all elevators to obtain the modeling data set DataSet = {(X1, y1), (X2, y2), ..., (X N ,y N )}, where X i is the characteristic of a certain fault, y i is the remaining life corresponding to the fault;

[0064] (3) The “holdout method” is used to randomly split the data set into two mutually exclusive sets: training set S and test set T, where the training set is used to train the model, accounting for 70%; the test set is used to evaluate the performance of the model, accounting for 30%. When training the regression model, the S-fold cross validation method is used to split the training set for learning the parameters in the model;

[0065] (4) The regression model in sub-model 2 is constructed using four machine learning algorithms: support vector regression, multilayer perceptron, random forest and decision tree regression. The SVR, MLP Regressor, Random Forest Regressor and Decision Tree Regressor models in the Python scikit-learn library are used for modeling. The mean absolute error (MAE), mean squared error (MSE) and model score (Score) are used to evaluate the prediction effect of the model. The model score is obtained based on the prediction error of the model. For the prediction error of each data in the test set, the score S of the data is obtained. i , and then average the scores of all data to get the score of the test set. The lower the score, the smaller the prediction error. If the prediction is ahead or behind by the same number of days, the score of the prediction ahead is lower, that is, it tends to predict ahead of time. The smaller the values ​​of these three indicators, the smaller the average prediction error and the higher the accuracy of the model. The calculation method is as follows:

[0066]

[0067] (5) Grid Search algorithm is used for parameter tuning. The grid search algorithm can perform parameter tuning according to the loss function within the set search range to find the parameter value with the minimum loss function. Here, the loss function is defined as the score above. At the same time, since the mean absolute error and mean square error have no bias for the advance and lag of prediction, and the model score tends to predict in advance, it is more conducive to arranging maintenance in advance and eliminating potential hidden dangers in fault prediction. Therefore, in the comparison of the four different models of the hybrid model, the comparison of model scores will be the main focus, and the comparison of mean absolute error and mean square error will be supplemented. Considering the three comparison results comprehensively, it is believed that the support vector regression model in the hybrid model has the best prediction effect.

Claims

1. An elevator failure prediction method based on a proportional risk model and machine learning, characterized in that The following steps are involved: Step 1: Import elevator parameter data and elevator maintenance data as data sets. Step 2: Extract the covariates that affect the elevator life, as well as the missing status and elevator life from the elevator parameter data and elevator maintenance data. Step 3 calculates the estimated life of the risk function, using the sub-model 1-Cox proportional hazard model (Weibull accelerated failure model), taking the covariates affecting the elevator life extracted from the elevator parameter data and elevator maintenance data, as well as the missing status and elevator life as input, and outputting the estimated life of the risk function. Step 4: Extract historical maintenance time series from elevator maintenance data. Step 5 predicts the remaining life of the elevator. Use sub-model 2 - machine learning model (support vector regression model), take the output of sub-model 1 and the historical maintenance time series extracted from the elevator maintenance data as input, and output the predicted remaining life of the elevator.

2. The elevator failure prediction method based on the combination of proportional risk model and machine learning according to claim 1 is characterized in that In the step 1, elevator parameter data and elevator maintenance data are imported and data preprocessing is performed. Since the original data is manually entered, there are a large number of repeated records, some record contents are missing, etc., and it is necessary to remove outliers. First, remove the single-day repeated records in the records. Such records are generally due to repeated maintenance in a short period of time due to the first maintenance not being completely repaired, or repeated records by maintenance personnel. Secondly, remove the elevator records with a total maintenance number of less than 5 times. Such elevators are generally elevators with a short service life, good quality, and low frequency of use, or some maintenance records are lost or missed. In the original data, there are a large number of elevators with only one or two maintenance records, and it is impossible to extract fault records in the past period of time. The maintenance records of such elevators cannot be used for modeling; finally, use the box plot method to remove outliers, that is, records that exceed 1.5 times the interquartile range of the upper quartile or are lower than 1.5 times the interquartile range of the lower quartile. In the original data, there are more data that exceed 1.5 times the interquartile range of the upper quartile, and most of these data come from elevators with low frequency of use.

3. The elevator failure prediction method based on the combination of proportional risk model and machine learning according to claim 1 is characterized in that In step 2, the data required for training the proportional hazard model are covariates and life expectancy. The covariate X consists of two parts: the first part is the elevator parameter covariate X1 extracted from the elevator parameter data = (X 11 , X 12 ,…,X 1m ), such covariates are related to the equipment parameters and working environment of the elevator. The corresponding regression coefficients represent the influence of different elevator star signs and operating environments on the elevator risk function, including elevator use, elevator design maximum operating speed, elevator design maximum load capacity, elevator age, contract type, region and city level; the second part is the maintenance information covariate X2=(X 21 ,X 22 ,...,X 2n ), this type of covariate is related to the fault and maintenance information of the elevator. The corresponding regression coefficient represents the degree of influence of different fault types on the elevator life risk function, including faulty components, fault causes, maintenance operations and cumulative number of faults. The two types of covariates are combined to form the covariate X = (X1, X2). In survival analysis, the life of the equipment can be divided into two types according to the missing state E: one is the life data of the next fault that has occurred during the observation period, and its missing state E = 0; the other is the life data of the next fault that has not occurred during the observation period, and its missing state E = 1. Both types of life data can be obtained from the elevator maintenance data: for the case of missing state E = 0, the life T is the length of time between this fault and the next fault; for the case of missing state E = 1, the life T is the length of time between this fault and the end of the observation.

4. The elevator failure prediction method based on the combination of proportional risk model and machine learning according to claim 1 is characterized in that In step 3. The proportional hazard model, also known as the Cox model, was proposed by Cox in 1972. It was first used in survival analysis in the medical field and has now been widely used in reliability, economics, sociology and other fields. The basic form of the proportional hazard model is: where h(t, X) is the risk function, also known as the instantaneous failure rate; X = [X1, X2, ..., X m ] is a covariate; β = [β1, β2, ..., β m ] is the regression coefficient, h0(t) is the benchmark risk function, and its expression is: By integrating the risk function, we can get the cumulative risk function: The proportional hazard model can use both the censored state E and the life span T, which is specifically reflected in the likelihood function of the model, and its expression is: Among them, I is the fault data with censoring state E = 0, and set J is the censoring data set with censoring state E = 1. The estimated value of the parameter can be obtained by using maximum likelihood estimation Using the covariate data X, the censored state data E, and the life data T, a proportional hazard model is constructed, and the cumulative hazard function H(t, X) is obtained. i , the corresponding covariate X i And the cumulative risk function H(t,X) can be used to calculate the cumulative risk function value H of the equipment when the next failure occurs i =H(T i , X i ). For all fault records, the cumulative risk average value of the next fault occurrence is calculated according to the above method. This average value is set as the cumulative risk threshold of the equipment. i , cumulative risk threshold And the cumulative risk function H(t,X), construct the equation The solution to this equation That is, the risk function estimates the life of the equipment after the failure.

5. The elevator failure prediction method based on the combination of proportional risk model and machine learning according to claim 1 is characterized in that In the step 4, a historical maintenance time series is extracted from the elevator maintenance data. Elevator maintenance data records the relevant information of a maintenance. The fault time information is recorded in the form of time points, and the time format is "year / month / day". From the elevator maintenance data, all the historical maintenance time series of a certain elevator {d0, d1, ..., d n }, where d i represents the time of occurrence of the i-th fault. For a certain fault with a time of d i The time between the next failure and the next failure can be regarded as the remaining life after this failure, that is, y i =d i+1 -d i , we can calculate the time difference between adjacent faults to obtain the historical fault interval time series {y1, y2, ..., y n }, this sequence also represents the historical remaining life sequence of the elevator. The historical fault information of the elevator can reflect the health of the elevator to a certain extent. If an elevator has frequent faults in the past, the interval between its next fault will be relatively short, and vice versa. For each fault, its fault information in the past period of time can reflect the health of the equipment to a certain extent. Therefore, for each fault, the historical fault interval time sequence with a time span of T is extracted forward to correspond to it, that is, for the i-th fault interval time y i , extract the fault interval time series within the time span T forward Where: k i represents the number of historical failures within the time span T, y i This is the output of the machine learning model. Then, the distribution features are extracted from the fault interval time series within the time span T. The extracted features are common features in time domain analysis and can better describe the distribution characteristics of a time interval series, including the last fault interval (previous), the cumulative number of faults (count), the minimum fault interval (min), the lower quartile (Q1), the median (median), the mean (mean), the upper quartile (Q3), the maximum (max) and the standard deviation (std). These features constitute the second part of the input X of the support vector regression model. Ti =(previous i ,count i , min i , Q1, …).

6. The elevator failure prediction method based on proportional risk model and machine learning according to claim 1 is characterized in that In step 5, the output of step 3 is used as the first part of the input of the machine learning model, that is, the risk function of the equipment after the failure is estimated to be the life span. The output of step 4 is fed into the second part of the machine learning model X Ti =(previous i , count i , min i , Q1, ...), the combination of the two features is the input of the model This gives the model input and output The above operation is performed for the historical maintenance records of all elevators to obtain the modeling data set DataSet = {(X1, y1), (X2, y2), ..., (XN, y N )}, where X i is the characteristic of a certain fault, y i is the remaining life corresponding to the fault. The "holdout method" is used to randomly split the data set into two mutually exclusive sets: training set S and test set T, where: the training set is used to train the model, accounting for 70%; the test set is used to evaluate the performance of the model, accounting for 30%. When training the regression model, the S-Fold Cross Validation method is used to split the training set for learning the parameters in the model. The regression model in sub-model 2 is constructed using four machine learning algorithms: support vector regression, multilayer perceptron, random forest, and decision tree regression. The mean absolute error (MAE), mean square error (MSE), and model score (Score) are used to evaluate the prediction effect of the model. The model score is obtained based on the prediction error of the model. For the prediction error of each data in the test set, the score S of the data is obtained. i , and then average the scores of all data to get the score of the test set. The lower the score, the smaller the prediction error. If the prediction is ahead or behind by the same number of days, the score of the prediction ahead is lower, that is, it tends to predict ahead of time. The smaller the values ​​of these three indicators, the smaller the average prediction error and the higher the accuracy of the model. The calculation method is as follows: The grid search algorithm is used for parameter tuning. The grid search algorithm can perform parameter tuning according to the loss function within the set search range to find the parameter value with the minimum loss function. Here, the loss function is defined as the score above. Since the mean absolute error and mean square error have no bias for the advance and lag of prediction, and the model score tends to predict in advance, it is more conducive to arranging maintenance in advance and eliminating potential hidden dangers in fault prediction. Therefore, in the comparison of the four different models of the hybrid model, the comparison of model scores will be the main focus, supplemented by the comparison of the mean absolute error and mean square error. Considering the three comparison results comprehensively, it is believed that the support vector regression model in the hybrid model has the best prediction effect.

Images