Elevator maintenance period prediction method based on internally and externally generated variable multi-granularity embedding

Abstract

The invention provides an elevator maintenance period prediction method based on endogenous and exogenous variable multi-granularity embedding. The elevator maintenance period prediction method comprises the steps that collected elevator operation parameters are divided into endogenous variables and exogenous variables; dividing the endogenous variable into a plurality of non-overlapping time slices, and mapping the time slices into patch-level tokens for processing; and the exogenous variables are mapped to the low-dimensional feature space to form respective series-level tokens. According to the method, a learnable global token is introduced to represent the macroscopic state of the whole endogenous variable, and information fusion among different granularities is promoted; according to the method, the strong capacity of a Transform architecture is utilized, complex association between endogenous and exogenous variables is captured through a cross attention mechanism, and the association is further strengthened through interaction between global and local tokens, so that the model can more accurately understand the influence of internal and external factors on the elevator maintenance period. Through the method, prediction precision and robustness are improved, and higher accuracy and reliability are shown in elevator maintenance period prediction.

Description

Technical Field

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

[0002] With the rapid development of smart cities and Internet of Things technologies, elevators, as an indispensable vertical transportation tool in modern buildings, have a direct impact on public safety and quality of life due to their safety and reliability. Elevator failures and accidents may not only cause casualties, but also cause significant inconvenience to commercial and residential life. Therefore, elevator maintenance is particularly important. However, there are still many problems with traditional elevator maintenance management methods. First, many elevator maintenance systems rely on fixed-cycle inspections and repairs. This "regular inspection" method does not fully consider the actual use of the elevator and the probability of failure, and is prone to excessive or insufficient maintenance. Secondly, traditional maintenance methods are usually based on experience or human predictions, resulting in low accuracy of prediction results, and often fail to detect potential problems in a timely manner, resulting in delayed maintenance responses, and increasing elevator maintenance costs and downtime. In addition, existing prediction methods fail to effectively utilize the large amount of data generated during the operation of the elevator, and fail to fully explore the potential laws behind it, resulting in the accuracy and efficiency of the prediction failing to meet 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, operating status, environmental conditions, etc.), accurate prediction of elevator maintenance cycles can be achieved, thereby optimizing maintenance plans, reducing unnecessary maintenance costs, and improving elevator efficiency and safety.

[0003] Although 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 existing prediction models adopt the same embedding granularity for endogenous and exogenous variables, but ignore the embedding characteristics of different data, thereby introducing noise and affecting model performance; 2) Existing technologies often use relatively simple feature engineering or linear combination methods to try to integrate 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 the integration of endogenous and exogenous variables. Summary of the invention

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

[0005] In order to achieve the above object, the present invention provides a method for predicting elevator maintenance cycle based on multi-granularity embedding of endogenous and exogenous variables, and the specific steps include:

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

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

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

[0009] S4: Use self-attention mechanism to capture endogenous variable patch-level granularity fragments Internal time dependencies , while capturing the fragments of patch-level granularity through the cross-attention mechanism Global tokens with series-level granularity The relationship between And global tokens at series level granularity With exogenous variable token The interaction between ;

[0010] S5: Using cross-attention mechanism to capture interactive relationships and relationship information interaction to obtain the fusion embedding representation of endogenous and exogenous variables ;

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

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

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

[0014] Preferably, in step S2, a time token is obtained and global tokens at series level granularity The specific steps include:

[0015] S2.1: Endogenous variables When performing segmentation, a series-level global toke is learned to represent the overall macro information of the endogenous variables. The expression of the endogenous variable block operation is:

[0016]

[0017] in, is the sequence length, For a time series at time step The value at is the time series of failure data, Perform block operations on time series;

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

[0019]

[0020] in, The operation of embedding the divided patch sequence.

[0021] Preferably, obtain a time token and global tokens at series level granularity The specific steps also include:

[0022] S2.3: Using learnable linear projections for endogenous variables Do a full sequence-level projection to get a global embedding:

[0023]

[0024] in, is a learnable linear projection, It is a global embedding at series level granularity.

[0025] Preferably, construct an exogenous variable series-level embedding sequence The expression is:

[0026]

[0027] in, is the exogenous variable sequence, Indicates exogenous variables, series-level embedding sequence , is the number of exogenous variables, is variable-level embedding.

[0028] Preferably, in step S4, the interactive information dependency relationship is obtained ,relation and interactive relationships The method is: global token at series level granularity And time token Cross-attention is applied to different levels for processing, including:

[0029] Information interaction within patch-level embedding: Self-attention mechanism is used to fuse and interact information within N patch levels;

[0030] Information interaction between patch level and series level: global token at series level Aggregate the entire endogenous variable sequence using cross attention Patch-level information;

[0031] Exogenous variables series-level embedding and global tokens at series-level granularity Information interaction: global token at series level Aggregating exogenous variables sequence-level embeddings using criss-cross attention information;

[0032] The specific expression is:

[0033]

[0034]

[0035]

[0036] in, Indicates Layer TimeXer, is the total number of layers, Indicated The endogenous variable embedding of the layer TimeXer represents the input; For the The global token input of the layer TimeXer, and: ; Representation layer normalization, Represents the cross attention mechanism.

[0037] Preferably, in step S5, a fusion embedding representation of endogenous and exogenous variables is obtained The specific steps include:

[0038] Will As the query vector query, As the key vector key and the value vector value, to establish the cross attention between the two types of variables, and finally interact with the internal information of the patch-level embedding Perform superposition operations to form the final joint information fusion embedding , and its specific implementation formula is as follows:

[0039]

[0040] in, represents the cross attention mechanism, Representation layer normalization.

[0041] Preferably, in step S6, the predicted value is obtained The method is: use a fully connected layer to fusion embedding representation Output:

[0042]

[0043] in, It is a standard Transformer model.

[0044] Preferably, the predicted probability is obtained The steps include:

[0045] S7.1: Convert the predicted value Input to the first fully connected layer of the multilayer perceptron and mapped to the hidden dimension ,Right now:

[0046] ,

[0047] in, represents the output of the first layer, represents the first layer weight matrix, represents the first layer bias, Represents a nonlinear activation function, which is used to introduce nonlinear features;

[0048] S7.2: The output of the first layer The fully connected layers are passed to the next layer in sequence to gradually learn the features, namely:

[0049] ,

[0050] in, Indicates The output of the layer, Indicates The output of the layer, Indicates The layer weight matrix, Indicates Layer bias.

[0051] Preferably, the predicted probability is obtained The steps also include:

[0052] S7.3: Map features to output dimensions through fully connected layers and through The activation function generates the binary classification probability at each time step:

[0053] ,

[0054] in, Indicates The output of the layer, that is, the output of the second-to-last layer, represents the output layer weight matrix, represents the output layer bias;

[0055] Said The formula for the activation function is:

[0056] ,

[0057] in, is the input value, and is applied element by element to the input matrix .

[0058] The elevator maintenance cycle prediction method based on multi-granularity embedding of exogenous and endogenous variables provided by the present invention has the following beneficial effects compared with the prior art:

[0059] 1. The present invention divides the elevator operation parameters into endogenous variables (fault variables) and exogenous variables (other variables). For endogenous variables, they are divided 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 used, which emphasizes capturing the overall trend and avoids introducing noise due to excessive granularity.

[0060] 2. The present invention also introduces a learnable global token to represent the macro state of the entire endogenous variable, and cross-attention is calculated with the embeddings of different granularities. As a bridge, it captures the complex relationship between endogenous and exogenous variables and promotes information fusion between different granularities.

[0061] 3. The present invention uses the powerful capabilities of the Transformer architecture to capture the complex associations between endogenous and exogenous variables through a cross-attention mechanism, and further strengthens this connection through the interaction between global and local tokens, so that the model can 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 complex dependencies in time series and integrate multiple types of data, thereby showing higher accuracy and reliability in elevator maintenance cycle prediction. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0064] Figure 3 This is a schematic diagram of the steps of data cleaning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

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

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

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

[0069] S3: Map each exogenous variable to a low-dimensional feature space through linear projection, and embed each exogenous variable into a token with series-level granularity , after splicing, a complete exogenous variable embedding sequence of series-level granularity is formed ;

[0070] S4: Using self-attention mechanism to capture endogenous variable patch-level granularity fragments Internal time dependencies , while capturing the fragments of patch-level granularity through the cross-attention mechanism Global tokens with series-level granularity The relationship between And global tokens at series level granularity With exogenous variable token The interaction between , thereby improving the information fusion between endogenous and exogenous variables. ;

[0071] S5: Using cross-attention mechanism to capture interactive relationships and relationship information interaction, further strengthen the association between exogenous and endogenous variables, ensure that the influence of exogenous variables on endogenous variables is effectively transmitted, and obtain the fusion embedding representation of exogenous and endogenous variables ;

[0072] S6: Embedding Fusion into Representation Input the standard Transformer architecture, perform information fusion through Encoder-Decoder, and output the predicted value , to achieve effective integration of endogenous and exogenous variables and accurate prediction of time series;

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

[0074] Specifically, the present invention divides the elevator operation parameters into endogenous variables (fault variables) and exogenous variables (other variables). The endogenous variables are divided into multiple non-overlapping time segments and mapped to patch-level tokens to capture internal details. The exogenous variables are embedded with series-level coarse-grained embedding to capture the overall trend, avoiding the introduction of noise due to excessive granularity. The present invention introduces a learnable global token to represent the macro state of the entire endogenous variable, and calculates cross-attention with embeddings of different granularities. As a bridge, it captures the complex relationship between endogenous and exogenous variables and promotes information fusion between different granularities. The present invention uses the powerful capabilities of the Transformer architecture to capture the complex relationship between endogenous and exogenous variables through a cross-attention mechanism, and further strengthens this connection through the interaction between global and local tokens, so that the model can more accurately understand the impact of internal and external factors on the elevator maintenance cycle.

[0075] In the present invention, the elevator operation parameters obtained include multiple data types, such as basic information of the elevator, IoT sensor data, fault and entrapment event data, maintenance records, detection and inspection data, and external environmental data such as weather and meteorology. Since these data come from different data sources and devices, they often have problems such as noise, missing values, inconsistency, and outliers, which affect subsequent analysis and modeling, so effective data preprocessing must be performed. 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 used to ensure that each fault record is counted only once, thereby avoiding erroneous statistics caused by duplicate data and ensuring the accuracy of fault frequency and maintenance response. In response to the problem of missing data, KNN Imputation is used to fill missing values ​​based on the similarity of data to reduce the interference of missing data on the analysis results. For obvious erroneous data, such as illogical installation time or equipment information, outliers are automatically identified by setting rules and constraints, and erroneous data are corrected in combination with manual review. For abnormal data with extreme fluctuations, an outlier detection method based on local weighted regression (such as LOF) is used to identify and process data points that deviate from the normal pattern. After these cleaning steps, the data will eventually be normalized, and the maximum and minimum normalization method will be used to scale the data to the range of 0 to 1 to ensure that different types of data can be compared and analyzed on the same scale. The processed data will be more consistent and accurate, providing a reliable basis for subsequent predictions and decisions. 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 source of the 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, with a 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). SMOTE oversampling is applied to the training set: the number of positive samples is oversampled to 1 / 10 of the negative samples, effectively avoiding overfitting and underfitting at the data level, so the sampling_strategy parameter of the SMOTE class is set to 0.1.

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

[0087] S2.1: Endogenous variables When segmenting (target variable / elevator fault condition), a series-level global toke is learned to represent the overall macro information of the endogenous variable. The expression of the endogenous variable block operation is:

[0088]

[0089] in, is the sequence length, For a time series at time step The value at is the time series of failure data, The time series is divided into blocks. In this embodiment, the time series is divided into patches. For the patches, and the length of each patch is , P is used as the hyperparameter of the model and is manually updated according to the effect during training and verification. .

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

[0091]

[0092] in, is the operation of embedding the patch sequence after segmentation. Specifically, it embeds { } is projected into a low-dimensional feature space and added to its own position embedding code (a standard Transformer operation used to enhance the global view), thus obtaining ——Token embedding with N patch-levels.

[0093] S2.3: Using learnable linear projections for endogenous variables Do a full sequence-level projection to get a global embedding:

[0094]

[0095] in, is a learnable linear projection, It is a global embedding at series level granularity.

[0096] In step S3, construct the exogenous variable series-level embedding sequence The specific method is:

[0097]

[0098] in, is the exogenous variable sequence, Indicates exogenous variables, series-level embedding sequence , is the number of exogenous variables, is variable-level embedding. For example, if there are three exogenous variables: temperature, weather, and holidays, then . It is a variable-level embedding, which is implemented by using a trainable linear projection Projection into sequence-level feature space.

[0099] In step S4, the interaction information dependency relationship is obtained ,relation and interactive relationships The method is: global token at series level granularity And time token Cross-attention is applied to different levels for processing, including:

[0100] Information interaction within patch-level embedding: Self-attention mechanism is used to fuse and interact information within N patch levels;

[0101] Information interaction between patch level and series level: global token at series level Aggregate the entire endogenous variable sequence using cross attention Patch-level information;

[0102] Exogenous variables series-level embedding and global tokens at series-level granularity Information interaction: global token at series level Aggregating exogenous variables sequence-level embeddings using criss-cross attention information;

[0103] The specific expression is:

[0104]

[0105]

[0106]

[0107] in, Indicates Layer TimeXer, is the total number of layers, Indicated The endogenous variable embedding of the layer TimeXer represents the input; For the The global token input of the layer TimeXer, and: ; Representation layer normalization, Represents the cross attention mechanism.

[0108] In step S5, the fusion embedding representation of the endogenous and exogenous variables is obtained The specific steps include:

[0109] Will As the query vector query, As the key vector key and the value vector value, to establish the cross attention between the two types of variables, and finally interact with the internal information of the patch-level embedding Perform superposition operations to form the final joint information fusion embedding , this embedding combines the self-attention of internal information and the cross-attention of endogenous and exogenous variables. Its specific implementation formula is as follows:

[0110]

[0111] in, represents the cross attention mechanism, Representation layer normalization.

[0112] In step S6, the predicted value is obtained The method is: use a fully connected layer to fusion embedding representation Output:

[0113]

[0114] in, It is a standard Transformer model.

[0115] Based on the output prediction value obtained , the output prediction value of the encoder When input into the multi-layer perceptron, the network will go through multiple layers of fully connected layers and nonlinear transformations of activation functions, and finally pass The activation layer converts the output of the network into a binary classification probability for each time step. This output is , represents the model's predicted probability of the label. The specific steps include:

[0116] S7.1: Input to the first fully connected layer of the multilayer perceptron and mapped to the hidden dimension ,Right now:

[0117] ,

[0118] in, represents the output of the first layer, represents the first layer weight matrix, represents the first layer bias, Represents a nonlinear activation function, which is used to introduce nonlinear features;

[0119] S7.2: The output of the first layer It is passed to the subsequent fully connected layers in sequence to gradually learn the features, namely:

[0120] ,

[0121] in, Indicates The output of the layer, Indicates The output of the layer, Indicates The layer weight matrix, Indicates Layer bias;

[0122] S7.3: Map features to output dimensions through fully connected layers and through The activation function generates the binary classification probability at each time step:

[0123] ,

[0124] in, Indicates The output of the layer, that is, the output of the second-to-last layer, represents the output layer weight matrix, represents the output layer bias, Represents the classification probability of the corresponding label of the sample in multiple future time steps.

[0125] The formula for the activation function is:

[0126] ,

[0127] in, is the input value, and is applied element by element to the input matrix .

[0128] During the training phase, the model output is calculated and the true label The binary cross entropy loss function between the two is used, and the back propagation algorithm is used to update the weight parameters in the network to minimize the error. The optimization algorithm is used to adjust the parameters of the model so that it keeps approaching the real label distribution. The model is trained and verified multiple times, and the model hyperparameters are adjusted to obtain the final prediction model. The trained and optimized model will be used to predict the elevator maintenance cycle. This task is regarded as a multi-label classification problem. The prediction results will cover multiple cycles in the future and help formulate a more accurate elevator maintenance plan.

[0129] The test data set of this embodiment preferably uses 26,190 IoT elevators in use 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 some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

Claims

1. A method for predicting elevator maintenance cycle based on multi-granularity embedding of endogenous and exogenous variables, characterized in that: The specific steps include: S1: Obtaining the operating parameters of the elevator, wherein the operating parameters include: endogenous variables and exogenous variables; S2: Split the endogenous variables into multiple non-overlapping time segments of patch-level granularity , each segment is mapped to a time token through linear projection , the endogenous variables are projected as a global token at the series level ; S3: Map each exogenous variable to a low-dimensional feature space through linear projection, and embed each exogenous variable into a token with series-level granularity , after splicing, a complete exogenous variable embedding sequence of series-level granularity is formed ; S4: Use self-attention mechanism to capture endogenous variable patch-level granularity fragments Internal time dependencies , while capturing the fragments of patch-level granularity through the cross-attention mechanism Global tokens with series-level granularity The relationship between And global tokens at series level granularity With exogenous variable token The interaction between ; S5: Using cross-attention mechanism to capture interactive relationships and relationship information interaction to obtain the fusion embedding representation of endogenous and exogenous variables ; S6: Embedding Fusion into Representation Input the standard Transformer architecture, perform information fusion through Encoder-Decoder, and output the predicted value ; S7: The predicted value Input into the multi-layer perceptron, after conversion and activation, output the Transformer model's predicted probability for the label .

2. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1 is characterized in that: The operation data of the elevator is stratified sampled by using the train_test_split function in the scikit-learn package, and is divided into a training set, a validation set, and a test set according to a set ratio. The training set is oversampled by using the SMOTE method in the imblearn library.

3. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1 is characterized in that: In step S2, a time token is obtained. and global tokens at series level granularity The specific steps include: S2.1: Endogenous variables When performing segmentation, a series-level global toke is learned to represent the overall macro information of the endogenous variables. The expression of the endogenous variable block operation is: in, is the sequence length, For a time series at time step The value at is the time series of failure data, Perform block operations on time series; S2.2: Learn a time token for each segment: in, The operation of embedding the divided patch sequence.

4. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 3 is characterized in that: Get time token and global tokens at series level granularity The specific steps also include: S2.3: Using learnable linear projections for endogenous variables Do a full sequence-level projection to get a global embedding: in, is a learnable linear projection, It is a global embedding at series level granularity.

5. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1 is characterized in that: Constructing exogenous variable series-level embedding sequence The expression is: in, is the exogenous variable sequence, Indicates exogenous variables, series-level embedding sequence , is the number of exogenous variables, is variable-level embedding.

6. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1 is characterized in that: In step S4, the interactive information dependency relationship is obtained. ,relation and interactive relationships The method is: global token at series level granularity And the time token at the patch level Cross-attention is applied at different levels for processing, including: information interaction within patch-level embedding: information fusion and interaction between N patch-levels using self-attention mechanism; information interaction between patch-level and series-level: global token at series-level granularity Aggregate the entire endogenous variable sequence using cross attention Patch-level information; exogenous variables series-level embedding and global tokens at series-level granularity Information interaction: global token at series level Aggregating exogenous variables sequence-level embeddings using criss-cross attention The specific expression is: ,in, Indicates Layer TimeXer, is the total number of layers, Indicated The endogenous variable embedding of layer TimeXer represents the input; For the The global token input of the layer TimeXer, and: ; Representation 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 is characterized in that: In step S5, the fusion embedding representation of the endogenous and exogenous variables is obtained. The specific steps include: As the query vector query, As the key vector key and the value vector value, to establish the cross attention between the two types of variables, and finally interact with the internal information of the patch-level embedding Perform superposition operations to form the final joint information fusion embedding , and its specific implementation formula is as follows: in, represents the cross attention mechanism, Representation layer normalization.

8. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 1 is characterized in that: In step S6, the predicted value is obtained. The method is: Use a fully connected layer to fusion embedding representation Output: in, It 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 is characterized in that , get the predicted probability The steps include: S7.1: Convert the predicted value Input to the first fully connected layer of the multilayer perceptron and mapped to the hidden dimension ,Right now: ,in, represents the output of the first layer, represents the first layer weight matrix, represents the first layer bias, Represents a nonlinear activation function, which is used to introduce nonlinear features; S7.2: The output of the first layer The fully connected layers are passed to the next layer in sequence to gradually learn the features, namely: ,in, Indicates The output of the layer, Indicates The output of the layer, Indicates The layer weight matrix, Indicates Layer bias.

10. The elevator maintenance cycle prediction method based on multi-granularity embedding of endogenous and exogenous variables according to claim 9 is characterized in that , get the predicted probability The steps also include: S7.3: Map features to output dimensions through fully connected layers and through The activation function generates the binary classification probability at each time step: ,in, Indicates The output of the layer, that is, the output of the second-to-last layer, represents the output layer weight matrix, represents the output layer bias; The formula for the activation function is: ,in, is the input value, and is applied element by element to the input matrix .

Images