An elevator risk early warning method and system based on time sequence blocking and dynamic-static fusion

Abstract

The application discloses an elevator risk early warning method and system based on time sequence block and dynamic and static fusion, comprising: acquiring dynamic time sequence data and static attribute characteristics of an elevator and performing normalization processing; utilizing a block coding mechanism to divide the dynamic time sequence characteristics into time sequence blocks, extracting local semantic representations and long-range time sequence dependencies of each block through a channel-independent deep neural network backbone; eliminating a reverse normalization step at an output end of the backbone network, and directly retaining a normalized distribution representation of the dynamic characteristics; splicing and fusing the retained dynamic characteristic vector and the normalized static attribute characteristics to construct a multivariate joint interaction feature space; utilizing a flattening fusion mechanism to perform full-dimensional space-time feature interaction mapping; and finally outputting an elevator fault probability in a future preset period. The application can effectively solve the problem of missing cross-variable interaction in long time sequence data, and significantly improve the accuracy and robustness of elevator fault prediction under non-stationary data flow.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of intelligent operation and maintenance and artificial intelligence technology, specifically involving an elevator risk early warning method and system based on time-series segmentation and dynamic-static fusion. Background Technology

[0002] With the acceleration of urbanization and the popularization of Internet of Things (IoT) technology, elevators, as a core transportation tool in high-rise buildings, are experiencing a continuous increase in their number. To ensure elevator safety and reduce operation and maintenance costs, data-driven predictive maintenance has become an industry consensus. Modern elevators are equipped with numerous sensors that can collect real-time operational status data (such as speed, load, door operator current, etc.) as well as discrete data such as maintenance records and fault codes. Utilizing deep learning models to uncover potential patterns in this multi-source heterogeneous data and provide early warnings of future elevator failure risks has significant social and economic value.

[0003] Despite significant progress in time series forecasting, deep learning still has limitations when applied to the specific task of elevator malfunction risk early warning:

[0004] (1) The semantic missing problem of traditional point-wise modeling Early recurrent neural networks (such as LSTM and GRU) and Transformer models based on point attention mechanisms mostly used a single time step as the input unit. However, in elevator operation data, the numerical value at a single time point (such as the instantaneous speed at a certain moment) often lacks clear semantic meaning. Only a continuous trajectory (such as a complete start-stop process) can reflect the health status of the equipment. Point-by-point modeling is not only computationally complex but also struggles to capture local temporal context information, resulting in low efficiency in feature extraction from long historical sequences.

[0005] (2) Noise interference and interaction modeling dilemmas among multivariable channels Elevator data typically contains dozens of sensor variables with complex coupling relationships, but also a large amount of irrelevant noise. On the one hand, models employing a channel-mixing strategy are easily affected by noise from a single anomalous sensor, making it difficult for the model to converge during training and even leading to overfitting. On the other hand, while existing channel-independent models (such as the original PatchTST [Patch Time Series Transformer]) alleviate the above problems by sharing weights, their network structure completely severs the connections between variables. Since elevator malfunctions are often the result of multiple concurrent variables (e.g., continuous abnormal temperature combined with high-frequency operation leading to shutdown), the lack of a spatiotemporal interaction mechanism across variables during the prediction phase will severely limit the accuracy of fault identification.

[0006] (3) Low utilization rate of static attribute features and rigid fusion method Elevator failure risk depends not only on real-time dynamic time-series data but also heavily on static attribute data (such as elevator age, brand, model, and load capacity). For example, the failure probability of an older elevator under the same operating conditions is obviously different from that of a newer elevator. Most existing time-series prediction models only focus on dynamic data or simply copy and paste static features to each time step. This approach not only increases computational redundancy but also has deep technical flaws: dynamic features (such as failure fluctuations and sudden weather changes) are high-frequency fluctuating signals, while static features (such as elevator age and rated load capacity) are global constants or extremely low-frequency signals. When calculating gradients during backpropagation, the high-frequency variance of dynamic features can easily "overwhelm" the low-frequency weights of static features (i.e., "feature collapse"), causing the model to lose its basic risk bias judgment ability. Furthermore, this early hybrid strategy failed to effectively apply static features as a "context bias" to deep decision networks, resulting in poor generalization ability of the model to different individual differences.

[0007] (4) Distribution adaptation problem when migrating regression models to classification tasks Currently, mainstream time-series large-scale models are primarily designed for continuous numerical prediction (regression tasks), and their architecture typically includes symmetrical "input normalization" and "output denormalization (RevIN Output)." However, elevator risk warning is essentially a classification task (predicting whether a malfunction will occur in the future). If a regression model architecture is directly adopted, the denormalization operation at the output end will restore the hidden layer features to their original physical dimensions. When these physical quantities with huge numerical differences are directly input into the classification activation function (such as Sigmoid or Softmax), it will lead to severe gradient instability or even gradient vanishing, hindering the model's convergence and classification performance. At the same time, direct denormalization will amplify the numerical variance of high-amplitude sensors, which can easily cause gradient oscillations or even gradient explosions when calculating the binary cross-entropy loss.

[0008] (5) Misjudgment and loss oscillation caused by extreme class imbalance In real-world elevator operation data, the number of samples in normal operation is overwhelmingly large, while positive samples representing fault risk are extremely scarce. Faced with this extreme disparity in class proportions, conventional classification loss functions can cause the model to rapidly collapse towards the majority class (normal samples), leading to gradient explosion and loss value oscillations, making it difficult for the model to accurately capture early, weak fault signals.

[0009] In summary, there is an urgent need for an elevator risk warning method that can effectively extract local semantics from long sequences, balance channel independence and interaction between variables, reasonably integrate dynamic and static features, and adapt to the distribution characteristics of classification tasks and the current state of imbalanced data. Summary of the Invention

[0010] In order to overcome the shortcomings of the prior art, the present invention provides an elevator risk early warning method and system based on time-series segmentation and dynamic-static fusion.

[0011] The primary objective of this invention is to provide an elevator risk early warning method based on time-series segmentation and dynamic-static fusion, the specific steps of which include:

[0012] Data acquisition and segmentation: Acquire historical operating data of the elevator to be monitored, and after analysis, divide the historical operating data into dynamic time-series features that change over time and static non-time-series features that describe the inherent properties of the elevator; Dynamic feature block encoding: The dynamic temporal features are normalized, and the normalized multivariate temporal data is divided into multiple overlapping time blocks along the time dimension, and mapped to generate multi-channel block embedding vectors. Channel Independent Feature Extraction: The multi-channel block embedding vector is input into a preset channel independent feature extraction network. Under the premise of maintaining the sharing of parameters and independent computation of each channel, the local semantic representation and long-range temporal dependency of each channel are extracted, and the dynamic hidden layer representation that maintains the normalized distribution state is output. Distribution feature preservation and recombination: Extract the dynamic hidden layer representation, and directly flatten the dynamic hidden layer representation of all channels and time steps that have not undergone inverse normalization processing to break the channel independence and generate a global dynamic feature vector containing cross-variable interaction information. Static and dynamic feature enhancement fusion: The static attribute features are encoded, and the encoded static attribute features are concatenated and fused with the global dynamic feature vector in the feature dimension to construct a full-dimensional enhanced feature vector containing multi-source heterogeneous information; Risk probability mapping and early warning: The full-dimensional enhanced feature vector is globally mapped using a fully connected classification network to capture the spatiotemporal interaction dependency between variables, calculate the predicted probability of elevator failure within a preset time window, and generate a risk early warning signal based on the probability.

[0013] The second objective of this invention is to provide an elevator risk early warning system based on time-series segmentation and dynamic-static fusion, comprising: The data acquisition and segmentation unit is used to acquire historical operating data of the elevator to be monitored, and after parsing, the historical operating data is divided into dynamic time-series features that change over time and static non-time-series features that describe the inherent properties of the elevator. The dynamic feature block encoding unit is used to normalize the dynamic temporal features and divide the normalized multivariate temporal data into multiple overlapping time blocks along the time dimension, and map them to generate multi-channel block embedding vectors. The channel-independent feature extraction unit is used to input the multi-channel block embedding vector into a preset channel-independent feature extraction network, extract the local semantic representation and long-range temporal dependency of each channel, and output a dynamic hidden layer representation that maintains a normalized distribution state. The distribution feature preservation and recombination unit is used to extract the dynamic hidden layer representation. It directly flattens the dynamic hidden layer representation of all channels and time steps that have not undergone inverse normalization processing to generate a global dynamic feature vector containing cross-variable interaction information. The static and dynamic feature enhancement and fusion unit is used to encode the static non-temporal features and then concatenate and fuse the encoded static non-temporal features with the global dynamic feature vector in the feature dimension to construct a full-dimensional enhanced feature vector containing multi-source heterogeneous information. The risk probability mapping and early warning unit is used to perform global mapping on the full-dimensional enhanced feature vector using a fully connected classification network, capture the spatiotemporal interaction dependency between variables, calculate the predicted probability of elevator failure within a future preset time window, and generate a risk early warning signal based on the probability.

[0014] Compared with the prior art, the advantages of this invention are:

[0015] 1. A temporal patching self-attention mechanism is introduced to achieve efficient capture of long-range local semantics and early identification of subtle faults. This invention uses a patching mechanism to aggregate single-time-point values ​​that originally lacked independent physical semantics into local semantic time slices containing the complete device action cycle (such as start-stop, acceleration). This reduces the computational complexity of the attention mechanism from quadratic... Reduce to linear While significantly improving inference efficiency, it also preserves the local micro-patterns of the sequence. This enables the model to accurately capture subtle waveform variations caused by minor wear and tear on components over a long time window, providing core algorithmic support for the "prevention of elevator problems" (i.e., providing early warnings before malfunctions occur).

[0016] 2. A heterogeneous network of "front-end channel decoupling - back-end global flattening" was constructed, achieving a perfect balance between noise resistance and multivariate spatial interaction. Addressing the dilemma of existing technologies in processing multivariate time series data—either easily amplifying single-sensor noise due to channel mixing or severing the physical connections between variables due to complete channel independence—this invention employs an innovative "separate-then-combine" mapping strategy. Maintaining channel independence in the feature extraction backbone network effectively isolates cross-channel contamination from abnormal fluctuations (noise) from a single sensor and prevents model overfitting. Furthermore, introducing flattening (the process of converting multidimensional input data into a one-dimensional vector) and fully connected interactive mapping before the decision layer forcibly reconstructs the high-dimensional coupling relationships between different physical variables. This architecture perfectly matches the nonlinear disaster-causing mechanism of "multi-factor concurrency" in elevator faults, significantly improving the identification accuracy of complex faults.

[0017] 3. A dual-branch bypass fusion architecture for dynamic and static features is proposed, achieving decoupling and complementarity between deep physical attributes and dynamic operating states. Unlike traditional methods that simply copy and brute-force static features at time steps (which easily leads to computational redundancy and the curse of feature dimensionality, and the high-frequency variance of dynamic features can easily "overwhelm" the low-frequency weights of static features, resulting in "feature collapse"), this invention decouples features through bypass branches. The dynamic backbone network focuses on extracting high-frequency waveform trends and operating cycle patterns, while low-frequency static features representing the inherent attributes of the equipment (such as elevator age, brand, rated load, etc.) bypass complex temporal attention calculations and are directly fused as "basic risk biases" at the end full-dimensional interaction layer. This mechanism not only optimizes computational overhead, but more importantly, during backpropagation, the gradient can be evenly and fairly distributed to dynamic temporal features and static attribute features, avoiding the weight of one side being "overwhelmed" by the extreme dimensions of the other. This allows the model to adaptively adjust the warning sensitivity threshold according to the "healthy base" of different elevators, greatly enhancing the model's generalization evaluation capability on elevators in different batches and under different operating conditions.

[0018] 4. This invention innovatively removes the output-side denormalization mechanism, solving the gradient oscillation problem during cross-domain migration of time-series representations to classification tasks. Addressing the issue of numerical reconstruction discrepancies caused by the inconsistencies in the dimensions of various physical sensors (such as voltage, rotational speed, and temperature) when traditional channel-independent time-series prediction models migrate to classification tasks, this invention directly retains and utilizes the normalized hidden layer distribution representation for probability mapping. This mechanism avoids the impact of large numerical fluctuations caused by denormalization on the classification layer gradient, significantly improving the convergence speed and fault prediction stability of the classification network without losing the relative trend of the original sequence.

[0019] 5. By combining business rules with weighted dynamic loss assessment, the reliability of early warnings under extreme long-tail distributions is significantly improved. Addressing the technical bias of extremely scarce elevator fault samples (extreme imbalance between positive and negative samples) in real-world industrial scenarios, this invention introduces deeply customized maintenance statistical features (such as maintenance frequency within a specific period) at the input end and employs a joint loss function with positive sample weights at the loss calculation end. Through a dual mechanism of business feature guidance and algorithmic loss penalty, the model is forced to break free from the local optimum of "predicting the majority class (normal)," significantly improving the recall rate for low-frequency, high-risk faults caused by maintenance deficiencies or overload operation, effectively reducing the system's false negative risk. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the core business process of an elevator risk warning method based on time-series segmentation and dynamic-static fusion, provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the overall network architecture of the model provided in an embodiment of the present invention. Detailed Implementation

[0022] The following specific examples illustrate the implementation 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 embodiments, 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, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0023] Example 1

[0024] like Figure 1 As shown, this embodiment provides an elevator risk early warning method based on time-series segmentation and dynamic-static fusion, which highly summarizes its core business process. The specific steps include:

[0025] S1. Data Acquisition and Classification: Acquire historical operating data of the elevator to be monitored, and after analysis, divide the historical operating data into dynamic time-series features that change over time and static non-time-series features that describe the inherent properties of the elevator.

[0026] S2. Dynamic feature block encoding: The dynamic temporal features are normalized, and the normalized multivariate temporal data is divided into multiple overlapping time blocks along the time dimension, and mapped to generate multi-channel block embedding vectors.

[0027] S3, Channel Independent Feature Extraction: The multi-channel block embedding vector is input into a preset channel independent feature extraction network. Under the premise of maintaining the sharing of parameters of each channel and independent computation, the local semantic representation and long-range temporal dependency of each channel are extracted, and the dynamic hidden layer representation that maintains the normalized distribution state is output.

[0028] S4. Distribution feature preservation and recombination: Extract the dynamic hidden layer representation, and directly flatten the dynamic hidden layer representation of all channels and time steps that have not undergone inverse normalization processing to break the channel independence and generate a global dynamic feature vector containing cross-variable interaction information.

[0029] S5. Enhanced Fusion of Static and Dynamic Features: The static non-temporal features are encoded, and the encoded static non-temporal features are concatenated and fused with the global dynamic feature vector in the feature dimension to construct a full-dimensional enhanced feature vector containing multi-source heterogeneous information.

[0030] S6. Risk Probability Mapping and Early Warning: The full-dimensional enhanced feature vector is globally mapped using a fully connected classification network to capture the spatiotemporal interaction dependency between variables, calculate the predicted probability of elevator failure within a preset time window, and generate a risk warning signal based on the probability.

[0031] Specifically, this embodiment provides an elevator risk early warning method based on time-series block coding and dynamic / static feature fusion, which is applied to an intelligent operation and maintenance cloud platform or an IoT elevator monitoring server. In actual business applications, the system uses real-time signals from multiple IoT sensors in the elevator and static device profiles recorded in the business form system as multi-source inputs, and uses the final output fault probability early warning signal as the guiding basis for front-end operation and maintenance actions. Figure 2 As shown in the schematic diagram of the overall network architecture of the model, by performing a deep adaptation of the PatchTST architecture to the classification task at the bottom layer, this method aggregates single-point time data into patches with complete mechanical semantics, and maintains channel independence at the network front end to isolate abnormal noise pollution from a single sensor. Based on this, the architecture removes the inverse normalization mechanism that easily causes gradient oscillations at the output layer, forcibly reconstructs the high-dimensional coupling of different physical variables, and introduces static features such as elevator age and load as the "basic risk bias" for evaluation at the end. This system effectively fits the nonlinear disaster-causing mechanism of "multi-factor concurrency" in elevator malfunctions, significantly improving the early warning accuracy and robustness in non-stationary data flow environments.

[0032] The steps of this embodiment will now be described in detail.

[0033] S1. Data Acquisition and Division This step aims to separate dynamic and static features from the raw data, laying the foundation for subsequent branch processing. Addressing the issue of "feature collapse" caused by early mixing of dynamic and static features as described in the background technology, this embodiment employs a front-end decoupling strategy. The system first acquires the historical operating data of the elevator to be monitored. Through a parsing engine, the data is rigorously separated into independent dynamic temporal feature streams and static non-temporal feature streams at the physical level. Specifically, this includes the following sub-steps:

[0034] S11. Construction and Selective Normalization of Dynamic Temporal Feature Matrix

[0035] Extract the time window length from the IoT database. L Historical observation sequences are used to construct dynamic time-series features. As an example, these dynamic time-series features specifically encompass the following multidimensional high-frequency and periodic fluctuation indicators:

[0036] 1) Core status and historical fault signals: including current fault status identifiers of various types (such as specific fault classification identifiers from is1 to is6) and the frequency of occurrence of various types of faults within the corresponding historical time steps (such as failureNum1 to failureNum6).

[0037] 2) Multi-scale maintenance statistics: Includes maintenance frequency statistics for half-month, quarter, semi-annual and annual periods (such as halfMonth_maintained_num, season_maintained_num), and can further construct derived features such as "cumulative non-maintained time step" to quantify the non-linear relationship between equipment health degradation and maintenance gaps.

[0038] 3) Meteorological environment aggregation features: Extract historical meteorological information of the elevator's geographical location, and calculate the mean temperature, standard deviation (representing temperature fluctuation rate), maximum temperature, and comprehensive weather index weighted score (weather_scores) within the time window.

[0039] 4) Time period characteristics: such as the number of holidays and other calendar characteristics that affect the frequency of elevator use.

[0040] To facilitate subsequent model processing, the aforementioned dynamic time-series features need to be normalized. Let the aforementioned dynamic time-series features include... C Given several physical observation variables, the constructed dynamic feature input matrix is... X d It can be represented as: in, x t ∈ R C Indicates at time step t The vector of observations from all dynamic sensors at any given time. The system extracts the mean μ ∈ [value] of each dynamic feature channel from the historical training set. R C and standard deviation σ ∈ R C Perform zero-mean normalization on the current input sequence: in, To prevent extremely small constants with a denominator of zero, this step maps all dynamic features to a standard normal distribution space with stable gradients, generating a normalized dynamic feature matrix. X d (norm) ∈ R L×C .

[0041] S12. Vectorized representation of static attribute features

[0042] In parallel with the processing of dynamic features, the system extracts static profile data describing the inherent attributes of elevators from the business ledger system. As an example, these static, non-time-series features mainly include low-frequency or immutable indicators that serve as a "benchmark" for assessing the health status of elevators:

[0043] 1) Mechanical performance and design parameters: such as elevator rated load, rated speed, number of floors, and elevator type.

[0044] 2) Equipment operating environment and status: such as the type of place of use (Place_type) and the current approved equipment status (Status).

[0045] 3) Elevator age profile features: The actual service life (elevator age) of the elevator is dynamically calculated by subtracting the installation date (installation_date) from the current time.

[0046] The above D static Each feature label is converted into a fixed-length static attribute vector through categorical encoding (such as One-Hot or Label Encoding) and numerical scaling. X S : in, s i Representing the i Static feature values ​​in each dimension.

[0047] S13, Decoupling Flow Direction Control In the deep network backbone constructed by this method, the normalized dynamic feature matrix X d (norm) The degradation trend over time is extracted by feeding the backbone network (i.e., the time-series block coding module); while the static attribute vector... X S The features are "frozen" and bypassed, skipping the complex temporal feature extraction layer until fusion occurs at the end of the network. Through this front-end decoupling strategy, the high-frequency variance of dynamic features does not interfere with the weight updates of static features, thus preserving pure "basic risk bias" information for subsequent fusion.

[0048] S2, Dynamic Feature Block Coding

[0049] This step aims to aggregate consecutive time steps into local time slices with complete mechanical semantics, in order to address the problems of lack of semantic meaning for single-point values ​​and excessive computational complexity of attention in traditional point-wise modeling.

[0050] Traditional time-series prediction models (such as the standard Transformer or Informer) typically employ a "point-wise" input approach, where each discrete time step in the time series is input as an independent token into the attention mechanism for computation. However, in elevator operation and maintenance scenarios, this point-wise input mechanism has two significant technical drawbacks: First, the single-point sampling values ​​of elevator sensors often lack complete mechanical semantics, and point-wise self-attention computation tends to allocate weights to local noise rather than the overall degradation trend; second, the time and space complexity of the standard attention mechanism is the square of the sequence length. O(L 2 ) The increasing frequency makes it impossible for the model to extract historical maintenance and feature data with extremely long backtracking periods (such as the past six months).

[0051] To overcome the aforementioned shortcomings, this embodiment innovatively introduces a channel-independent time patching and encoding mechanism when processing dynamic temporal features, aggregating point-level unordered discrete data into patch-level feature representations with local mechanical semantics. Specifically, it includes the following steps:

[0052] S21, Channel Independence and Overlapping Block Mapping

[0053] To avoid feature crosstalk between sensor variables of different dimensions in the initial stage, this network adopts a channel-independence strategy. This is applied to the normalized dynamic feature matrix output in step S1. X d (norm) ∈ R L×C First, decouple it along the channel dimension. C An independent univariate one-dimensional time series .

[0054] Subsequently, a length of P A sliding time window (i.e., block length) with step sizes. S (i.e., sliding step size) for each univariate time series x (i) Continuous segmentation is performed. To preserve temporal continuity at block boundaries, this embodiment employs an overlapping block design (i.e., S < PAfter appropriate zero-padding at the beginning and end of the sequence to ensure data integrity, the univariate time series is divided into... N The time is divided into overlapping patches. As an optional implementation, the number of patches... N The calculation formula is: After the block operation, the original univariate time series x (i) Reconstructed into two-dimensional block tensors X patch (i) ∈ R N×P Each row of the tensor represents a complete time block, containing... P A series of time steps of data.

[0055] S22, High-dimensional spatial projection of local semantics In order for the deep network to process the above-mentioned block data, this step performs linear feature mapping on each time block, transforming it from the original... P Projecting the 3D physical observation space to a dimension of D model The implicit semantic space. As an optional implementation, for the first... i Block tensors of each channel X patch (i) Its linear embedding computation is expressed as: X embed (i) = X patch (i) ·W emb + b emb in, W emb ∈ R P×Dmodel The weight matrix is ​​a learnable linear mapping. b emb ∈ R Dmodel This is the bias vector. After projection, the dimension of the output embedding tensor is increased to [dimensionality]. X embed (i) ∈ R N×Dmodel .

[0056] S23, Absolute position coding and timing fusion Since the linear projection process and subsequent self-attention mechanism lack the ability to perceive the temporal order, this step adds a positional code to each block to preserve the unidirectional evolution of the elevator over time. As an optional implementation, learnable positional encoding can be used. The positional encoding matrix is ​​then constructed. W pos ∈ R N×Dmodel This is added to the block embedding tensor to obtain the final dynamic block coding tensor input to the backbone network. X in (i) : X in (i) = X embed (i) + W pos Ultimately, all C After each channel independently completes the above encoding, they are processed in parallel at the batch level, outputting a dynamic block encoding tensor, which will serve as the input to the subsequent channel-independent feature extraction network.

[0057] S3, Channel Independent Feature Extraction

[0058] This step aims to deeply explore the long-term temporal dependencies within each variable while maintaining independent modeling of each physical variable through a "front-end channel decoupling" strategy.

[0059] In traditional multivariate time series prediction models, cross-variable self-attention calculations (i.e., calculating the correlation weights between different sensors) are typically introduced at the shallow layers of the network. However, in the complex electromechanical systems of elevators, the physical characteristics and decay periods of different sensors are drastically different. If cross-channel mixing calculations are performed too early, the drastic variance of high-frequency signals can easily interfere with the feature extraction of low-frequency signals, causing the model to fail to capture the pure internal degradation trend of a single physical quantity.

[0060] To address the aforementioned pain point of cross-variable feature crosstalk, this embodiment employs a "separate-then-combine" mapping strategy: a channel-independent strategy is implemented in the backbone network, combined with a multi-head self-attention (MHSA) mechanism to deeply mine the long-range temporal dependencies of each independent variable. The specific derivation is as follows:

[0061] S31. Dimensional Reconstruction and Isolation of Physical Channels

[0062] Assuming the current batch of input data contains B One elevator sample. After block encoding in step S2, the system obtains the total input tensor. X in ∈ R B×C×N×Dmodel ,in C The number of dynamic time-series characteristic variables, N Number of time blocks D model For hidden layer dimensions.

[0063] To achieve independent channel computation and maximize parallel computing efficiency in engineering, this step restructures the input tensor by merging the physical variable dimension (Channel) and the batch dimension (Batch), collapsing them into an effective batch dimension: X CI = Reshape( X in ) ∈ R (B×C)×N×Dmodel Through this reconstruction operation, when the network backbone performs matrix multiplication, it will treat different sensors in the same elevator as completely independent samples, thereby achieving strict channel isolation in the mathematical dimension and preventing abnormal cross-flow of gradients between variables of different dimensions.

[0064] S32. Block-based multi-head self-attention computation For the reconstructed independent feature tensor X CI This embodiment uses a multi-layer Transformer encoder to extract its temporal features. For the first... l Layer encoder ( l = 1,2,…, L layers Its core is the multi-head self-attention (MHSA) operator.

[0065] First, the first is mapped using a linear mapping matrix. l - Output of layer 1 H (l-1) (when l When = 1, H （0） = X CI The projections are respectively onto a query matrix (denoted as Query). Q ), key matrix (Key, denoted as K) and value matrix (Value, denoted as V ): Q h (l) = H (l-1) W h,Q (l) K h (l) = H (l-1) W h,k (l) V h (l) = H (l-1) W h,V (l) in, h ∈[1, M ] indicates the first h One point of attention, M This represents the total number of heads. W h,Q , W h,k , W h,V ∈ R Dmodel×dk For learnable linear projection weights, d k = D model / M It is a single-head dimension.

[0066] Next, the attention similarity matrix between time blocks is calculated. As an optional implementation, Scaled Dot-Product Attention can be used to evaluate the dependency between two time blocks separated by any time distance. in, This is a scaling factor used to prevent the dot product result from becoming too large, which could cause the gradient of the Softmax function to vanish; the attention weight matrix generated at this time... A ∈ R N×N The physical meaning is extremely important: it measures the implicit causal relationship between the current time block and the time blocks of the past weeks or months for the same sensor.

[0067] Subsequently, the multi-head features are concatenated and linearly fused: in, This is for outputting the projection matrix.

[0068] S33, Residual Normalization and Nonlinear Mapping To avoid gradient vanishing or exploding during backpropagation in deep networks, this step introduces a residual connection around the MHSA operator. Simultaneously, considering the non-stationary nature of elevator time-series signals, as an optional implementation, this embodiment uses batch normalization (BN) for feature stabilization, replacing the traditional layer normalization standard in Transformers. Subsequently, a feed-forward network (FFN) is used for dimensionality upscaling and downscaling of the nonlinear features. The calculation process can be represented as follows: As an optional implementation, FFN can be represented as: , where GELU is the activation function, used to enhance the network's ability to express nonlinear degradation features.

[0069] S34, Output Mechanism of Dynamic Hidden Layer Representation go through L layers After layer depth extraction, the system outputs the final dynamic hidden layer representation. H out ∈ R (B×C)×N×Dmodel Finally, the system performs a reverse dimension reconstruction operation to restore it to its original state. H out_restored ∈ R B×C×N×Dmodel ( H out_restored It is the dynamic hidden layer representation tensor recovered after dimension unfolding. The reason for performing this inverse recovery operation is that in step S31, the batch and channel dimensions were merged to improve the efficiency of parallel computing; after feature extraction is completed, it must be restored to an independent dimension to ensure that different physical variables of the same elevator sample can be accurately aligned in the subsequent flattening and fusion steps.

[0070] Through the joint operator architecture of "channel independent isolation + temporal block attention + batch normalization (BatchNorm) stabilization" described above, this step endows the model with a capability similar to "independent expert consultation." Each independent attention channel acts as a "sensor expert" in a specific domain, retrospectively examining subtle signs of degradation in its own historical data without interference from other physical dimensions. Simultaneously, the introduction of BatchNorm ensures that each expert does not miss crucial amplitude mutation signals when assessing risk. This significantly improves the model's convergence stability and high-risk fault recall capability on non-stationary multivariate elevator datasets. The dynamic hidden layer representation in this step will serve as input to step S4 for subsequent flattening and recombination operations.

[0071] S4. Preservation and recombination of distribution characteristics

[0072] In the original independent channel time series prediction architecture, the network is usually equipped with a denormalization module at the output end, which aims to restore the predicted values ​​back to the true physical dimensions and output the future time series sequence through independent linear layers. However, when the task objective shifts from "numerical regression" to "fault risk classification", the above-mentioned original mechanism exposes two serious business pain points: First, direct denormalization will amplify the numerical variance of high-amplitude sensors, which can easily cause gradient oscillations or even explosions when calculating binary or multi-class cross-entropy loss, causing the model to fail to converge; Second, if the channel independence is still maintained at the output end, the model will never be able to learn "cross-variable concurrency anomalies", thus missing high-dimensional spatiotemporal interaction information.

[0073] To completely resolve the aforementioned deficiencies, this step involves innovative structural reorganization and mapping mechanism modification at the network backend. The specific steps are as follows:

[0074] S41. Removing inverse normalization and retaining normalized distribution characteristics.

[0075] Extract the dynamic hidden layer representation output in step S3 H out_restored ∈ R B×C×N×Dmodel Using a linear mapping head to divide the block dimensions N With hidden layer dimensions D model Flatten it and project it along the time dimension to the target time step length. L out Then, dimension permutation is performed to generate the dimensionality-reduced dynamic hidden layer feature tensor. X feat ∈ R B×Lout×C (in B For batch size, Lout To output the time step length or mapping length, C (Number of dynamic sensor channels).

[0076] To address the classification attributes of elevator risk warnings, this method removes and discards all inverse normalization calculation branches from the computation flow. Through this operation, X feat All feature values ​​are strictly constrained and preserved within the standardized hidden layer distribution space formed during the network's forward propagation (i.e., a stationary space with mean μ≈0 and standard deviation σ≈1). This mechanism provides an excellent numerical stability foundation for subsequent access to fully connected classification heads and the application of a binary cross-entropy loss function with logistic regression, effectively avoiding the weight contamination of the classification gradient by differences in physical dimensions.

[0077] S42, Flattening and Dimensional Reduction and Breaking Channel Independence

[0078] In order to capture the concurrent coupling relationships between variables from multiple sensor sources, this step performs a full-dimensional flattening operation in the feature space to break down the strict channel-independence barrier in the front-end network.

[0079] Specifically, the system operates along all dimensions other than the batch dimension (i.e., the time dimension). L out and channel dimensions C ) for hidden layer feature tensor X feat Perform high-dimensional flattening:

[0080] By changing the dimension from Flattened into a one-dimensional long vector The system forcibly places the historical evolution characteristics of all time steps and all physical variables on the same topological level, laying the data structure foundation for subsequent use of fully connected layers to capture the nonlinear cross weights between different variables.

[0081] S43, Architecture reservation for classification mapping High-dimensional dynamic feature vector generated after flattening X flatten It includes a spatiotemporal cross-representation of all dynamic operating signals of the elevator during the observation period. Traditional retrofitting schemes would directly feed this data into a linear classification layer for probability prediction, but this embodiment argues that relying solely on dynamic signals still lacks the "basic risk bias" of the equipment.

[0082] Therefore, in the architectural design of this method, X flattenInstead of immediately triggering Logits mapping calculation, it serves as a "full-dimensional dynamic basis" and waits in the system cache for the static attribute feature vectors that were bypassed and preserved in step S1. X s Perform the final feature-level stitching.

[0083] Through a combined innovation of "elimination and denormalization + flattening and full-dimensional reconstruction," this step successfully bridges the gap between time series prediction models and industrial-grade classification tasks. The distribution preservation mechanism ensures efficient convergence of the network when facing non-stationary elevator multi-sensor data; while the flattening and reconstruction mechanism gives the model the ability to understand the complex nonlinear disaster-causing mechanism of "concurrent degradation of multiple elevator components," significantly improving the early warning recall rate for complex fault types.

[0084] The global dynamic feature vector output by this step X flatten This will be used as input to S5 for subsequent dynamic and static feature enhancement and fusion.

[0085] S5, Enhanced Fusion of Dynamic and Static Features

[0086] This step aims to jointly diagnose the device's "basic physical condition" and "apparent symptoms" to address the technical problems of low utilization rate of static attribute features and rigid fusion methods in the background technology.

[0087] Traditional equipment failure prediction models often rely excessively on time-series fluctuation signals from sensors, neglecting the inherent physical properties of the equipment. From the perspective of the objective laws of mechanical operation and maintenance, dynamic time-series signals only reflect the current "apparent symptoms" of the equipment, while the "basic condition" of the equipment is determined by static attributes (such as a ten-year-old elevator, a large rated load, and a complex installation and usage environment). Without static attributes as a priori benchmark for evaluation, the model is prone to misjudgment: for example, even if the sensor dynamic fluctuations of a 20-year-old elevator with a large load are similar in the short term, the confidence level of its occurrence of a serious failure is completely different from that of a newly installed low-rise residential elevator.

[0088] To endow the model with this combined diagnostic capability of "device constitution" and "apparent symptoms," this embodiment innovatively introduces a dynamic and static dual-branch feature enhancement fusion mechanism at the network end. The specific steps are as follows:

[0089] S51. Dimensional Alignment and Concatenation of Heterogeneous Features

[0090] In step S4, the network has output a global dynamic feature vector that breaks channel independence and completes the flattening operation. X flatten ∈ R B×(Lout×C)This vector highly condenses the spatiotemporal evolution and cross-coupling information of all sensor signals and maintenance statistics within the current observation period.

[0091] At this point, the system retrieves the static attribute vectors that were frozen in step S1 from the front-end bypass cache. X s ∈ R B×Dstatic (in B For batch size, D static This is a dimension of static features, including numerical features such as age of the ladder, load, and number of floors / stations.

[0092] Subsequently, before entering the final prediction classification head, the system performs a direct concatenation operation on the two heterogeneous vectors along the feature dimension: X fused =Concat( X flatten , X s ,dim=-1) After concatenation, the system constructs a full-dimensional enhanced feature vector. X fused ∈ R B×(Lout×C+Dstatic) This vector perfectly unifies high-frequency timing degradation information with low-frequency inherent risk bias in terms of data structure.

[0093] S52, Numerical Stability and Feature Weight Anti-Collapse Mechanism It is worth emphasizing that the strategy of "delaying" static features to the deepest layers of the network (before the classifier) ​​in this embodiment is not only to avoid the waste of computing power in the early attention mechanism, but also to solve the gradient collapse problem when fusing heterogeneous features.

[0094] If fusion is performed directly at the input, static features (such as age) will be completely diluted by the high-frequency variance of the dynamic sensor during multiplication of multiple self-attention matrices. However, in the architecture of this embodiment:

[0095] 1) Pure dynamic extraction: Without being disturbed by static constants, the dynamic branches, through batch normalization and temporal attention, extract the purest sequence degradation representation. X flatten ).

[0096] 2) Stable fusion of units of the same dimension: This is thanks to the innovative design of forced removal of anti-normalization in step S4. Xflatten Strictly maintained within the standardized distribution space with zero mean; and X s Numerical scaling and encoding were also performed during preprocessing. This ensures that the dynamic latent vector and the static attribute vector are in very similar numerical magnitudes and variance spaces when the concatenation operation is performed.

[0097] After this splicing X fused When fed into the subsequent fully connected classification network, the gradients of backpropagation can be evenly and fairly distributed to dynamic temporal features and static attribute features, avoiding the weights of one side being "overwhelmed" by the extreme dimensions of the other side (such as unnormalized thousands of amperes of current).

[0098] Through this bypass delay fusion strategy, this step not only injects a crucial "basic risk bias" into elevator risk prediction, greatly reducing the false alarm rate caused by ignoring individual equipment differences; at the same time, it cleverly utilizes a normalized distribution preservation mechanism at the mathematical level to ensure the efficient and stable fusion of heterogeneous data (low-frequency static ledgers and high-frequency dynamic waveforms) in the same hidden layer space, laying a solid high-dimensional feature foundation for the final nonlinear probability mapping.

[0099] The full-dimensional enhanced feature vector output by this step X fused This will be used as input for step S6, for risk probability mapping and early warning.

[0100] S6, Risk Probability Mapping and Early Warning Module

[0101] This step aims to map the fused high-dimensional features into risk probabilities with engineering guidance value, and to solve the technical problems in the background technology, such as the difficulty in adapting to classification tasks and the extreme imbalance between positive and negative samples causing the model to be biased towards the majority class.

[0102] In existing time series prediction models, the network terminals often use linear layers to directly output continuous physical values, and are optimized using regression loss functions such as mean squared error (MSE). However, in elevator safety monitoring and on-demand maintenance scenarios, the core business requirement is not to accurately predict sensor values ​​at the next moment (such as specific voltage values), but to assess the confidence probability of serious equipment failures within a certain period. Furthermore, elevator failures are typical high-risk, low-probability events, with the number of normally operating samples far exceeding the number of failed samples (i.e., extreme imbalance between positive and negative samples). If conventional cross-entropy loss is used, the model's gradient updates are easily dominated by the mundane information from the massive number of normal samples, causing the model to "ignore" the few high-risk failure signals.

[0103] To bridge the business gap between "sequence numerical prediction" and "binary / multi-class risk decision-making," this embodiment performs deep classification adaptation and probability mapping at the network output. The specific implementation steps are as follows:

[0104] S61, Dimensionality Reduction and Spatiotemporal Interaction Mapping of MLP Classification Heads

[0105] The full-dimensional enhanced feature vector generated in step S5 X fused ∈ R B×(Lout×C+Dstatic) The data is fed into a classification head composed of a multilayer perceptron (MLP). This classification head acts as a "decoder" for global feature interactions. The data is then processed through a weight matrix. W cls and bias b cls The high-dimensional feature space containing both dynamic and static information is projected onto the corresponding future preset time window (i.e., the prediction step size). L pred Logits space: Y logits = X fused· W cls +b cls The output tensor dimension is now... , among them Y logits It is the original logistic score without any nonlinear probabilistic activation function compression, preserving the widest gradient descent optimization space.

[0106] S62. Weighted Cross-Entropy Optimization Mechanism Based on Logits

[0107] During the model training phase, considering the extremely long-tailed distribution of elevator fault samples, this embodiment innovatively employs a binary logistic cross-entropy loss function with positive sample penalty weights. This mechanism merges the Sigmoid activation function with binary cross-entropy at the underlying operator level, and utilizes the Log-Sum-Exp mathematical technique to greatly improve the numerical stability under high-variance Logits input, completely eliminating the gradient explosion problem that may be caused when calculating Sigmoid separately.

[0108] As an optional implementation, the loss calculation formula is derived as follows:

[0109] in, y i ∈{0,1} represents the actual device status label (0 represents normal, 1 represents fault), and σ(·) is the internally encapsulated Sigmoid function. ω pos (i.e., pos_weight, positive sample weight coefficient) is the positive sample weight coefficient calculated based on the ratio of normal samples to faulty samples in the historical dataset. By assigning extremely high gradient weights to faulty samples in the minority class, the system forces the model to invest more computing power during backpropagation to "focus on" and "amplify" those hidden high-risk abnormal features, thereby significantly improving the model's recall rate for sudden faults.

[0110] S63. Risk Probability Quantification and Generation of Level 3 Security Index

[0111] During the actual model deployment and inference phase, the system extracts the network output. Y logits It explicitly uses the Sigmoid function to nonlinearly map the probability of risk to the physical interval (0,1), and then uses the Sigmoid function to map it to the absolute risk probability. P fault :

[0112] To better align the early warning signals with the cognitive habits of frontline maintenance personnel and eliminate distribution biases in the original probability, the system performs smoothing post-processing operations such as log-normalization on the output original probability distribution. Finally, the system bases the warning signals on the smoothed absolute risk probability. P fault A multi-level security threshold system is established, which is then mapped to generate a three-level safety index for business terminals. 1) Red code (high risk): When P fault > T red When triggered, it indicates that the elevator faces a high risk of downtime in the very short future, and the system will issue the highest-level work order of "immediately stop the elevator for maintenance" in conjunction with the Internet of Things platform.

[0113] 2) Yellow code (medium risk): When T yellow < P fault ≤ T red When triggered, it indicates that the equipment is showing signs of sub-health and deterioration (such as increased wear on the door slider). The system then issues a "time-limited on-demand maintenance" instruction to guide maintenance personnel to conduct key inspections in the next inspection cycle.

[0114] 3) Green code (low risk): When P fault ≤ T yellow When triggered, it indicates that the equipment is in good operating condition, and its daily maintenance cycle can be dynamically extended, saving labor costs.

[0115] Through this risk probability mapping and multi-level early warning mechanism, this embodiment ensures the convergence stability of model training under high class imbalance at the underlying architecture level; at the top-level business logic level, the temporal hidden features of deep learning are transformed into a "red, yellow, and green" three-color safety index with extremely high engineering guidance value, which completely opens up the closed loop of the entire industrial implementation process from "IoT underlying data collection" to "intelligent prediction" and then to "front-end on-demand maintenance action guidance".

[0116] To verify the effectiveness of the method proposed in this embodiment, actual operating data from 21,407 elevators provided by a company were used for testing. The prediction results of the model for the high-risk, extremely small sample failure type "trapped" (is1) are shown in Table 1. The model used discrete data from April 1, 2022 to September 16, 2024, for a total of 60 half-monthly time steps, and strictly divided the training set, validation set, and test set according to a 7:2:1 ratio.

[0117] Regarding feature window settings, the model uses historical multi-source data from the past 24 time steps (i.e., 12 months) to predict the probability of failure in the next 6 time steps (i.e., 3 months). Historical backtesting was performed on six consecutive half-monthly nodes from early April to late June 2024 in the test set. With the red code (high-risk) warning threshold set at the top 3000 elevators in terms of predicted probability, the prediction results are shown in Table 1. Table 1. Prediction results of the model on the trapped person fault (is1)

[0118] In real-world industrial scenarios, elevator entrapment incidents are typical examples of a very small number of high-risk, "long-tail" events, where the ratio of normal cases to malfunctioning cases is often extremely disparate. Conventional time-series prediction models are prone to collapsing towards normal cases in such tasks, leading to missed detections of high-risk incidents.

[0119] However, as shown in the table above, the model proposed in this embodiment exhibits extremely high robustness and effectiveness. This system sets the "red code" high-risk warning threshold to the top 3000 elevators, which represents only about 14% of the total 21407 elevators. Under the premise of limiting the top 3000 elevators to key maintenance targets, the model maintains an exceptionally high recall capability for elevators that actually experience entrapment malfunctions. For example, on April 15, 2024, out of 100 elevators that actually experienced entrapment malfunctions, the model successfully predicted 78 elevators in advance, achieving a recall rate as high as 78.00%; in subsequent prediction periods, the recall rate remained stable between 66.30% and 75.22%.

[0120] The elevator risk early warning big data model in this embodiment comprises a perception layer, a data layer, an algorithm layer, and an application layer in its overall structure. The perception layer relies on elevator IoT sensor components, environmental and meteorological sensing nodes, and maintenance and fault reporting terminals. It utilizes IoT technology to collect multi-source heterogeneous data in real time, such as elevator operating waveforms, fault status, and meteorological scores, and combines this data with the business ledger system to obtain inherent attribute records such as equipment brand and age.

[0121] In the initial multi-source heterogeneous data aggregation stage of the data layer, the original dataset contains 114 redundant discrete fields. To ensure the effective convergence of the "temporal block coding" and "dynamic and static feature fusion" modules in the subsequent algorithm layer, and to eliminate the "curse of dimensionality" and gradient crosstalk caused by redundant features, this embodiment performs in-depth missing value analysis and business logic simplification on the massive original data during the data preprocessing stage (after processing by the algorithm layer, the application layer finally outputs red, yellow, and green three-level safety health codes). The specific data layer simplification and feature reconstruction process is as follows:

[0122] 1. Hard removal based on data quality and missing rate

[0123] Statistical analysis of missing data revealed severe deficiencies in some ledgers and detection fields. This implementation method forcibly discarded low-quality fields with missing rates exceeding 83%, such as reform date (reform_date), reform unit (gzdw), detection conclusion (detect_conclusion), detection unit (detect_DW), detection date (detect_date), next detection date (next_detect_date), and detection personnel (detect_person).

[0124] 2. Business logic dimensionality reduction based on equipment degradation mechanism:

[0125] 1) Simplification of Subject Associations: The user unit / property (sydw) field was removed because it contains hundreds of fragmented enumeration values ​​and has a weak direct impact on the mechanical health of the equipment. Instead, the maintenance unit (wbdw) field, which plays a decisive role in repairing equipment health degradation rates, was retained. Simultaneously, the county (county) geographical field, which has minimal impact on the micro-operational patterns of individual elevators, was removed.

[0126] 2) Static attribute mapping: Instead of directly inputting discrete installation date strings, the system dynamically converts them into the continuous numerical feature variable "Age", enabling deep learning models to more intuitively infer the physical differences between old and new elevators in terms of vibration, noise and failure frequency.

[0127] 3. Aggregation and semantic enhancement of dynamic temporal features:

[0128] 1) Preservation of time series covariates: For lunar holidays (such as the Spring Festival), the time encoder built into the standard time series model is difficult to automatically identify. Therefore, this system focuses on preserving the highly valuable time feature of "number of holidays" to guide the model to accurately capture abnormal traffic patterns in office buildings or residential buildings during special periods.

[0129] 2) Statistical aggregation of maintenance and meteorology: The originally complex maintenance details are compressed, extracting only the maintenance frequency statistics (num) for semi-monthly, quarterly, semi-annual, and annual periods, directly reflecting the health degradation trajectory of the equipment. At the same time, daily weather scatter data is calculated on-site within a 15-day sliding time window and aggregated into mean temperature (avg_temp), standard deviation of temperature (std_temp, representing volatility), maximum temperature (max_temp), and weather index weighted score (weather_scores), replacing the original weather fields.

[0130] After the above simplification and reconstruction, the final high-quality feature data is strictly divided into static non-temporal features (input to the bypass fusion module at the end of the network) and dynamic temporal features (input to the front-end block coding module).

[0131] The following data defines the elevator's physical limits, operating characteristics, and basic health foundation, serving as the "basic bias" for model risk assessment, as shown in Table 2: Table 2. Static non-temporal feature table extracted by the model (static portrait data)

[0132] The following type of data contains high-frequency time-series signals such as the elevator's historical status, maintenance actions, and external weather conditions. The model captures the abnormal evolution trends within these signals through time-series block coding, as shown in Table 3: Table 3. Dynamic time-series features extracted by the model (dynamic waveform data) Time Context holiday Statistics on the number of holidays within the observation window Core timing signals is1, is2, is4, is5, is6 The corresponding fault status identifier (such as trapped person, speeding, etc.) that occurred at the current time step. failureNum1, failureNum2, failureNum4, failureNum5, failureNum6 Statistics on the number of occurrences of various types of faults within the corresponding historical period Maintenance statistics halfMonth_maintained_num Bi-weekly maintenance frequency statistics season_maintained_num Quarterly maintenance frequency statistics halfYear_maintained_num Semi-annual maintenance frequency statistics Year_maintained_num Annual maintenance frequency statistics Weather aggregation avg_temp Average temperature over a preset period (15 days) std_temp The standard deviation of temperature over a preset period (15 days) represents the temperature volatility. max_temp The highest temperature within the preset period (15 days) weather_scores Weighted score of weather condition index within a preset period (15 days)

[0133] Example 2

[0134] This embodiment provides an elevator risk early warning system based on temporal block coding and dynamic / static feature fusion, used to execute the steps in the above method embodiments. The system includes:

[0135] The data acquisition and segmentation unit is used to acquire historical operating data of the elevator to be monitored, and after parsing, the historical operating data is divided into dynamic time-series features that change over time and static non-time-series features that describe the inherent properties of the elevator. The dynamic feature block encoding unit is used to normalize the dynamic temporal features and divide the normalized multivariate temporal data into multiple overlapping time blocks along the time dimension, and map them to generate multi-channel block embedding vectors. The channel-independent feature extraction unit is used to input the multi-channel block embedding vector into a preset channel-independent feature extraction network. Under the premise of maintaining the sharing of parameters of each channel and independent computation, it extracts the local semantic representation and long-range temporal dependency of each channel and outputs a dynamic hidden layer representation that maintains the normalized distribution state. The distribution feature preservation and recombination unit is used to extract the dynamic hidden layer representation. It directly flattens the dynamic hidden layer representation of all channels and time steps that have not undergone inverse normalization processing, breaks the channel independence, and generates a global dynamic feature vector containing cross-variable interaction information. The static and dynamic feature enhancement and fusion unit is used to encode the static non-temporal features and then concatenate and fuse the encoded static non-temporal features with the global dynamic feature vector in the feature dimension to construct a full-dimensional enhanced feature vector containing multi-source heterogeneous information. The risk probability mapping and early warning unit is used to perform global mapping on the full-dimensional enhanced feature vector using a fully connected classification network, capture the spatiotemporal interaction dependency between variables, calculate the predicted probability of elevator failure within a future preset time window, and generate a risk early warning signal based on the probability.

[0136] In a preferred implementation, the channel-independent feature extraction unit employs a Transformer encoder architecture. All channels share the same set of model parameters but independently perform multi-head self-attention calculations, extracting implicit causal relationships between time blocks through this multi-head self-attention mechanism. The distribution feature preservation and recombination unit breaks channel independence by globally flattening the dynamic hidden layer representation without performing inverse normalization. The risk probability mapping and early warning unit trains the model using a binary cross-entropy loss function with positive sample weights and generates different levels of risk warning signals based on comparisons between predicted probabilities and preset multi-level safety thresholds.

[0137] For the specific implementation of each functional unit of the system described in this embodiment, please refer to the relevant descriptions of S1 to S6 in the aforementioned method embodiment 1, which will not be repeated here.

[0138] Example 3

[0139] This embodiment provides an electronic device, including: Memory, used to store computer programs; The processor is used to execute the computer program stored in the memory to implement the steps of the elevator risk warning method based on time-series segmentation and dynamic-static fusion as described in any of the above-described method embodiments 1.

[0140] For details on the specific implementation of each step and related explanations, please refer to the aforementioned method implementation examples, which will not be repeated here.

[0141] The memory of the electronic device mentioned in this embodiment may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. The processor mentioned above may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0142] Example 4

[0143] This embodiment provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the elevator risk warning method based on time-series segmentation and dynamic-static fusion as described in any of the above-described method embodiments 1. For specific implementation details and explanations of each step of the method, please refer to the aforementioned method embodiment 1, which will not be repeated here.

[0144] The computer-readable storage medium can be any tangible medium capable of storing computer programs, including but not limited to: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, and other media capable of storing program code.

[0145] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. A method for elevator risk early warning based on time-series segmentation and dynamic-static fusion, characterized in that, Includes the following steps: Historical operating data of the elevator to be monitored is acquired and analyzed to divide the historical operating data into dynamic time-series features that change over time and static non-time-series features that describe the inherent properties of the elevator. The dynamic time series features are normalized, and the normalized multivariate time series data is divided into multiple overlapping time blocks along the time dimension, and mapped to generate multi-channel block embedding vectors. The multi-channel block embedding vector is input into a preset channel-independent feature extraction network to extract the local semantic representation and long-range temporal dependency of each channel, and output a dynamic hidden layer representation that maintains a normalized distribution state. Extract the dynamic hidden layer representation, and directly flatten the dynamic hidden layer representation of all channels and time steps that have not undergone inverse normalization processing to generate a global dynamic feature vector containing cross-variable interaction information. The static attribute features are encoded, and the encoded static attribute features are concatenated and fused with the global dynamic feature vector in the feature dimension to construct a full-dimensional enhanced feature vector containing multi-source heterogeneous information. The fully connected classification network is used to globally map the full-dimensional enhanced feature vector, capture the spatiotemporal interaction dependency between variables, calculate the elevator failure prediction probability within a future preset time window, and generate a risk warning signal based on the probability.

2. The elevator risk early warning method based on time-series segmentation and dynamic-static fusion according to claim 1, characterized in that, The dynamic time-series features include multivariate time-series data collected at preset time steps, while the static non-time-series features include device inherent attributes and environmental profile data.

3. The elevator risk early warning method based on time-series segmentation and dynamic-static fusion according to claim 1, characterized in that, The normalization process employs an instance normalization strategy and performs an asymmetric de-normalization process.

4. The elevator risk early warning method based on time-series segmentation and dynamic-static fusion according to claim 1, characterized in that, The channel-independent feature extraction network adopts a Transformer encoder architecture. All channels share the same set of model parameters for time series, but perform multi-head self-attention calculation independently. For each channel, the implicit causal relationship between time blocks is extracted through the multi-head self-attention mechanism to generate a dynamic hidden layer representation containing local semantics and long-range dependencies.

5. The elevator risk early warning method based on time-series segmentation and dynamic-static fusion according to claim 1, characterized in that, The flattening process involves globally flattening the dynamic hidden layer representation of dimension Batch×C×N×D along the channel dimension and the block dimension to obtain a one-dimensional global dynamic feature vector.

6. The elevator risk early warning method based on time-series segmentation and dynamic-static fusion according to claim 1, characterized in that, The splicing and fusion process involves directly splicing the global dynamic feature vector with the encoded static non-temporal features along the feature dimension to form a full-dimensional enhanced feature vector.

7. An elevator risk early warning system based on time-series segmentation and dynamic-static fusion, characterized in that, include: The data acquisition and segmentation unit is used to acquire historical operating data of the elevator to be monitored, and after parsing, the historical operating data is divided into dynamic time-series features that change over time and static non-time-series features that describe the inherent properties of the elevator. The dynamic feature block encoding unit is used to normalize the dynamic temporal features and divide the normalized multivariate temporal data into multiple overlapping time blocks along the time dimension, and map them to generate multi-channel block embedding vectors. The channel-independent feature extraction unit is used to input the multi-channel block embedding vector into a preset channel-independent feature extraction network, extract the local semantic representation and long-range temporal dependency of each channel, and output a dynamic hidden layer representation that maintains a normalized distribution state. The distribution feature preservation and recombination unit is used to extract the dynamic hidden layer representation. It directly flattens the dynamic hidden layer representation of all channels and time steps that have not undergone inverse normalization processing to generate a global dynamic feature vector containing cross-variable interaction information. The static and dynamic feature enhancement and fusion unit is used to encode the static non-temporal features and then concatenate and fuse the encoded static non-temporal features with the global dynamic feature vector in the feature dimension to construct a full-dimensional enhanced feature vector containing multi-source heterogeneous information. The risk probability mapping and early warning unit is used to perform global mapping on the full-dimensional enhanced feature vector using a fully connected classification network, capture the spatiotemporal interaction dependency between variables, calculate the predicted probability of elevator failure within a future preset time window, and generate a risk early warning signal based on the probability.

8. The system according to claim 7, characterized in that, The channel-independent feature extraction unit adopts a Transformer encoder architecture. All channels share the same set of model parameters for time series, but independently perform multi-head self-attention calculations. The implicit causal relationships between time blocks are extracted through the multi-head self-attention mechanism.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute a computer program stored in the memory to implement the steps of the elevator risk warning method based on time-series segmentation and dynamic-static fusion as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the elevator risk warning method based on time-series segmentation and dynamic-static fusion as described in any one of claims 1 to 6.

Images