A method for predicting the elevator maintenance cycle based on multi-granularity embedding of endogenous and exogenous variables

Abstract

The present invention provides a method for predicting the elevator maintenance period based on multi-granularity embedding of endogenous and exogenous variables. The steps include: dividing the collected elevator operation parameters into endogenous variables and exogenous variables; processing the endogenous variables by splitting them into multiple non-overlapping time segments and mapping them in the form of patch-level tokens; while the exogenous variables are mapped to a low-dimensional feature space to form their respective series-level tokens. Among them, the present invention introduces a learnable global token to represent the macroscopic state of the entire endogenous variable, promoting information fusion between different granularities; utilizing the powerful capabilities of the Transformer architecture, capturing the complex correlations between endogenous and exogenous variables through the cross-attention mechanism, and further strengthening this connection through the interaction between global and local tokens, enabling the model to more accurately understand the influence of internal and external factors on the elevator maintenance period. Through this method, the prediction accuracy and robustness are improved, showing higher accuracy and reliability in the prediction of the elevator maintenance period.

Description

Technical Field

[0001] The present invention relates to a method for predicting the maintenance cycle of intelligent elevators, and in particular to a method for predicting the maintenance cycle of elevators based on multi-granularity embedding of endogenous and exogenous variables. Background Art

[0002] With the rapid development of smart city and Internet of Things technologies, elevators, as indispensable vertical transportation tools in modern buildings, the safety and reliability of their operation directly affect public safety and quality of life. The occurrence of elevator failures and accidents may not only lead to casualties but also cause significant inconvenience to business and residents' lives. Therefore, the maintenance work of elevators is particularly important. However, there are still many problems in the traditional elevator maintenance management method. First of all, many elevator maintenance systems rely on inspections and repairs at fixed intervals. This "regular inspection" method does not fully consider the actual usage and failure probability of elevators, and it is easy to have the phenomenon of over-maintenance or under-maintenance. Secondly, the traditional maintenance method is usually based on experience or manual prediction, resulting in low accuracy of prediction results, and often unable to detect potential problems in a timely manner. The maintenance response lags, increasing the elevator maintenance cost and outage time. In addition, the existing prediction methods fail to effectively utilize the large amount of data generated during the operation of elevators and cannot fully explore the potential laws behind them, resulting in the prediction accuracy and efficiency not meeting the requirements of modern elevator maintenance needs. In recent years, with the rapid development of big data and artificial intelligence technologies, data-driven intelligent prediction methods have gradually become an effective means to solve this problem. Through in-depth analysis of elevator operation data (such as fault logs, operation status, environmental conditions, etc.), accurate prediction of the elevator maintenance cycle can be achieved, thereby optimizing the maintenance plan, reducing unnecessary maintenance costs, and improving the use efficiency and safety of elevators.

[0003] Although the elevator fault prediction methods based on machine learning and deep learning have achieved some results in recent years, these methods still face many challenges, namely: 1) Most of the existing prediction models adopt the same embedding granularity for endogenous and exogenous variables, ignoring the embedding characteristics of different data, thus introducing noise and affecting the model performance; 2) The existing technologies often use relatively simple feature engineering or linear combination methods to try to fuse these internal and external factors, but this method is difficult to fully reveal the deep interaction between endogenous and exogenous variables, resulting in difficulties in fusing endogenous and exogenous variables. Summary of the Invention

[0004] In order to overcome the deficiencies of the prior art, the present invention provides a method for predicting the maintenance cycle of elevators based on multi-granularity embedding of endogenous and exogenous variables.

[0005] To achieve the above object, a method for predicting the maintenance cycle of elevators based on multi-granularity embedding of endogenous and exogenous variables provided by the present invention specifically includes the following steps:

[0006] S1: Obtain the operating parameters of the elevator, where the operating parameters include: endogenous variables and exogenous variables;

[0007] S2: Divide the endogenous variables into multiple non-overlapping time segments at the patch level granularity , and each segment is mapped to a time token through linear projection , and the overall projection of the endogenous variables is a global token at the series level granularity ;

[0008] S3: Map each exogenous variable to a low-dimensional feature space through linear projection, and each exogenous variable is embedded as a token at the series level granularity , and after splicing, a complete exogenous variable embedding sequence at the series level granularity is formed ;

[0009] S4: Use the self-attention mechanism to capture the temporal dependencies within the endogenous variable patch-level granularity segments , and at the same time, use the cross-attention mechanism to capture the relationship between the patch-level granularity segments and the global token at the series level granularity as well as the interaction relationship between the global token at the series level granularity and the exogenous variable token ; and the interaction relationship between the global token at the series level granularity and the exogenous variable token ;

[0010] S5: Use the cross-attention mechanism to capture the information interaction of the interaction relationship and the relationship , and obtain the fused embedding representation of the endogenous and exogenous variables ;

[0011] S6: Input the fused embedding representation into the standard Transformer architecture, perform information fusion through Encoder-Decoder, and output the predicted value ;

[0012] S7: Input the predicted value into the multi-layer perceptron, and after conversion and activation, output the predicted probability of the Transformer model for the label .

[0013] Preferably, the operation data of the elevator is stratified sampled by using the train_test_split function in the scikit-learn package and divided into a training set, a validation set, and a test set according to a set ratio. The SMOTE method in the imblearn library is used to oversample the training set.

[0014] Preferably, in step S2, obtain the time token and the global token at the series level The specific steps include:

[0015] S2.1: When splitting the endogenous variable Learn a global token at the series level to represent the overall macroscopic information of the endogenous variable. The expression for the block operation of the endogenous variable is:

[0016]

[0017] Where is the sequence length, is the value of the time series at the time step , is the time series of fault data, Perform a block operation on the time series;

[0018] S2.2: Learn a time token for each segment:

[0019]

[0020] Where is the operation of embedding the patched sequence after blocking.

[0021] Preferably, the specific steps for obtaining the time token and the global token at the series level also include:

[0022] S2.3: Use a learnable linear projection to perform a projection on the entire sequence level of the endogenous variable to obtain the global embedding:

[0023]

[0024] Where is the learnable linear projection, is the global embedding at the series level.

[0025] Preferably, the expression for constructing the exogenous variable series-level embedding sequence is:

[0026]

[0027] Among them, is an exogenous variable sequence, represents the th exogenous variable, and the series-level embedding sequence , is the number of exogenous variables, is the variable-level embedding.

[0028] Preferably, in step S4, the method for obtaining the interaction information dependencies , relationships and interaction relationships is as follows: Applying cross-attention to the global tokens at the series level and the time tokens separately at different levels, specifically including:

[0029] Information interaction inside the patch-level embedding: Using the self-attention mechanism for information fusion and interaction among N patch-level insides;

[0030] Information interaction between the patch level and the series level: The global token at the series level aggregates the patch-level information of the entire endogenous variable sequence ;

[0031] Information interaction between the exogenous variable series-level embedding and the global token at the series level: The global token at the series level aggregates the information of the exogenous variable sequence-level embedding ;

[0032] The specific expression is:

[0033]

[0034]

[0035]

[0036] Among them, represents the th layer of TimeXer, is the total number of layers, represents the input of the endogenous variable embedding of the th layer of TimeXer; is the Global token input of layer TimeXer, and: ; represents layer normalization, represents the cross-attention mechanism.

[0037] Preferably, in step S5, obtaining the fused embedding representation of endogenous and exogenous variables specific steps include:

[0038] Taking as the query vector query, and taking as the key vector key and value vector value to establish cross-attention between the two types of variables, and finally interact with the information inside the patch-level embedding Perform a stacking operation to form the final joint information fusion embedding , and its specific implementation formula is as follows:

[0039]

[0040] Among them, represents the cross-attention mechanism, represents layer normalization.

[0041] Preferably, in step S6, obtaining the predicted value The method is: using a fully connected layer to output the fused embedding representation :

[0042]

[0043] Among them, is the standard Transformer model.

[0044] Preferably, the steps of obtaining the predicted probability include:

[0045] S7.1: Input the predicted value into the first fully connected layer of the multi-layer perceptron and map it to the hidden dimension , that is:

[0046] ,

[0047] Among them, represents the output of the first layer, represents the weight matrix of the first layer, represents the bias of the first layer, represents the non-linear activation function, which is used to introduce non-linear features;

[0048] S7.2: Output of the first layer Passed to the subsequent fully connected layer in sequence to gradually learn features, that is:

[0049] ,

[0050] Among them, represents the output of the th layer, represents the output of the th layer, represents the weight matrix of the th layer, represents the bias of the th layer.

[0051] Preferably, the step of obtaining the prediction probability further includes:

[0052] S7.3: Map the features to the output dimension through the fully connected layer, and generate the binary classification probability for each time step through the activation function:

[0053] ,

[0054] Among them, represents the output of the th layer, that is, the output of the penultimate layer, represents the output layer weight matrix, represents the output layer bias;

[0055] The formula of the activation function is:

[0056] ,

[0057] Among them, is the input value, and is applied element-wise to the input matrix.

[0058] An elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables provided by the present invention, compared with the prior art, its beneficial effects are as follows:

[0059] 1. By dividing the elevator operation parameters into endogenous variables (fault variables) and exogenous variables (other variables), for endogenous variables, they are segmented into multiple non-overlapping time segments and mapped into patch-level tokens to capture internal details. For exogenous variables, series-level coarse-grained embedding is adopted, which emphasizes capturing the overall trend and avoids introducing noise due to overly fine-grained.

[0060] 2. The present invention also introduces a learnable global token to represent the macroscopic state of the entire endogenous variable, which cross - attentions with the embedding calculations at different granularities respectively. As a bridge, it captures the complex associations between endogenous and exogenous variables and promotes the information fusion between different granularities.

[0061] 3. The present invention utilizes the powerful capabilities of the Transformer architecture to capture the complex associations between endogenous and exogenous variables through the cross - attention mechanism, and further strengthens this connection through the interaction between global and local tokens, enabling the model to more accurately understand the impact of internal and external factors on the elevator maintenance cycle. Compared with traditional methods, the present invention not only significantly improves the prediction accuracy and robustness, but also can effectively handle the complex dependencies in time series and fuse various types of data, thus showing higher accuracy and reliability in elevator maintenance cycle prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a flowchart of a method for predicting elevator maintenance cycle based on multi - granularity embedding of endogenous and exogenous variables provided by the present invention.

[0063] Figure 2 is a schematic diagram of the original data category information mentioned in the embodiments of the present invention.

[0064] Figure 3 is a schematic diagram of the step information for data cleaning in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] The following uses specific specific examples to illustrate the implementation manners 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 implementation manners, 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, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0066] As Figure 1 shown, a method for predicting elevator maintenance cycle based on multi - granularity embedding of endogenous and exogenous variables provided by the present invention, its specific steps include:

[0067] S1: Obtain the operating parameters of the elevator, and the operating parameters include: endogenous variables and exogenous variables;

[0068] S2: Divide the endogenous variables into multiple non - overlapping time segments at the patch - level granularity , and each segment is mapped to a time token through linear projection , the overall projection of the endogenous variables is a global token at the series level , for subsequent interaction with exogenous variables;

[0069] S3: Map each exogenous variable to a low-dimensional feature space through linear projection, and each exogenous variable is embedded as a token at the series level , and they are concatenated to form a complete exogenous variable embedding sequence at the series level ;

[0070] S4: Use the self-attention mechanism to capture the time-dependent relationships inside the patch-level segments of the endogenous variables , and at the same time use the cross-attention mechanism to capture the relationships between the patch-level segments and the global token at the series level as well as the relationships between the global token at the series level and the exogenous variable tokens , thereby improving the information fusion between endogenous and exogenous variables.;

[0071] S5: Use the cross-attention mechanism to capture the interaction relationships and relationships for information interaction, further strengthening the association between endogenous and exogenous variables, ensuring that the influence of exogenous variables on endogenous variables is effectively transmitted, and obtaining the fused embedding representation of endogenous and exogenous variables ;

[0072] S6: Input the fused embedding representation into the standard Transformer architecture, perform information fusion through Encoder-Decoder, and output the predicted value , realizing the effective fusion of endogenous and exogenous variables and the accurate prediction of time series;

[0073] S7: Input the predicted value into the multi-layer perceptron, and after conversion and activation, output the predicted probability of the Transformer model for the label .

[0074] Specifically, in the present invention, elevator operation parameters are divided into endogenous variables (fault variables) and exogenous variables (other variables). For endogenous variables, internal details are captured by splitting them into multiple non-overlapping time segments and mapping them into patch-level tokens. For exogenous variables, series-level coarse-grained embeddings are used to capture the overall trend, avoiding the introduction of noise due to overly fine-grainedness. The present invention introduces a learnable global token to represent the macroscopic state of the entire endogenous variable, calculates cross-attention with embeddings of different granularities respectively, and as a bridge, captures the complex correlations between endogenous and exogenous variables, promoting information fusion between different granularities. The present invention utilizes the powerful capabilities of the Transformer architecture, captures the complex correlations between endogenous and exogenous variables through the cross-attention mechanism, and further strengthens this connection through the interaction between global and local tokens, enabling the model to more accurately understand the impact of internal and external factors on the elevator maintenance cycle.

[0075] In the present invention, the obtained elevator operation parameters include various data types, such as basic information of the elevator, Internet of Things sensor data, fault and entrapment event data, maintenance records, inspection and testing data, and external environment data such as weather and meteorology. Since these data come from different data sources and devices, problems such as noise, missing values, inconsistencies, and outliers often exist, affecting subsequent analysis and modeling. Therefore, effective data preprocessing must be carried out. First, to ensure the accuracy and uniqueness of the data, duplicate records need to be identified and deleted.

[0076] Specifically, a deduplication method based on timestamps and device identifiers is adopted to ensure that each fault record is only counted once, thereby avoiding incorrect statistics caused by duplicate data and ensuring the accuracy of fault frequency and maintenance response. For the problem of missing data, K-nearest neighbor imputation (KNN Imputation) is adopted to fill in the missing values according to the similarity of the data, so as to reduce the interference of data missing on the analysis results. For obvious incorrect data, such as illogical installation time or device information, outlier identification is automatically performed by setting rules and constraint conditions, and the incorrect data is corrected in combination with manual review. For abnormal data with extreme fluctuations, an outlier detection method based on locally weighted regression (such as LOF) is used to identify and process data points that deviate from the normal pattern. After the data undergoes these cleaning steps, it will finally be normalized, and the data is scaled to the interval from 0 to 1 through the maximum-minimum normalization method, ensuring that different types of data can be compared and analyzed on the same scale. The data processed in this way will be more consistent and accurate, providing a reliable basis for subsequent prediction and decision-making. The relevant data fields after processing are shown in Table 1-4.

[0077]

[0078] Table 1

[0079]

[0080] Table 2

[0081]

[0082]

[0083]

[0084] Table 4

[0085] Based on the sources of elevator operation parameters, they are divided into endogenous variables (i.e., elevator fault variables) and exogenous variables (other variables). Based on the processed elevator operation parameters, the train_test_split function in the scikit-learn package is used for stratified sampling, and the division is carried out according to the ratio of 70% training set, 15% validation set, and 15% test set. After the division is completed, the SMOTE method in the imblearn library is used to oversample the training set to increase the number of minority class samples (i.e., fault class samples). Apply SMOTE oversampling to the training set: oversample the number of positive samples to 1 / 10 of the negative samples, effectively avoiding overfitting and underfitting at the data level. Therefore, the sampling_strategy parameter of the SMOTE class is set to 0.1.

[0086] In step S2, obtain the time token and the global token at the series level The specific steps include:

[0087] S2.1: When splitting the endogenous variable (target variable / elevator fault situation), learn a series-level global token to represent the overall macroscopic information of the endogenous variable. The expression for the block operation of the endogenous variable is:

[0088]

[0089] Among them, is the sequence length, is the value of the time series at the time step is the time series of fault data, Perform a block operation on the time series. In this embodiment, when performing the block operation, the time series is decomposed to the patch-level, is the th patch, and the length of each patch is set to , taking P as a hyperparameter of the model and manually updating it according to the effect during training and validation, then .

[0090] S2.2: Learn a temporal token for each segment:

[0091]

[0092] Among them, is the operation of embedding the segmented patch sequence. Specifically, through a learnable linear projection, project { } into a low-dimensional feature space and add it to its own position embedding encoding (the operation of the standard Transformer, used to enhance the global view), that is, to obtain - with N patch-level token embeddings.

[0093] S2.3: Use a learnable linear projection to perform a full sequence-level projection on the endogenous variable to obtain the global embedding:

[0094]

[0095] Among them, is the learnable linear projection, is the global embedding at the series level.

[0096] In step S3, the specific method for constructing the exogenous variable series-level embedding sequence is as follows:

[0097]

[0098] Among them, is the exogenous variable sequence, represents the th exogenous variable, the series-level embedding sequence , is the number of exogenous variables, is the variable-level embedding. For example, if there are three exogenous variables: temperature, meteorology, and holiday, then . is the variable-level embedding. The specific implementation method is to use a trainable linear projection to project into the sequence-level feature space.

[0099] In step S4, the way to obtain the interaction information dependency , the relationship and the interaction relationship is as follows: For the global token at the series level and time tokens Cross-attention is applied separately at different levels, specifically including:

[0100] Information interaction within patch-level embeddings: Self-attention mechanism is used for information fusion and interaction among N patch-level internals;

[0101] Information interaction between patch-level and series-level: Global tokens at the series level Use cross-attention to aggregate the entire endogenous variable sequence of patch-level information;

[0102] Information interaction between exogenous variable series-level embeddings and global tokens at the series level of: Global tokens at the series level Use cross-attention to aggregate the exogenous variable sequence-level embeddings of information;

[0103] The specific expression is:

[0104]

[0105]

[0106]

[0107] Among them, represents the layer of TimeXer, is the total number of layers, represents the input of the endogenous variable embedding of the layer of TimeXer; is the global token input of the layer of TimeXer, and: ; represents layer normalization, represents the cross-attention mechanism.

[0108] In step S5, the specific steps to obtain the fused embedding representation of endogenous and exogenous variables include:

[0109] Take as the query vector query, take as the key vector key and value vector value to establish cross-attention between the two types of variables, and finally perform a stacking operation with the information interaction within the patch-level embedding to form the final joint information fusion embedding , this embedding incorporates self-attention that fuses internal information and cross-attention of internal and external variables. Its specific implementation formula is as follows:

[0110]

[0111] Among them, represents the cross-attention mechanism, represents layer normalization.

[0112] In step S6, the predicted value is obtained in the following way: Use a fully connected layer to output the fused embedding representation :

[0113]

[0114] Among them, is the standard Transformer model.

[0115] Based on the obtained output predicted value , the output predicted value of the encoder is input into a multi-layer perceptron. The network will undergo non-linear transformation through multiple fully connected layers and activation functions. Finally, through the activation layer, the output of the network is converted into the binary classification probability for each time step. This output is , representing the prediction probability of the model for the label. The specific steps include:

[0116] S7.1: Input into the first fully connected layer of the multi-layer perceptron and map it to the hidden dimension , that is:

[0117] ,

[0118] Among them, represents the output of the first layer, represents the weight matrix of the first layer, represents the bias of the first layer, represents the non-linear activation function, which is used to introduce non-linear features;

[0119] S7.2: Sequentially pass the output of the first layer to the subsequent fully connected layers to gradually learn features, that is:

[0120] ,

[0121] Among them, represents the output of the th layer, represents the output of the th layer, denotes the layer weight matrix, denotes the layer bias;

[0122] S7.3: Map the features to the output dimension through a fully connected layer , and generate the binary classification probability for each time step through the activation function:

[0123] ,

[0124] where denotes the output of the layer, that is, the output of the penultimate layer, denotes the output layer weight matrix, denotes the output layer bias, denotes the classification probability of the labels corresponding to multiple future time steps of the sample.

[0125] The formula of the activation function is:

[0126] ,

[0127] where is the input value, and is applied element-wise to the input matrix.

[0128] In the training phase, calculate the binary cross-entropy loss function between the model output and the true label Use the backpropagation algorithm to update the weight parameters in the network to minimize the error. Use an optimization algorithm to adjust the model's parameters to continuously approximate the true label distribution. Train and validate the model multiple times, adjust the model hyperparameters, and obtain the final prediction model. The model after training and optimization will be used to predict the elevator maintenance cycle. This task is regarded as a multi-label classification problem, and the prediction results will cover multiple future cycles to help formulate a more accurate elevator maintenance plan.

[0129] In this embodiment, the test data set preferably uses 26,190 in-use IoT elevators as the data source, uses the data from 2015 to 2023 as the training set to train the model, and uses the data from July to December 2024 as the test set. The specific experimental verification results are shown in Table 5.

[0130]

[0131] Table 5

[0132] Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

Claims

1. A method for predicting the elevator maintenance cycle based on multi-granularity embedding of endogenous and exogenous variables, characterized in that The specific steps include: S1: Obtain the operating parameters of the elevator, where the operating parameters include endogenous variables and exogenous variables; S2: Split the endogenous variable into multiple non-overlapping time segments at the patch-level granularity, and map each segment to a temporal token through linear projection. The entire endogenous variable is projected into a global token at the series-level granularity; ;​​ S3: Map each exogenous variable to a low-dimensional feature space through linear projection, and each exogenous variable is embedded as a token at the series level granularity , and they are concatenated to form a complete exogenous variable embedding sequence at the series level granularity ; where is the th exogenous variable; S4: Use the self-attention mechanism to capture the patch-level granularity segments of the endogenous variables Internal temporal dependencies , and at the same time use the cross-attention mechanism to capture the segments at the patch-level granularity And the global tokens at the series-level granularity Relationships between And the global tokens at the series-level granularity And the exogenous variable tokens Interaction relationships between ; S5: Capture interaction relationships using a cross-attention mechanism and relationship for information interaction to obtain a fused embedding representation of endogenous and exogenous variables ; S6: Embed the fused representation Input it into the standard Transformer architecture, perform information fusion through Encoder-Decoder, and output the predicted value ; S7: Input the predicted value into the multi-layer perceptron. After conversion and activation, output the predicted probability of the Transformer model for the label .

2. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1, wherein For the operating data of the elevator, stratified sampling is performed using the train_test_split function in the scikit-learn package, and it is divided into a training set, a validation set, and a test set according to a set ratio. The SMOTE method in the imblearn library is used to oversample the training set.

3. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1, characterized in that In the said step S2, obtain the time token and the global token at the series level of granularity The specific steps are as follows: S2.1: When splitting the endogenous variable a series-level global token is learned to represent the overall macro information of the endogenous variable. The expression for the block operation of the endogenous variable is: Among them, is the sequence length, is the value of the time series at time step , is the time series of fault data, perform a chunking operation on the time series; S2.2: Learn a time token for each segment: Among them, is an operation to embed the patched sequence after chunking.

4. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 3, wherein, Obtain the time token and the global token at the series level The specific steps also include: S2.3: Perform a complete sequence-level projection on the endogenous variables using a learnable linear projection to obtain a global embedding: ​ Among them, is a learnable linear projection, is a global embedding at the series level of granularity.

5. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1, characterized in that Construct the embedded sequence at the series level of exogenous variables The expression of which is: Among them, is an exogenous variable sequence, represents the th exogenous variable, and the series-level embedding sequence , is the number of exogenous variables, is the variable-level embedding.

6. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1, characterized in that In the step S4, obtaining the interaction information dependencies , relationships and interaction relationships are obtained by applying cross-attention to the global tokens at the series level granularity and the temporal tokens at the patch level granularity separately at different levels, specifically including: Information interaction inside the patch-level embedding: Self-attention mechanism is used for information fusion and interaction among N patch-level insides; Information interaction between patch level and series level: Global tokens at series level granularity Aggregate the entire sequence of endogenous variables using cross-attention patch-level information of; External variable series-level embedding and global tokens at the series level of granularity Information interaction of: global tokens at the series level of granularity Aggregate the external variable sequence-level embedding using cross-attention of information; The specific expression is: ， Among them, represents the layer of TimeXer, is the total number of layers, is the input of the endogenous variable embedding of the layer of TimeXer, represents the input of the exogenous variable embedding of the layer of TimeXer; is the global token input of the layer of TimeXer, and: ; represents layer normalization, represents the cross-attention mechanism.

7. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1, wherein In the step S5, a fused embedding representation of the endogenous and exogenous variables is obtained. The specific steps include: Take as the query vector query, and take as the key vector key and the value vector value to establish cross-attention between the two types of variables, and finally interact with the information inside the patch-level embedding Perform a stacking operation to form the final combined information fusion embedding , and its specific implementation formula is as follows: Among them, represents the cross-attention mechanism, represents layer normalization.

8. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1, wherein In the step S6, the predicted value is obtained in the following manner: using a fully connected layer to output the fused embedding representation : Among them, is a standard Transformer model.

9. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1, wherein , Obtaining the predicted probability The steps include: S7.1: Input the predicted value into the first fully-connected layer of the multi-layer perceptron and map it to the hidden dimension , that is: ， Among them, represents the output of the first layer, represents the weight matrix of the first layer, represents the bias of the first layer, represents the non-linear activation function, which is used to introduce non-linear features; S7.2: Pass the output of the first layer to the fully connected layers of the subsequent layers in sequence, and gradually learn features, that is: ， Among them, represents the output of the th layer, represents the output of the th layer, represents the weight matrix of the th layer, represents the bias of the th layer.

10. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 9, wherein , obtaining the predicted probability The steps also include: S7.3: Map the features to the output dimension through the fully connected layer , and generate the binary classification probability for each time step through the activation function: ， Among them, represents the output of the layer, that is, the output of the penultimate layer, represents the output layer weight matrix, represents the output layer bias; The formula of the activation function is: ， Among them, is the input value, and is applied element-wise to the input matrix .

Images