Elevator risk early warning method based on channel clustering of Transform

Abstract

The invention provides an elevator risk early warning method based on channel clustering of Transform, which comprises the following steps of: disassembling multivariable time sequence data into a trend item and a seasonal item through a time sequence decomposition module, clustering channels with internal similarity in a frequency domain space, and adopting a cross attention module with a routing mechanism to realize early warning of elevator risk. Channel interaction in the clusters and between the clusters is realized, so that the relevance between different channels is effectively captured, and the accuracy of elevator risk prediction is improved; a maintenance unit can formulate a more scientific maintenance plan accordingly, the elevator operation and maintenance efficiency is improved, and the maintenance cost is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of elevator fault warning, and in particular relates to an elevator risk warning method based on Transformer and channel clustering. Background Art

[0002] Extensive engineering experience demonstrates that elevator operation data contains a wealth of information that indicates potential fault precursors. This includes time-series features such as abnormal traction motor vibration spectra, sudden changes in wire rope wear, and voltage fluctuations in the control cabinet, as well as discrete features such as manufacturing unit process variations and the integrity of maintenance records. Deep learning models, leveraging their multi-layer nonlinear mapping capabilities, can predict faults by mining latent correlations in historical data without requiring a precise physical model. Therefore, researching and implementing a deep learning-based elevator fault warning system has significant research significance and social value.

[0003] In practical applications such as elevator risk warning, time series often exhibit multivariate patterns, and their inherent correlations present a key modeling challenge. Current multivariate time series forecasting methods suffer from two main limitations: channel-independent strategies optimize single-channel modeling but break potential cross-channel dependencies; while channel-dependent strategies incorporate global correlations but introduce heterogeneous noise by conflating all channels. Summary of the Invention

[0004] In order to overcome the deficiencies of the prior art, the present invention provides an elevator risk warning method based on Transformer channel clustering.

[0005] To achieve the above objectives, the present invention provides an elevator risk warning method based on Transformer channel clustering, which specifically includes the following steps: S1: Obtain IoT elevator data and generate multivariate time series features; S2: In the feature extraction layer of the Transformer model, the time series decomposition module is used to decompose the multivariate time series features into trend terms and seasonal terms; S3: Use the Patch block operation to block the decomposed time series features; S4: Use Fourier transform to extract the frequency domain amplitude features of time series data, calculate the similarity distance between each channel of the multivariate time series and the learnable cluster center, and convert it into a probabilistic relationship matrix to generate a binary mask matrix; S5: Learn the channel dependencies within the cluster through the dual-route cross-attention mechanism and output the corresponding vector; S6: In the prediction layer, the clustered vectors are used to generate the corresponding fault types and obtain the corresponding fault probability distribution; the trained model is used to extract specific maintenance requirements.

[0006] Preferably, in step S2, when the time series decomposition module performs seasonal term decomposition, the time series is converted from the time domain to the frequency domain, and the conversion formula is: , in, is the seasonal pattern, Φ and A are the phase and amplitude obtained by Fourier transform DFT(X), respectively. The first K with the largest amplitude f frequency components.

[0007] Preferably, in step S2, when the time series decomposition module decomposes the trend term, the residual signal X after removing the seasonal term is extracted based on the sliding average method. rem =X—X sea The trend pattern in is as follows: , in, is the trend pattern extracted after padding and average pooling, Padding(·) is the padding operation, AvgPool(·) is the average pooling operation, X is the time series, and the local average value is calculated by sliding a fixed window in the time dimension; X sea It is a seasonal pattern; The extracted seasonal patterns and trend patterns are superimposed on the original input time series to obtain the final input time series.

[0008] Preferably, the specific steps of generating the binary mask matrix in step S4 include: Initialize a set of learnable cluster embeddings whose number is smaller than the number of time series data variables; Perform a real fast Fourier transform on each channel data and project it into a finite frequency space with a fixed Fourier basis, whose spatial dimension is d m / 2; The channel feature vector is obtained by norm calculation. The specific formula is as follows: , , in, is the time series of the i-th channel in the p-th Patch, is the eigenvector of the jth cluster center in the pth patch, rFFT(·) represents the real fast Fourier transform, norm(·) represents the calculation of the amplitude value of the complex frequency domain vector; chan is the eigenvector after Fourier transform; The Euclidean distance metric is used to quantify the similarity between the channel and the cluster embedding in the frequency domain, and its calculation formula is: , in, is the value of the frequency domain vector of the ith channel in the pth Patch at the nth frequency component, is the embedding value of the pth Patch in the jth cluster embedding on the nth frequency component, d m / 2 is the dimension of frequency domain features.

[0009] Preferably, the specific step of generating a binary mask matrix in step S4 further includes: Similarity distance based on channel and cluster embedding in frequency domain , calculate the relationship matrix between channels and clusters and normalize it to the interval [0,1]. The formula is as follows: , , , Where D, R, and P are distance matrix, relationship matrix, and probability matrix, respectively; γ∈ (0, 1) is the discount factor. 、 and are the probability, distance, and relationship of channel i belonging to cluster j in the prediction task. Through the above process, the relationship between channel and cluster embedding is probabilistic; The probability matrix P is converted into a binary channel mask to obtain the cluster membership matrix M.

[0010] Preferably, in step S5, the specific steps include: Routing clustering stage: embed the initial clustering As a router, it learns the channel dependencies within the cluster through the cross-attention mechanism and shields the noise outside the cluster. Its expression is: , in, is the updated cluster embedding, M is the cluster membership matrix, represents element-wise multiplication for sparse masking of the attention matrix, W Q 、W K and W V is a learnable parameter matrix, exp(·) is a nonlinear transformation of the attention score to simulate the scaling effect of the Softmax function; Normalize(⋅) is a row-by-row normalization operation to control the sum of the attention weights to 1; P is the number of patches; is the channel representation of the time series; T is the transpose of the matrix, d m is the dimension of frequency domain features.

[0011] Preferably, in step S5, when performing the clustering operation, the steps further include: Routing distribution phase: embed the updated clusters into As key vector K and value vector V, represented by the channels of the time series As the query vector Q, information exchange between channels is carried out to build a fully connected relationship between different clusters; the formula of the routing distribution stage is as follows: , in, 、 、 is a learnable parameter, is the output of a single Patch block, which has the relationship between different channels; all Patch blocks share the same cross attention mechanism to obtain the output of all Patch blocks. ,Right now ; X dim Add the features after time series decomposition.

[0012] Preferably, the multivariate time series features in step S1 include: elevator running speed, elevator running times and temperature features.

[0013] Preferably, in step S6, the prediction layer converts the elevator aggregate features into elevator failure probability through an MLP structure and a Sigmod function, wherein the MLP structure has two fully connected layers, and the calculation steps include: The first fully connected layer is used to cross the time series channel X out Calculate, specifically: , Among them, W1 and b1 are the weight matrix and bias vector of the first layer, is the activation function, and T1 is the output intermediate hidden layer feature.

[0014] Preferably, the step of calculating the failure probability in the prediction layer further includes: Generate prediction scores through the second fully connected layer: , Among them, W2 and b2 are the second layer parameters, is the original score vector of each fault category, which is normalized using the Softmax function to convert the predicted score into a probability distribution.

[0015] The present invention provides an elevator risk warning method based on Transformer and channel clustering. Compared with the existing technology, the method has the following advantages: To balance fine-grained modeling and correlation capture, the present invention constructs a channel clustering module. First, channels with high structural similarity are dynamically aggregated in the frequency domain (through FFT amplitude features). Then, a dual-route cross-attention mechanism is designed to model the complex dependencies between multiple channels through collaborative enhancement between clusters and refined interactions within clusters. This framework significantly improves the accuracy and robustness of risk prediction while retaining the local characteristics of the channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of an elevator risk warning method based on Transformer channel clustering provided by the present invention.

[0017] Figure 2 A system diagram of an elevator risk warning method based on Transformer channel clustering provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. The 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 the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0019] like Figure 1 As shown, the present invention provides an elevator risk warning method based on Transformer channel clustering, which is applied to the elevator risk warning big data model system. The specific steps include: S1: Obtain IoT elevator data and generate multivariate time series features; S2: In the feature extraction layer of the Transformer model, the time series decomposition module is used to decompose the multivariate time series features into trend terms and seasonal terms; S3: Use the Patch block operation to block the decomposed time series features; S4: Use Fourier transform to extract the frequency domain amplitude features of time series data, calculate the similarity distance between each channel of the multivariate time series and the learnable cluster center, and convert it into a probabilistic relationship matrix to generate a binary mask matrix; S5: Learn the channel dependencies within the cluster through the dual-route cross-attention mechanism, block the noise outside the cluster, and output the corresponding vector; S6: In the prediction layer, the clustered vectors are used to generate the corresponding fault types and obtain the corresponding fault probability distribution; the trained model is used to extract specific maintenance requirements.

[0020] Specifically, based on the multivariate time series vectors of IoT elevators (i.e., the time series vectors formed by processing the raw data of IoT elevators), this paper employs a time series decomposition module to decompose the multivariate time series data into trend terms and non-trend terms (i.e., seasonal terms). Channels with inherent similarity in the frequency domain are then clustered. During the clustering phase, the frequency domain amplitude features extracted by Fourier transform are used to calculate the similarity distance between each channel of the multivariate time series and the learnable cluster center. This is converted into a probabilistic relationship matrix, which then generates a binary mask matrix. A cross-attention module with a routing mechanism is then employed to enable channel interaction within and between clusters, effectively capturing the correlations between different channels while simultaneously shielding against noise interference outside of the clusters. This improves the accuracy of elevator risk prediction. This allows maintenance units to formulate more scientific maintenance plans, enhance elevator operation and maintenance efficiency, and reduce maintenance costs. During model training, performance is enhanced through iterative adjustments using an optimization algorithm. When predicting maintenance needs, predictions are further refined using multi-layer convolutional blocks based on fault type and probability distribution, resulting in a clearer and more efficient early warning process.

[0021] In machine learning and clustering algorithms, "learnable cluster centers" typically refer to cluster centers that can be adaptively adjusted and optimized during model training. These centers are not fixed in advance, but are learned from the training data in a data-driven manner to better fit the data distribution characteristics. While cluster centers in traditional clustering algorithms (such as K-means) are often determined through initialization or heuristic methods, learnable cluster centers are dynamically optimized through model parameter updates (such as gradient descent) to more closely match the actual data distribution.

[0022] In this embodiment, the raw data of the IoT elevator is cleaned and preprocessed before training and prediction. Data cleaning is mainly used to deal with abnormal information in the raw data. Data preprocessing includes: missing value processing and discretization of elevator-related text information.

[0023] Data cleaning primarily targets abnormal information in the raw data, including garbled characters, mixed fields, numerical anomalies, and inconsistent formats. Differentiated processing strategies are required based on the type of anomaly: For garbled data, valid information is retained through manual screening or directly eliminated; field mixing problems are often caused by data export errors, and the field correspondence can be corrected through manual verification; numerical anomalies are caused by missing values or automatic filling, and are repaired using rule-based detection methods; format inconsistencies are handled collaboratively by humans and machines to ensure that data at different time points and standards can be integrated and used. Data preprocessing covers missing value processing and the discretization of elevator-related text information. Specifically, fields with clear identifiers, such as dates and fault codes, can be quickly converted through automated rules. Text information that lacks identifiers relies on keyword extraction and field reorganization for structured processing, thereby constructing a standardized input data set.

[0024] Based on the multivariate time series feature vectors, this paper decomposes the time series to capture its periodicity (i.e., seasonal data) and trend. Periodicity refers to a pattern of fluctuations in a time series that recurs at regular intervals, with predictable fluctuation periods. For example, elevator operation data may exhibit regular variations on a daily, weekly, or monthly basis. Trend refers to the overall increasing or decreasing trend in a time series over a longer time span, reflecting the direction of long-term data change.

[0025] The time series decomposition module includes seasonal term decomposition and trend term decomposition to extract periodic patterns and long-term trends. When the time series decomposition module performs seasonal term decomposition, in order to convert the time series from the time domain to the frequency domain, the discrete Fourier transform (DFT) is used to decompose the input time series X so that it can be represented as a linear combination of the Fourier basis. Specifically, the K f Fourier basis to maintain the sparsity of the frequency domain, and then reconstruct the seasonal pattern X by inverse Fourier transform (IDFT) sea The specific process is as follows: , Where Φ and A are the phase matrix and amplitude vector calculated by discrete Fourier transform DFT (X), respectively, and the set {f1,...,f_K f} means the first K selected after sorting in descending order of amplitude f frequency components.

[0026] When the time series decomposition module decomposes trend items, it uses the sliding average principle to highlight long-term trend characteristics by smoothing periodic fluctuations. Specifically, it extracts trend patterns based on the residual signal Xrem=X-Xsea after removing seasonal components. The specific process is as follows: , Among them, Xtrend The trend pattern is extracted through padding and average pooling. Padding (·) represents the boundary padding operation, and AvgPool (·) is an average pooling operation that calculates the local mean by sliding a fixed window in the time dimension.

[0027] Finally, the extracted seasonal pattern and trend pattern are superimposed on the original input time series X to obtain the final input time series.

[0028] After the above time series decomposition process, the time series data already contains seasonal and trend characteristics. To avoid the attention mechanism modeling only a single point in time and better capture the temporal relationship in the sequence, the time series data is divided into blocks, specifically: Assume that the input multivariate time series , where R is the vector space, T is the length of the time series, and C is the feature dimension. For a given Patch size S, the time series is divided into P patches, where P = T / S, and the Patch sequence is obtained. This method solves the problem that traditional attention mechanisms have difficulty capturing long-range temporal dependencies when directly processing single-point temporal features. The block operation aggregates S consecutive time steps into a local temporal unit, enabling the attention mechanism to model in "time block" units. For example, by learning the changing pattern of elevator vibration signals in short periods through the interaction of time steps within a patch, single-point modeling can be expanded to local temporal correlation analysis.

[0029] After the block-based time series data is input into the channel clustering module to learn cross-channel (cross-dimensional) dependencies. The channel clustering module aims to capture the dependencies between multi-dimensional features in different clusters. This module first projects the time series information into the frequency domain using a real Fourier transform, providing a frequency-domain feature foundation for channel clustering analysis. Then, using the dual mechanisms of route aggregation and route distribution, it constructs an interactive correlation network between multiple variables (channels), achieving efficient aggregation and distribution of cross-channel features. This process is divided into three stages: clustering (step S4), route clustering, and route distribution (step S5).

[0030] (1) Clustering stage In the clustering phase, the model initializes a set of K learnable cluster embeddings (K is less than the number of channels) and performs a patch operation on them. To enhance cross-channel modeling capabilities, a frequency domain modeling strategy is used to mine discriminative features. Specifically, for the multivariate time series processed by the dual time series module, for each patch block Xp,:, a real fast Fourier transform is performed on each channel data and projected into a finite frequency space with a fixed Fourier basis (dimension d m / 2). The channel feature vector is obtained by calculating the norm (amplitude) , perform the same transformation on the cluster embedding to obtain , the specific calculation logic is as follows: , , in, is the time series of the i-th channel in the p-th Patch, is the eigenvector of the jth cluster center in the pth Patch, rFFT(·) represents the real fast Fourier transform, norm(·) represents the calculation of the amplitude value of the complex frequency domain vector; chan is the eigenvector through Fourier transform.

[0031] In order to evaluate the similarity between channel and cluster embedding in the frequency domain, the Euclidean distance is used to quantify the similarity between channel and cluster embedding in the frequency domain, namely: , in, is the value of the frequency domain vector of the ith channel in the pth Patch at the nth frequency component, is the embedding value of the pth Patch in the jth cluster embedding on the nth frequency component, d m / 2 is the dimension of frequency domain features.

[0032] Similarity distance based on channel and cluster embedding in frequency domain , calculate the relationship matrix between channels and clusters and normalize it to the interval [0,1]. The formula is as follows: , , , Where D, R, and P are distance matrix, relationship matrix, and probability matrix, respectively; γ∈ (0, 1) is the discount factor. 、 and are the probability, distance, and relationship of channel i belonging to cluster j in the prediction task. Through the above process, the relationship between channel and cluster embedding is probabilistic; In addition, in order to effectively filter out irrelevant channel interference and strengthen the feature aggregation of similar channels, the Bernoulli resampling mechanism is introduced to convert the probability matrix P into a binary channel mask to generate the clustering member matrix M. Follows Bernoulli distribution ≈Bernoulli ( ), through probability sampling, the high probability value P_i,j( ) is more easily mapped to 1, ensuring that channels with high correlation are divided into corresponding clusters with a greater probability, realizing adaptive screening of channel features and clustering attribution.

[0033] (2) Routing clustering stage In the routing clustering stage, the initial clusters generated in the early stage are embedded into As a router, , d m The embedding dimension is used as the core component of routing, and the improved cross-attention mechanism is used to strengthen the channel dependency within the cluster while suppressing the noise interference outside the cluster. As the query vector (Query), the time series processed by the dual time series module is represented as At the same time, it serves as both key and value, and aggregates feature information of similar channels through the cross-attention mechanism. Its mathematical expression is as follows: , Among them, M is the cluster membership matrix, which is an approximate binary mask matrix used to implement sparse attention on the channels within the cluster, that is, sparse masking is implemented on the attention matrix through element-by-element multiplication, allowing only the channels within the cluster to participate in the attention calculation. represents element-wise multiplication for sparse masking of the attention matrix, W Q 、W K and W V is a learnable parameter matrix. The attention scores are nonlinearly transformed using the exp function to simulate the scaling effect of the Softmax function. The weights are then normalized using Normalize() to sum to 1. T is the transpose of the matrix. The calculated updated cluster embeddings are used to iteratively optimize the cluster probability matrix and provide a feature foundation for cross-cluster information exchange in the subsequent routing and distribution phase.

[0034] (3) Routing distribution phase.

[0035] In the routing distribution phase, the updated clustering embedding Key vector K and value vector V, represented as channels of time series As the query vector Q, cross-channel information sharing is achieved through the cross-attention mechanism, thereby building a fully connected interactive relationship between different clusters. The specific calculation logic of the routing distribution stage is as follows: , in, 、 、 is a learnable parameter, A single Patch block aggregates the output of the relationship between different channels; all Patch blocks share the same cross attention mechanism to obtain the output of all Patch blocks. ,Right now .

[0036] Finally, X dim The features after time series decomposition are added together to obtain the final time series cross-channel representation Xout, which captures the dependencies across time and dimensions at the same time.

[0037] Based on the feature similarity weights learned by the channel clustering module, this paper implements fault probability prediction for IoT elevators through a prediction layer. The prediction layer employs a two-layer fully connected MLP structure for dimensionality reduction. Compared to the linear fitting capability of a single fully connected layer, this structure effectively captures nonlinear mapping relationships in the data. The specific calculation process is as follows: First, the time series cross-channel representation X out Calculated by the first fully connected layer: , Among them, W1 and b1 are the weight matrix and bias vector of the first layer, is the activation function (such as ReLU), and T1 is the output intermediate hidden layer feature.

[0038] The prediction score is then generated through the second fully connected layer: , Where W2 and b2 are the second layer parameters, is the original score vector of each fault category. To convert the predicted score into a probability distribution, the Softmax function is used for normalization.

[0039] In its implementation, the output vector of the channel clustering module first enters a fully connected layer for dimensionality reduction. Through weighted summation and bias calculation, high-dimensional time series features are mapped to a low-dimensional space. This process reduces feature dimensionality by learning the underlying patterns in the data, improving computational efficiency while avoiding information loss. For example, multi-channel features such as the elevator traction motor vibration spectrum and door motor current can be compressed into a low-dimensional representation that includes fault correlations after MLP processing. A softmax function is then used to generate a probability distribution for each fault type.

[0040] The advantages of this prediction mechanism are: the two fully connected layers introduce nonlinear transformations through activation functions, which can capture complex associations; the Softmax output has a clear fault probability distribution, which makes it easier for maintenance personnel to formulate strategies based on risk levels; and the interpretability is enhanced: key fault indicators can be traced through intermediate features. For example, when the weight of a channel is high, the corresponding sensor data is prompted as the main warning basis.

[0041] During the model training phase, both hyperparameters and internal model parameters need to be optimized to improve prediction performance. In response to the common sample imbalance problem in elevator fault warning scenarios, the traditional cross-entropy loss function easily causes the model to favor the majority class, reducing the recognition accuracy of minority class faults. To this end, this paper introduces a weighted cross-entropy loss function, Loss, which balances the training contribution of each class by adaptively adjusting the class weights. The specific calculation formula is as follows: , Among them, N is the total number of samples, M is the number of categories, and w c is the weight coefficient of the cth category, y ic is the true probability of the sample, p ic is the probability that the i-th sample belongs to category c. The weight of category c and the weight coefficient w c The calculation formula is: , Among them, n c is the number of samples belonging to category c.

[0042] Tables 1-5 show the prediction results of the proposed system for different fault types (occupant entrapment, door opening, door opening / closing, overspeeding, and other). This study used 8,000 KONE elevator units as test subjects, analyzing data from four time periods: early May, late May, early June, and late June. The results show that the model demonstrates excellent stability in predicting door opening / closing faults, achieving a hit rate of approximately 28%. The prediction accuracy for other fault types also remained around 20%, fully validating the model's effectiveness. Even with a small sample size, the model was able to effectively predict critical faults such as occupant entrapment, door opening, and overspeeding, highlighting its generalization capabilities and fault warning value in small sample scenarios.

[0043]

[0044] Table 1 Prediction results of IoT elevator when people are trapped

[0045] Table 2 Prediction results of IoT elevator door opening failure

[0046] Table 3 Prediction results of IoT elevator door opening and closing failures

[0047] Table 4 Prediction results of IoT elevator overspeed fault

[0048] Table 5 Prediction results of other faults in IoT elevators like Figure 2 As shown, the raw data consists of three types of tables: elevator information, fault logs, and weather temperature. The elevator information and fault logs are generated by IoT elevators, while the weather temperature table is derived from historical Hangzhou meteorological data from the 2345 Weather King platform. As shown in Tables 6-9, the dataset used in this paper consists of three structured tables: elevator information, fault logs, and weather temperature. Table 9 lists common anomalies and their corresponding cleaning methods.

[0049]

[0050] Table 6 Elevator information table

[0051] Table 7 Elevator Fault Record

[0052] Table 8 Weather Temperature Table

[0053] Table 9 Most raw elevator data is in text format, making it difficult for models to process directly. To improve training efficiency and model robustness, text data needs to be discretized and encoded. This discretized data is then transformed into dense, multivariate time-series feature vectors. Common discretization encoding methods include enumeration encoding and one-hot encoding. Enumeration encoding assigns a unique number to each unique value, while one-hot encoding creates a separate field for each unique value, with a value of either 0 or 1.

[0054] Normalization mainly solves the problem of inconsistent dimensions of different features. The normalization formula is as follows: ,

[0055] Where x represents the input feature value. After normalization, all feature values are mapped to the range [0, 1] to avoid model bias caused by differences in numerical scales. To avoid data inflation during the preprocessing phase, normalization is performed in the prediction module, recording only the maximum and minimum values of each field in advance.

[0056] Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

Claims

1. A Transformer-based channel clustering elevator risk warning method, characterized by: The specific steps include: S1: Obtain IoT elevator data and generate multivariate time series features; S2: In the feature extraction layer of the Transformer model, the time series decomposition module is used to decompose the multivariate time series features into trend terms and seasonal terms; S3: Use the Patch block operation to block the decomposed time series features; S4: Use Fourier transform to extract the frequency domain amplitude features of time series data, calculate the similarity distance between each channel of the multivariate time series and the learnable cluster center, and convert it into a probabilistic relationship matrix to generate a binary mask matrix; S5: Learn the channel dependencies within the cluster through the dual-route cross-attention mechanism and output the corresponding vector; S6: In the prediction layer, the clustered vectors are used to generate the corresponding fault types and obtain the corresponding fault probability distribution; the trained model is used to extract specific maintenance requirements.

2. The elevator risk warning method based on Transformer channel clustering according to claim 1 is characterized in that: In step S2, when the time series decomposition module performs seasonal term decomposition, the time series is converted from the time domain to the frequency domain.

3. The elevator risk warning method based on Transformer channel clustering according to claim 1 is characterized in that: In step S2, when the time series decomposition module performs trend term decomposition, the trend pattern in the residual signal after removing the seasonal term is extracted based on the sliding average method; The extracted seasonal patterns and trend patterns are superimposed on the original input time series to obtain the final input time series.

4. The elevator risk warning method based on Transformer channel clustering according to claim 1 is characterized in that: The specific steps of generating the binary mask matrix in step S4 include: Initialize a set of learnable cluster embeddings whose number is smaller than the number of time series data variables; Perform real fast Fourier transform on each channel data and project it into a finite frequency space with a fixed Fourier basis; Obtain channel feature vectors through norm calculation; The Euclidean distance metric is used to quantify the similarity between channel and cluster embeddings in the frequency domain; Based on the similarity distance between the channel and cluster embeddings in the frequency domain, the relationship matrix between the channel and cluster is calculated and normalized to the interval [0,1] to probabilize the relationship between the channel and cluster embeddings; Convert the probability matrix to a binary channel mask to obtain the cluster membership matrix.

5. The elevator risk warning method based on Transformer channel clustering according to claim 1 is characterized in that: In step S5, the specific steps include: Routing clustering stage: The initial cluster embedding is used as a router, and the channel dependencies within the cluster are learned through the cross-attention mechanism to block noise outside the cluster; Routing distribution phase: The updated cluster embedding is used as the key vector and value vector, and the channel representation of the time series is used as the query vector to exchange information between channels and build fully connected relationships between different clusters. All Patch blocks share the same cross attention mechanism to obtain the output of all Patch blocks and the features after temporal decomposition and add them together.

6. The elevator risk warning method based on Transformer channel clustering according to claim 1 is characterized in that: The multivariate time series features in step S1 include: elevator running speed, elevator running times and temperature features.

7. The elevator risk warning method based on Transformer channel clustering according to claim 1 is characterized in that: In step S6, the prediction layer converts the elevator aggregate features into elevator failure probability through an MLP structure and a Sigmod function. The MLP structure has two fully connected layers, and its calculation steps include: The first fully connected layer is used to calculate the time series across channels; The prediction scores are generated through the second fully connected layer.

Images