An lstm-based elevator on-demand maintenance time series prediction method and system

Abstract

The application discloses an elevator on-demand maintenance time sequence prediction method and system based on LSTM, which comprises the following steps: collecting various feature data of the elevator and performing pretreatment; selecting key features of the elevator maintenance from the pretreated data, assigning weight values according to the importance of the features, obtaining real values of the features, and constructing a data set; constructing and training a multivariate LSTM model to capture the mutual relationship and influence between the key features; the input of the model is the time sequence data of each feature, and the output is the prediction of the maintenance state of the next time step; real-time collection of the elevator feature data, input of the pretreated data into the trained multivariate LSTM model, output of the time sequence prediction result, and formulation of the elevator maintenance plan according to the prediction result. The application realizes accurate prediction of the elevator maintenance time sequence by collecting and analyzing multiple elevator features and combining the time sequence prediction capability of the multivariate LSTM model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of elevator on-demand maintenance, in particular to an elevator on-demand maintenance time sequence prediction method and system based on LSTM. BACKGROUND

[0002] Elevator on-demand maintenance is a maintenance mode that arranges maintenance plans flexibly according to the actual use, the running state and the demand of the elevator. The traditional periodic maintenance mode cannot fully consider the actual running state and the demand of the elevator, which may lead to waste of resources and poor maintenance effect. The elevator on-demand maintenance can arrange personalized maintenance according to the actual situation of the elevator, and the longest maintenance cycle can reach 3 months, thereby improving the maintenance effect and resource utilization efficiency.

[0003] There are also some researches on elevator maintenance prediction based on big data analysis. For example, Chinese patent application CN201711168356.6 discloses an elevator on-demand maintenance system and method based on big data analysis, which predicts the fault risk through elevator fault data and performs elevator on-demand maintenance based on fault warning. The method provides an energy-saving and time-saving intelligent maintenance idea, but also has some defects in technology and feasibility. First, the fault risk calculation model proposed in the above patent uses less data: the data of 1 hour before the fault time point is brought into the fault analysis prediction model to obtain the fault risk coefficient within 5 minutes before the fault occurs; second, there are many fault factors involved in the elevator, and the mechanical fault analysis needs to consider multi-factor dependence. The single fault data analysis method will cause the safety deviation of the entire elevator maintenance. Therefore, a method for predicting the maintenance time sequence according to the multi-dimensional maintenance factors of the elevator and the long-term actual running state data is needed to improve the maintenance efficiency and reduce the fault risk. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides an elevator on-demand maintenance time sequence prediction method and system based on LSTM.

[0005] The specific technical solutions are as follows:

[0006] An elevator on-demand maintenance time sequence prediction method based on LSTM includes the following steps:

[0007] S1: Collecting various feature data of the elevator and performing preprocessing;

[0008] S2: Selecting key features of elevator maintenance from the preprocessed data and assigning weight values according to the importance of the features to obtain real values of the features and construct a data set;

[0009] S3: Construct a multivariate LSTM model to capture the mutual relationship and influence between key features; the input of the model is the time series data of each feature, and the output is the prediction of the maintenance state at the next time step;

[0010] S4: Train the multivariate LSTM model with the data set constructed in S2 to obtain a trained multivariate LSTM model;

[0011] S5: Real-time acquisition of elevator feature data, preprocessed by the method of S1, input into the trained multivariate LSTM model, output the time series prediction result and formulate the elevator maintenance plan according to the prediction result.

[0012] Further, in S1, the feature data of the elevator includes: machine room environment, operation device state, drive host vibration, brake state, encoder state.

[0013] Further, in S1, the preprocessing includes: data cleaning, missing value processing, and outlier processing.

[0014] Further, the multivariate LSTM model includes five layers: input layer, LSTM layer one, LSTM layer two, fully connected layer, and output layer.

[0015] The input layer is used to accept time series data of multiple features as input sequence;

[0016] The LSTM layer one includes multiple LSTM units for capturing the basic time sequence relationship between features; the number of LSTM units is between 20 and 100;

[0017] The LSTM layer two includes multiple LSTM units for further extracting deeper time sequence relationship between features; the number of LSTM units is between 20 and 100;

[0018] The fully connected layer is used to extract the potential interaction relationship between input features;

[0019] The output of the fully connected layer is connected to the output layer, and the number of neurons in the output layer is determined according to actual needs, including using one neuron for binary classification, and the basis of binary classification is whether maintenance is needed; or using multiple neurons for multi-classification, and the basis of this classification is maintenance priority.

[0020] Further, in S4, the mean square error loss function and the stochastic gradient descent optimization algorithm are used in the training process of the multivariate LSTM model to optimize the model parameters.

[0021] Further, the elevator maintenance plan includes: maintenance time node, maintenance content, and maintenance priority.

[0022] Furthermore, the maintenance priorities include three categories: high priority, medium priority, and low priority; for high-priority elevator maintenance needs, personnel are arranged to complete the maintenance of the elevator within 24 hours; for medium-priority elevator maintenance needs, personnel are arranged to complete the maintenance of the elevator within one week; for low-priority elevator maintenance needs, personnel are arranged to complete the maintenance of the elevator within two weeks.

[0023] An LSTM-based elevator on-demand maintenance time series prediction system is used to implement the LSTM-based elevator on-demand maintenance time series prediction method, including: a data collection and preprocessing module, an elevator maintenance demand prediction module, a maintenance plan formulation module, and a multivariate LSTM model update module;

[0024] The data collection and preprocessing module is used to collect user elevator feature data in real time and perform preprocessing;

[0025] The elevator maintenance demand prediction module is used to input the data output by the data collection and preprocessing module into the trained multivariate LSTM model to obtain a time series prediction result; the trained multivariate LSTM model is the output of the multivariate LSTM model update module;

[0026] The maintenance plan formulation module is used to formulate an elevator maintenance plan based on the time series prediction results, including maintenance time nodes, maintenance content and maintenance priority, etc.;

[0027] The multivariate LSTM model update module is used to regularly obtain user historical elevator maintenance data from the database, input it into the multivariate LSTM model after preprocessing, optimize and update the model, and output the trained multivariate LSTM model.

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

[0029] (1) The present invention can achieve accurate prediction: by collecting and analyzing multiple elevator features and combining the time series prediction capability of the multivariate LSTM model, accurate prediction of the elevator maintenance time series can be achieved.

[0030] (2) The present invention can realize on-demand maintenance: the maintenance timing is predicted according to the actual operating status of the elevator, avoiding the resource waste and low efficiency of fixed-cycle maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the LSTM-based elevator on-demand maintenance time series prediction method proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be described in detail below based on the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become more apparent. The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0033] like Figure 1 As shown in FIG, a method for predicting elevator maintenance time series on demand based on LSTM includes the following steps:

[0034] S1: Collect various characteristic data of the elevator and perform preprocessing.

[0035] The characteristic data of the elevator include: machine room environment, operating device status, drive host vibration, brake status, encoder status, etc.

[0036] Preprocessing includes: data cleaning, missing value processing, outlier processing, etc. It is used to ensure the quality and integrity of the data.

[0037] In this embodiment, data from a quarterly elevator maintenance form for an elevator company is used as an example to illustrate the present invention. Maintenance items are used as feature data, and the results of each item are categorized as normal (meets requirements, and testing and experimentation function normally), abnormal (non-compliance detected, requiring adjustment or component replacement), or absent (the elevator does not have this item or function). Based on the needs of elevator maintenance, feature data with abnormal results are assigned a feature value of 1, while feature data with normal or absent results are assigned a feature value of 0. In other words, if an item scores 1 or higher, it indicates that maintenance is required for that item.

[0038] S2: Select the key features of elevator maintenance from the preprocessed data and assign weights according to the importance of the features. According to the scoring rules given in S1, multiply the weight value by the score to obtain the true value of the feature (i.e., the label value). The true value set of all features is used to construct the data set.

[0039] In this embodiment, 45 key features of elevator maintenance are extracted, as shown in Table 1 below.

[0040] Table 1 Elevator maintenance feature number table

[0041]

[0042] In Table 1, the No. 1 "machine room, pulley room environment" is determined to be normal, and the requirements are clean, doors and windows are intact, and lighting is normal; the No. 2 "manual emergency operating device state" requires that the device be complete and located at the designated position; the No. 3 "drive main machine state" requires that there be no abnormal vibration and abnormal sound during operation; the No. 5 "brake clearance state" requires that there be no friction between the brake lining and the brake wheel when the brake is opened, and the clearance value meets the requirements of the manufacturing unit; the No. 6 "brake self-monitoring state" specifically refers to the self-monitoring state of the brake as a car accidental movement protection device stop system; the No. 7 "encoder state" requires cleaning and secure installation; the No. 13 "guide shoe oil cup state" specifically refers to whether the oil cup oil amount is appropriate and whether there is leakage; the No. 29 "landing door locking element engagement length" requires not less than 7 mm; the No. 30 "pit environment state" requires that the pit be clean, there be no water seepage or water accumulation, and the lighting be normal; the No. 33 "reducer lubricating oil state" requires that the lubricating oil amount be appropriate and there be no leakage; and the No. 37 "traction sheave groove, suspension device clean, steel wire rope state" requires that there be no serious oiliness, the tension be uniform, and it meet the requirements of the manufacturing unit.

[0043] According to the importance of these key features to elevator maintenance, they are divided into three levels of A, B, and C, with the weight of A-level features being 0.2, B-level features being 0.5, and C-level features being 1, to reflect the importance of the features to the maintenance time sequence prediction. The classification is shown in Table 2.

[0044] Table 2: Classification and weight assignment of elevator maintenance features

[0045]

[0046] S3: Construct a multivariate LSTM model that captures the mutual relationship and influence between the selected elevator maintenance key features; use the time series data of each feature as an input sequence to predict the maintenance state at the next time step.

[0047] When constructing a multivariate LSTM model, the number of network layers and the number of LSTM units (i.e., the number of neurons) in the LSTM layer need to be determined first, because a deeper network and more neurons can better capture the time sequence relationship between features, but also require more computing resources and longer training time. In this embodiment, the multivariate LSTM model sequentially includes five layers: an input layer, an LSTM layer one, an LSTM layer two, a fully connected layer, and an output layer.

[0048] Input layer: used to accept multiple feature time series data as input sequences.

[0049] LSTM layer one: contains multiple LSTM units to capture the basic temporal relationship between features; the number of LSTM units can be set according to requirements, generally between 20 to 100.

[0050] LSTM layer two: also contains multiple LSTM units to further extract deeper temporal relationships between features; the number of LSTM units can be the same as or different from LSTM layer one, and the number of LSTM units is also between 20 to 100.

[0051] Fully connected layer: used to extract potential interaction relationships between input features.

[0052] Output layer: connects the output of the fully connected layer to the output layer, and the number of neurons in the output layer is determined according to actual requirements. For example, a neuron can be used for binary classification (whether maintenance is needed or not), or multiple neurons can be used for multi-classification (maintenance priority).

[0053] S4: Train the multivariate LSTM model with the dataset constructed in S2 to obtain the trained multivariate LSTM model. Through training the model to learn the temporal relationship between features to accurately predict the maintenance needs of the elevator. Specifically, the time series data of each feature is taken as the input sequence, and they are sent into the multivariate LSTM model; in the multivariate LSTM, the sequence of each feature is input into different input channels or feature channels, so that the model can consider the temporal relationship between multiple features at the same time.

[0054] Further, the dataset is divided into a training set and a validation set in a ratio of 7:3. The multivariate LSTM model is trained using the data of the training set, and this process uses appropriate loss functions and optimization algorithms, such as the Mean Squared Error (MSE) loss function and the Stochastic Gradient Descent (SGD) optimization algorithm. During the training process, according to the input features and true values of the training set, the network parameters are updated through the backpropagation algorithm, so that the model gradually learns the temporal relationship between features. The performance and accuracy of the model are evaluated using the validation set, and the model is fine-tuned and verified.

[0055] Among them, the MSE loss function is used to measure the difference between the predicted value of the model and the actual value, and the expression is as follows:

[0056]

[0057] In the formula, n represents the number of training samples, y predy represents the predicted value of the model true y represents the true value of the sample.

[0058] SGD optimization algorithm is a commonly used optimization algorithm for updating the parameters of the model to minimize the loss function, expressed as follows:

[0059]

[0060] where θ represents the parameters of the model, θ' represents the updated model parameters, learning rate represents the learning rate, L represents the loss function, i.e. the MSE loss function. represents the gradient of the loss function with respect to the model parameters.

[0061] Backpropagation algorithm: used to calculate the gradient of the loss function with respect to the model parameters, in order to update the parameters. It passes the gradient layer by layer through the chain rule, and calculates the gradient from the output layer to the input layer in reverse.

[0062] Training process (Training Process): for each training iteration, input the data of the training set into the model for forward propagation, and calculate the predicted value. According to the predicted value and the actual value, the value of the loss function is calculated. Use the backpropagation algorithm to calculate the gradient, and update the parameters of the model according to the optimization algorithm. Repeat the above steps until the preset number of iterations or the loss function converges.

[0063] Validation process (Validation Process): after each training iteration, input the data of the validation set into the model for forward propagation, and calculate the predicted value. According to the predicted value and the actual value of the validation set, the value of the loss function on the validation set is calculated, and the performance of the model is evaluated. If the loss function is less than the threshold value, it is considered that the performance of the model is sufficient, otherwise the training is performed again.

[0064] S5: Real-time acquisition of elevator feature data, and after preprocessing by the method of S1, input into the trained multivariate LSTM model, output the time series prediction result and formulate the elevator maintenance plan according to the prediction result. The elevator maintenance plan includes: maintenance time node, maintenance content, maintenance priority, etc.

[0065] In this embodiment, the multivariate LSTM model will classify the elevator maintenance demand into three categories according to priority: high priority, medium priority and low priority. The high priority elevator maintenance demand will be arranged to be repaired within 24 hours; the medium priority elevator maintenance demand will be arranged to be repaired within a week; the low priority elevator maintenance demand will be arranged to be repaired within two weeks.

[0066] In order to realize the above-mentioned LSTM-based elevator on-demand maintenance time series prediction method, an LSTM-based elevator on-demand maintenance time series prediction system is further provided by the embodiment of the present application, comprising: a data collection and preprocessing module, an elevator maintenance demand prediction module, a maintenance plan making module, and a multivariate LSTM model updating module.

[0067] The data collection and preprocessing module is used for collecting user elevator feature data in real time and performing preprocessing.

[0068] The elevator maintenance demand prediction module is used for inputting the data output by the data collection and preprocessing module into the trained multivariate LSTM model to obtain a time series prediction result; the trained multivariate LSTM model is the output of the multivariate LSTM model updating module.

[0069] The maintenance plan making module is used for making an elevator maintenance plan according to the time series prediction result, including maintenance time nodes, maintenance contents, and maintenance priorities, etc.

[0070] The multivariate LSTM model updating module will obtain user historical elevator maintenance data from a database at regular intervals, input the data into the multivariate LSTM model after preprocessing, and optimize and update the model.

[0071] Those skilled in the art can understand that the above-mentioned is only the preferred examples of the present application, and is not used to limit the present application, although the present application is described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for some technical features. Any modification, equivalent replacement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting elevator maintenance time series on demand based on LSTM, characterized in that: The following steps are involved: S1: Collect various characteristic data of the elevator and perform preprocessing; S2: Select key features of elevator maintenance from the preprocessed data and assign weights based on the importance of the features to obtain the true value of each feature and construct a dataset; S3: Build a multivariate LSTM model to capture the interrelationships and influences between key features. The model takes as input the time series data for each feature and outputs a prediction of the maintenance status at the next time step. The multivariate LSTM model includes five layers: input layer, LSTM layer 1, LSTM layer 2, fully connected layer, and output layer; The input layer is used to accept time series data of multiple features as an input sequence; The LSTM layer 1 includes multiple LSTM units for capturing the basic temporal relationship between features; the number of LSTM units is between 20 and 100; The LSTM layer 2 includes multiple LSTM units for further extracting deeper temporal relationships between features; the number of LSTM units is between 20 and 100; The fully connected layer is used to extract the potential interaction relationship between input features; The output of the fully connected layer is connected to the output layer. The number of neurons in the output layer is determined according to actual needs, including using one neuron for binary classification, where the binary classification is based on whether maintenance is required; or using multiple neurons for multi-classification, where the classification is based on maintenance priority. S4: Use the dataset constructed in S2 to train the multivariate LSTM model to obtain a trained multivariate LSTM model; S5: Collect elevator feature data in real time, preprocess it using the method in S1, input it into the trained multivariate LSTM model, output the time series prediction results, and formulate an elevator maintenance plan based on the prediction results.

2. The LSTM-based elevator on-demand maintenance time series prediction method according to claim 1 is characterized in that: In S1, the characteristic data of the elevator include: machine room environment, operating device status, driving host vibration, brake status, encoder status.

3. The LSTM-based elevator on-demand maintenance time series prediction method according to claim 1 is characterized in that: In S1, the preprocessing includes: data cleaning, missing value processing, and outlier processing.

4. The LSTM-based elevator on-demand maintenance time series prediction method according to claim 1 is characterized in that: In S4, during the training process of the multivariate LSTM model, the mean square error loss function and the stochastic gradient descent optimization algorithm are used to optimize the model parameters.

5. The LSTM-based elevator on-demand maintenance time series prediction method according to claim 1 is characterized in that: The elevator maintenance plan includes: maintenance time nodes, maintenance content, and maintenance priorities.

6. The LSTM-based elevator on-demand maintenance time series prediction method according to claim 5 is characterized in that: The maintenance priorities include three categories: high priority, medium priority, and low priority. For high-priority elevator maintenance needs, personnel will be arranged to complete the maintenance of the elevator within 24 hours; for medium-priority elevator maintenance needs, personnel will be arranged to complete the maintenance of the elevator within one week; for low-priority elevator maintenance needs, personnel will be arranged to complete the maintenance of the elevator within two weeks.

7. An LSTM-based elevator on-demand maintenance time series prediction system, used to implement the LSTM-based elevator on-demand maintenance time series prediction method according to any one of claims 1 to 6, characterized in that: include: Data collection and preprocessing module, elevator maintenance demand prediction module, maintenance plan formulation module, and multivariate LSTM model update module; The data collection and preprocessing module is used to collect user elevator feature data in real time and perform preprocessing; The elevator maintenance demand prediction module is used to input the data output by the data collection and preprocessing module into the trained multivariate LSTM model to obtain a time series prediction result; the trained multivariate LSTM model is the output of the multivariate LSTM model update module; The maintenance plan formulation module is used to formulate an elevator maintenance plan based on the time series prediction results, including maintenance time nodes, maintenance content and maintenance priority, etc.; The multivariate LSTM model update module is used to regularly obtain user historical elevator maintenance data from the database, input it into the multivariate LSTM model after preprocessing, optimize and update the model, and output the trained multivariate LSTM model.

Images