A method and system for elevator data acquisition and analysis

Abstract

This invention discloses an elevator data acquisition and analysis method and system, relating to the field of intelligent elevator operation analysis technology. The method includes: acquiring multi-dimensional sensor data during elevator operation and performing data preprocessing to obtain a preprocessed data sequence; dividing the data into segments according to a preset time window, and performing embedding encoding on each data segment to obtain a corresponding sequence embedding representation; extracting the global dependencies between time segments and outputting time series feature representations; performing multi-task analysis and outputting predicted data; comparing the predicted data with a preset threshold to generate early warning information; and executing corresponding elevator maintenance operations. This invention, through multi-dimensional sensor data fusion and Transformer-based deep analysis, accurately captures the global dependencies of elevator operation, achieving early fault warning, accurate estimation of remaining life, and anomaly detection, significantly reducing false alarms and false negatives, and comprehensively improving the intelligence and safety of elevator operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for elevator operation, specifically to a method and system for elevator data acquisition and analysis. Background Technology

[0002] With urbanization and the development of high-rise buildings, elevators have become an indispensable vertical transportation tool in modern buildings, and their safety and operational reliability have attracted much attention. In recent years, with the rapid development of sensor technology, the Internet of Things, and artificial intelligence, real-time collection and analysis of elevator operation data has gradually become an important means of intelligent operation and maintenance. By acquiring multi-dimensional data such as elevator motor current, voltage, acceleration, vibration, and door operator status, and combining signal processing and machine learning models, the elevator's operating status can be monitored in real time and preliminary fault diagnosis can be performed. The application of this technology helps reduce sudden failures, improve maintenance efficiency, and promote the elevator industry's gradual transformation towards predictive maintenance and intelligentization.

[0003] However, existing elevator data acquisition and analysis methods still have significant shortcomings. First, traditional methods often rely on statistical features or simple time-series models (such as ARIMA or sliding window-based neural networks) for state assessment, making it difficult to capture global dependencies over long periods and limiting their ability to identify hidden and progressive fault modes. Second, current technologies are insufficient in the fusion analysis of multi-source heterogeneous sensor data; the temporal correlations and cross-modal features between different sensor data are often not fully utilized, resulting in insufficient accuracy and robustness in fault prediction. Third, most existing methods can only output a single result (such as whether a fault has occurred), lacking multi-task analysis capabilities and unable to simultaneously provide multi-dimensional information such as fault type, remaining life, and anomaly indicators, thus limiting the refinement of early warning and maintenance strategies.

[0004] Finally, existing technologies, after generating early warning information, lack effective linkage with actual elevator maintenance operations, and cannot dynamically adjust operating parameters or formulate targeted maintenance strategies based on different early warning levels and fault types. Therefore, existing technologies struggle to achieve truly intelligent and proactive operation and maintenance, and cannot achieve the technical effects brought about by this invention in terms of multi-dimensional feature extraction, comprehensive analysis, and early warning linkage. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention proposes an elevator data acquisition and analysis method and system.

[0006] Therefore, the technical problem solved by this invention is that existing elevator data acquisition and fault analysis methods have the following problems: they cannot effectively capture long-term sequence dependencies, have insufficient multi-source sensor data fusion capabilities, lack multi-task prediction output, and how to achieve intelligent early warning and maintenance linkage based on elevator operating status.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide an elevator data acquisition and analysis method, including acquiring multidimensional sensor data during elevator operation and performing data preprocessing on the sensor data to obtain a preprocessed data sequence;

[0009] The preprocessed data sequence is divided into data segments according to a preset time window, and each data segment is subjected to embedding encoding processing to obtain the corresponding sequence embedding representation;

[0010] The sequence embedding representation is input into a Transformer model based on a self-attention mechanism to extract the global dependencies between time segments and output a time series feature representation.

[0011] Multi-task analysis is performed based on the time series feature representation to output predicted data;

[0012] The predicted data is compared with a preset threshold, and an early warning message is generated based on the comparison result.

[0013] Perform the corresponding elevator maintenance operation based on the warning information.

[0014] As a preferred embodiment of the elevator data acquisition and analysis method of the present invention, the multidimensional sensing data includes the motor's current signal and voltage signal.

[0015] Acceleration and vibration signals of the elevator car;

[0016] The current signal of the elevator door mechanism and the door opening / closing status signal;

[0017] Position and speed signals of the floor encoder;

[0018] Temperature and operating status signals of the elevator braking device;

[0019] Noise and temperature signals in the elevator operating environment.

[0020] As a preferred embodiment of the elevator data acquisition and analysis method of the present invention, the data preprocessing includes time synchronization processing, signal denoising processing, missing value completion processing, normalization processing, and data segmentation processing.

[0021] As a preferred embodiment of the elevator data acquisition and analysis method of the present invention, the step of obtaining the corresponding sequence embedding representation includes dividing the preprocessed data sequence into data segments according to a preset time window;

[0022] For each data segment, a fixed-dimensional segment vector is generated through one-dimensional convolutional mapping and fully connected mapping;

[0023] Add time position encoding to the segment vector;

[0024] Add channel identifier codes based on the sensor channels corresponding to the segment vectors;

[0025] The sequence embedding representation is obtained by concatenating the segment vectors encoded by time position and channel identifier in chronological order.

[0026] As a preferred embodiment of the elevator data acquisition and analysis method of the present invention, the output time series feature representation includes: embedding the sequence representation into a Transformer model composed of multiple Transformer encoders in chronological order as an input sequence;

[0027] In each layer, a query vector, a key vector, and a value vector are generated from the input sequence, and multi-head self-attention operations are performed to obtain an attention output sequence.

[0028] The attention output sequence is processed by a feedforward network, and the output sequence of the current layer is obtained by residual connection and normalization.

[0029] The final encoded sequence is obtained after hierarchical concatenation.

[0030] The final encoded sequence is subjected to feature aggregation operations in the time dimension to obtain the time series feature representation.

[0031] As a preferred embodiment of the elevator data acquisition and analysis method of the present invention, the multi-task analysis includes performing fault classification analysis, remaining life estimation analysis and anomaly detection analysis based on the time series feature representation.

[0032] The fault classification analysis outputs a fault probability vector for a preset set of fault types.

[0033] The remaining lifetime estimation analysis outputs an estimated remaining lifetime value;

[0034] The anomaly detection and analysis outputs an anomaly index.

[0035] The predicted data includes a failure probability vector, an estimated remaining lifetime, and an anomaly index.

[0036] As a preferred embodiment of the elevator data acquisition and analysis method of the present invention, the generation of early warning information includes: comparing the fault probability vector, remaining life estimate and anomaly index in the predicted data with their respective preset thresholds to determine the corresponding early warning level and target fault type, and generating early warning information containing the early warning level and target fault type.

[0037] As a preferred embodiment of the elevator data acquisition and analysis method of the present invention, wherein: the step of performing corresponding elevator maintenance operations based on the early warning information includes selecting and performing corresponding elevator maintenance operations based on the early warning level and target fault type in the early warning information, wherein the elevator maintenance operations include:

[0038] When the warning information indicates that maintenance is required, a maintenance prompt message containing the target fault type is sent to the maintenance personnel.

[0039] When the warning information indicates that operation adjustment is required, the elevator operation parameters are adjusted, including the operating speed, maximum load and the range of floors to stop at.

[0040] When the warning message indicates that the elevator needs to be stopped, elevator stop control will be executed;

[0041] The warning information and corresponding operational data will be stored for subsequent fault tracing and model optimization.

[0042] Secondly, embodiments of the present invention provide an elevator data acquisition and analysis system, comprising:

[0043] Data acquisition and preprocessing module: Acquires multi-dimensional sensor data during elevator operation and preprocesses the sensor data to obtain a preprocessed data sequence;

[0044] Sequence segmentation module: Divides the preprocessed data sequence into data segments according to a preset time window, performs embedding encoding processing on each data segment, and obtains the corresponding sequence embedding representation;

[0045] Feature extraction module: Inputs the sequence embedding representation into the Transformer model based on self-attention mechanism, extracts the global dependencies between time segments through multi-head attention operation, and outputs the time series feature representation;

[0046] Multi-task analysis module: Performs multi-task analysis based on the time series feature representation and outputs predicted data;

[0047] Early warning generation module: compares the predicted data with a preset threshold and generates early warning information based on the comparison result;

[0048] Maintenance Operation Module: Executes corresponding elevator maintenance operations based on the warning information. Beneficial Effects of the Invention: By integrating multi-dimensional sensor data with a Transformer deep analysis model based on a self-attention mechanism, this invention can accurately capture global dependencies in elevator operation data, achieving multi-task predictive analysis of fault types, remaining lifespan, and abnormal states. Combined with threshold judgment to generate multi-level warning information, and linked with maintenance strategies, it can dynamically adjust operating parameters and accurately prompt maintenance needs, thereby significantly improving the accuracy and timeliness of fault prediction, reducing the risk of sudden failures, and enhancing the intelligence and reliability of elevator operation and maintenance. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0050] Figure 1 The first embodiment of the present invention provides an overall flowchart of an elevator data acquisition and analysis method. Detailed Implementation

[0051] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0052] Example 1, referring to Figure 1 As an embodiment of the present invention, an elevator data acquisition and analysis method is provided, comprising:

[0053] S1: Collect multi-dimensional sensor data during elevator operation and preprocess the sensor data to obtain a preprocessed data sequence.

[0054] Multidimensional sensing data includes the motor's current and voltage signals;

[0055] Acceleration and vibration signals of the elevator car;

[0056] The current signal of the elevator door mechanism and the door opening / closing status signal;

[0057] Position and speed signals of the floor encoder;

[0058] Temperature and operating status signals of the elevator braking device;

[0059] Noise and temperature signals in the elevator operating environment.

[0060] In the data section of step S1, the collected data include motor current and voltage signals, car acceleration and vibration signals, door mechanism current and door opening / closing status signals, floor encoder position and speed signals, braking device temperature and operating status signals, and noise and temperature signals from the elevator operating environment. To ensure the distinguishability and alignment of subsequent analyses, each channel is recorded with a unified timestamp and includes metadata such as sampling rate, range, unit, and channel identifier. Motor-side signals can be taken from the inverter output side or the motor side, and anti-aliasing filtering is configured before sampling to ensure the range covers the safety margin of the rated value. Acceleration / vibration sensors are installed near the car bottom frame or guide shoes, and the installation direction and sensitivity coefficient are recorded. Door mechanism current and door position status are collected synchronously, and the rise / fall time of the switching quantity is marked. Encoder data is directly read from the position pulse and speed estimate or only the pulse time interval is recorded. Braking temperature and operating status are provided by temperature probes and contact signals, respectively. Environmental noise and temperature sensors are arranged in the machine room or shaft, and the measurement point locations are marked. Continuous quantities (such as current, voltage, vibration, and noise) are recommended to be acquired at a sampling rate that sufficiently covers the target frequency band (e.g., motor current / voltage at kHz, vibration / acceleration at kHz, noise at 8–16 kHz, and temperature at 1–2 Hz). Discrete / event quantities (gate position, braking action) are recorded with event-timestamped data. The acquisition end simultaneously records overflow and communication anomaly flags to avoid subsequent misjudgments. The above data directly constitutes the "raw multi-channel time-series data" as input for preprocessing. Here, kHz stands for kilohertz, and Hz stands for hertz.

[0061] Data preprocessing includes time synchronization, signal denoising, missing value completion, normalization, and data segmentation.

[0062] In the preprocessing section of step S1, time synchronization is first performed: using the controller's master clock or network clock as a reference, the timestamps of each channel are drift-corrected, continuous signals are uniformly resampled to a common time base (e.g., 1 kHz), and discrete / event signals are inserted into the common time axis according to the timestamp alignment, thereby ensuring that cross-channel evidence such as "gate current peak - gate opening / closing edge" and "deceleration vibration - velocity curve" correspond one-to-one in the time domain. Subsequently, signal denoising is performed: current / voltage is de-DC and slowly drifted, and notch filtering or bandpass filtering is applied to the power frequency / switching frequency; acceleration / vibration uses mechanical correlation bandpass and robust methods to suppress spikes; noise signals are pre-emphasized and bandpassed to enhance abnormal noise characteristics; position / velocity signals are differentially smoothed and unified; gate position and braking status are uniformly soft-deboiled to eliminate contact jitter. For unavoidable sampling omissions, processing is differentiated based on duration: short-term omissions (e.g., less than 1 second) are handled using interpolation or Kalman smoothing for continuous quantities and time-limited forward hold for state quantities; long-term omissions are not backfilled, but only marked as missing for subsequent analysis to reduce weight or remove. Normalization is then performed: standardized parameters are independently calculated for each channel and z-score (standard score) or robust normalization is applied. These parameters are bound to the device and archived with the data to mitigate differences in dimensions and scales and improve cross-device comparability. Finally, data is segmented: windows are preferentially segmented based on event boundaries such as start / stop, acceleration, deceleration, leveling, door opening / closing, and braking actions; when event information is unavailable, fixed-length sliding windows and step sizes are used to form sequence segments, and the start and end times, event labels, and omission percentage are recorded for each window. After the above processing, a multi-channel "preprocessed data sequence" is obtained, arranged by window under a unified time axis and a unified sampling rate, along with a channel dictionary, normalization parameters, missing / abnormal markers, and event index. This sequence is the direct input for the subsequent embedding and encoding steps.

[0063] S2: Divide the preprocessed data sequence into data segments according to a preset time window, and perform embedding encoding processing on each data segment to obtain the corresponding sequence embedding representation.

[0064] The preprocessed data sequence is divided into data segments according to a preset time window;

[0065] For each data segment, a fixed-dimensional segment vector is generated through one-dimensional convolutional mapping and fully connected mapping;

[0066] Add time position encoding to the segment vector;

[0067] Add channel identifier codes based on the sensor channels corresponding to the segment vectors;

[0068] The sequence embedding representation is obtained by concatenating the segment vectors encoded by time position and channel identifier in chronological order.

[0069] In step S2, the preprocessed data sequence is first segmented using a preset time window to form data segments arranged in chronological order. The length and step size of the time window are set according to the typical duration and frequency band requirements of elevator operation events, ensuring that each data segment is statistically approximately stationary and that adjacent segments have moderate overlap to retain transition information. For tail segments that cannot be divided evenly, zero-padding or truncation can be used, and the padding positions are marked as missing in subsequent encoding to avoid introducing false features. Each segmented data segment maintains a multi-channel structure and a unified time base, thus providing consistent input for subsequent isomorphic vectorization processing.

[0070] Subsequently, embedding encoding is performed on each data segment to obtain a fixed-dimensional segment vector. Specifically, a one-dimensional convolutional mapping is first applied to the segment along the temporal dimension to extract local temporal patterns and perform dimensionality reduction compression. The size of the convolutional kernel is equal to the key dynamic scale within the segment covered by the receptive field, and the stride is kept to 1 or consistent with the sampling interval within the window to avoid losing fine-grained variations. The convolutional output is activated and flattened before being fed into a fully connected mapping to project local features of different channels and temporal positions onto a unified fixed-dimensional space, forming the segment vector corresponding to the data segment. For missing markers retained by S1, the corresponding positions are processed with zero weight or masking during the convolution and fully connected stages to prevent missing points from shifting the segment vector. To explicitly encode temporal and channel priors, temporal position encoding and channel identifier encoding are added to the segment vector: temporal position encoding is used to characterize the relative temporal position of the segment within the window, and channel identifier encoding is used to characterize the sensor channel category corresponding to the segment vector. Both types of encoding are linearly projected to the same dimension as the segment vector and then fused by vector addition to ensure consistency in the numerical space and to avoid introducing new dimensions. Finally, the segment vectors encoded by time position and channel identifier are concatenated in their natural order along the time axis to form a sequence embedding representation that corresponds one-to-one with the original window. This sequence embedding representation preserves the temporal relationships and channel semantics at the segment level and can be directly used as input for subsequent steps to extract global dependencies across segments. Through the above processing, the original multi-channel temporal series is normalized into a vector sequence with consistent length, fixed dimensions, and clear semantics, which not only preserves local dynamic features but also provides a comparable and verifiable representation basis for subsequent modeling.

[0071] S3: Input the sequence embedding representation into the Transformer model based on the self-attention mechanism, extract the global dependencies between time segments, and output the time series feature representation.

[0072] The sequence embedding representation is fed into a Transformer model consisting of a multi-layer Transformer (Transformer architecture, a sequence modeling framework based on self-attention) encoder in chronological order as the input sequence.

[0073] In each layer, a query vector, a key vector, and a value vector are generated from the input sequence, and multi-head self-attention operations are performed to obtain an attention output sequence.

[0074] The attention output sequence is processed by a feedforward network, and the output sequence of the current layer is obtained by residual connection and normalization.

[0075] The final encoded sequence is obtained after hierarchical concatenation.

[0076] The final encoded sequence is subjected to feature aggregation operations in the time dimension to obtain the time series feature representation.

[0077] In step S3, the sequence embedding representation obtained in step S2 is fed into the model consisting of a multi-layer Transformer encoder in chronological order as the input sequence. Each layer first generates query vectors, key vectors, and value vectors from the input sequence through linear mapping, and uses multi-head self-attention with scaled dot products to calculate the correlation weights between time segments. To avoid interference from invalid samples on the correlation measurement, two types of masking are introduced before the attention score enters the Softmax (soft maximum / softmax function, normalized exponential function): one is the padding mask generated by sequence length alignment, and the other is the masking matrix generated based on the missing and abnormal labels output in step S1. The attention score at the corresponding position is suppressed or set to zero, so that pseudo-correlation caused by sampling interruption, abnormal pulses, etc., does not participate in aggregation. The multi-head attention output is nonlinearly mapped through a feedforward network and combined with residual connections and layer normalization to form the output sequence of the current layer; after multi-layer concatenation, the final encoded sequence is obtained. Considering that step S2 has superimposed time location encoding and channel identifier encoding on the segment vector, the encoder can distinguish between different time locations and different sensor channels when generating query / key vectors, thereby reflecting the differentiated attention to key time periods and key channels in attention weight learning.

[0078] To simultaneously characterize both short-term impacts and long-term degradation during elevator operation, this step employs relative position bias in the multi-head self-attention model to enhance sensitivity to local neighborhoods, while retaining the attention head without distance penalty to capture long-range dependencies across windows. The outputs of both types of heads are fused within the same layer, enabling the model to respond to local coupling relationships such as "current peak at door closing and door position edge" and "vibration and velocity curve during deceleration," while also remaining aware of slow evolution patterns such as "slow drift of motor current root mean square and trend changes in acceleration variance." Finally, feature aggregation operations (such as weighted aggregation or pooling) are performed on the encoded sequence obtained from multi-layer concatenation in the time dimension, outputting a fixed-dimensional time-series feature representation as direct input for subsequent multi-task analysis. Compared with existing schemes that rely on sliding statistics or shallow time series models, the key technical point of this step is: to construct an attention mask using the missing and anomaly markers generated in step S1, and to combine it with position / channel priors and head functional differentiation, so that the temporal alignment of cross-channel evidence and the joint modeling of long and short-term dependencies can be completed within the same coding framework, thereby improving the robustness and discriminability of feature representation without changing the terminology of the claims.

[0079] S4: Perform multi-task analysis based on the time series feature representation and output predicted data.

[0080] Fault classification analysis, remaining lifetime estimation analysis, and anomaly detection analysis are performed based on the time series feature representation.

[0081] The fault classification analysis outputs a fault probability vector for a preset set of fault types.

[0082] The remaining lifetime estimation analysis outputs an estimated remaining lifetime value;

[0083] The anomaly detection and analysis outputs an anomaly index.

[0084] The predicted data includes a failure probability vector, an estimated remaining lifetime, and an anomaly index.

[0085] In step S4, the time-series feature representation output from step S3 is used as a unified input, and fault classification analysis, remaining life estimation analysis, and anomaly detection analysis are performed sequentially (or in parallel) to obtain usable quantitative results for subsequent threshold comparison and maintenance decisions. Specifically, fault classification analysis uses a preset set of fault types as the category space to discriminate the time-series feature representation and outputs a fault probability vector arranged in the order of the set. To ensure numerical comparability, the sum of each component of the probability vector is normalized, and the output corresponding to the same window is lightly smoothed with adjacent windows to suppress instantaneous fluctuations. Remaining life estimation analysis generates non-negative remaining life estimates on the same time-series feature representation. For the same elevator instance, monotonicity constraints and anomaly corrections are performed within continuous windows to ensure that the estimated value shows a non-increasing trend as operation progresses and to eliminate occasional fluctuations. When the fault probability vector increases significantly in a certain type, a consistency adjustment can be applied to the remaining life estimate to make the two semantically coordinated. Anomaly detection analysis is based on a measure of the difference between time series feature representation and normal operation reference mode, outputting a dimensionless anomaly index. This index is robustened between adjacent windows to ensure that it remains sensitive to abnormal upward trends under noise disturbances.

[0086] To ensure the engineering usability of the three results, this invention introduces consistency verification and numerical range constraints at the result level: when the fault type corresponding to the largest component of the fault probability vector remains stable within multiple windows and the anomaly index increases synchronously, the high-confidence judgment of that type is maintained; when the remaining lifetime estimate shows a short-term anomaly that contradicts the trend of the probability vector, it is corrected through neighborhood weighting to avoid the influence of single-window extreme values ​​on the overall judgment. All of the above processing is completed in the output space of the time series feature representation without changing the data structure and terminology of the preceding steps, ensuring the closure and verifiability of the technical chain. Finally, the fault probability vector, the remaining lifetime estimate, and the anomaly index together constitute the prediction data, serving as the direct input for threshold comparison and early warning information generation in step S5; among them, the probability vector provides typological discrimination, the remaining lifetime estimate provides time margin quantification, and the anomaly index provides a sensitive indication of unknown or early anomalies. The synergy of these three provides sufficient numerical basis and interpretive foundation for subsequent early warning classification and maintenance selection.

[0087] S5: Compare the predicted data with a preset threshold, and generate early warning information based on the comparison result.

[0088] By comparing the fault probability vector, remaining lifetime estimate, and anomaly index in the predicted data with their respective preset thresholds, the corresponding warning level and target fault type are determined, and warning information containing the warning level and target fault type is generated.

[0089] In step S5, the predicted data output from step S4 is used as input, including a fault probability vector arranged according to a preset fault type set, a non-negative remaining life estimate, and a dimensionless anomaly index. First, corresponding preset threshold groups are set for the three types of quantities: the probability thresholds for each fault type are set in three levels (Level 1, Level 2, and Level 3); the remaining life thresholds are set in three levels (Level 1, Level 2, and Level 3) according to the principle that "the smaller the value, the higher the risk"; and the anomaly thresholds are set in three levels (Level 1, Level 2, and Level 3) according to the principle that "the larger the value, the higher the risk". The system completes comparisons within the same time window: each component of the fault probability vector is compared with its corresponding level's probability threshold to obtain the triggering result on the probability side; the remaining life estimate is compared with the life threshold level by level to obtain the triggering result on the life side; and the anomaly index is compared with the anomaly threshold level by level to obtain the triggering result on the anomaly side. The warning level is then determined by combining the above three results according to the preset grading rules: if any of the third-level conditions are met, it is determined to be level three; otherwise, if at least two of the second-level conditions are met simultaneously, it is determined to be level two; otherwise, if any of the first-level conditions are met, it is determined to be level one; if none of them are met, no warning is generated.

[0090] The determination of the target fault type maintains consistency with the warning level: When the probability-based trigger is at the currently determined level, the component that reaches the probability threshold of that level is selected from the fault probability vector, sorted from largest to smallest probability value, and the fault type corresponding to the largest probability value is taken as the target fault type; when the probability-based trigger is not triggered but the warning is triggered by the life-cycle or anomaly side, the fault type corresponding to the largest component in the fault probability vector is still selected as the target fault type, ensuring that the warning information has a clear maintenance direction. To suppress false alarms caused by instantaneous fluctuations, the system superimposes consistency constraints and hysteresis strategies on the above judgments: before generating a warning, the triggering conditions must be repeatedly met within adjacent preset number windows for it to take effect; when deactivating a warning, the triggering conditions must not be met within adjacent preset number windows for it to be deactivated, or a numerical hysteresis of entry / exit is formed by using a method where the triggering threshold and deactivation threshold are not exactly the same. In the case where multiple fault types simultaneously meet the probability threshold within the same window, the system uses a preset priority (e.g., combined with historical occurrence frequency or maintenance strategy priority) or the one with the longer continuous triggering duration as arbitration to ensure that the target fault type is unique and executable.

[0091] The final generated early warning information includes at least the warning level and the target fault type, and may include triggering criteria (such as the category of the triggered indicator and its comparison with the corresponding threshold) and time interval, so that the corresponding elevator maintenance operation can be directly selected and executed in subsequent steps. Through the hierarchical judgment link of "multi-indicator - multi-level - consistency / hysteresis", step S5 transforms the model output into a stable and operable alarm result: the probability vector provides typified evidence, the lifetime estimation provides time margin, and the anomaly degree provides sensitive prompts for unknown or early anomalies. The organic integration of the three at the rule level ensures that the early warning does not overly rely on the instantaneous extreme value of a certain indicator, and can promptly upgrade the level when the risk is significant, thus seamlessly connecting with subsequent maintenance decisions.

[0092] S6: Perform the corresponding elevator maintenance operation based on the warning information.

[0093] Based on the warning level and target fault type in the warning information, select and execute the corresponding elevator maintenance operation, which includes:

[0094] When the warning information indicates that maintenance is required, a maintenance prompt message containing the target fault type is sent to the maintenance personnel.

[0095] When the warning information indicates that operation adjustment is required, the elevator operation parameters are adjusted, including the operating speed, maximum load and the range of floors to stop at.

[0096] When the warning message indicates that the elevator needs to be stopped, elevator stop control will be executed;

[0097] The warning information and corresponding operational data will be stored for subsequent fault tracing and model optimization.

[0098] In step S6, the warning information generated in step S5 is used as input. This warning information includes at least a warning level and a target fault type. In this embodiment, the warning information is first parsed according to a one-to-one corresponding handling strategy table: (warning level, target fault type) is mapped to a set of elevator maintenance operations to be executed. The operation types are limited to three categories—maintenance notification, operating parameter adjustment, and elevator stop control. To ensure the safety and verifiability of the handling, a timing judgment is performed before execution: if the elevator is in the process of traveling, the operating parameter adjustment takes effect only after the most recent leveling is completed and the door is closed; if the handling is elevator stop control, the elevator is stopped after the nearest leveling is completed and the door is opened, avoiding stopping at a non-leveling position. The above handling results are recorded along with a timestamp, forming a traceable handling link with the warning information.

[0099] The adjustment of operating parameters is only implemented for the three types of parameters defined in this invention, and follows the principle of monotonic load reduction. Specifically, the nominal operating speed is adjusted accordingly. Nominal load limit With nominal docking layer set Mapped to reduced configuration Among them, the speed limit is calculated using a proportional coefficient. Set as Load limits are calculated using a proportional coefficient. Set as The docking level range is constrained by set constraints. The proportionality coefficient and set constraints are jointly determined by (warning level, target fault type): for example, when the target fault type involves a door mechanism, tightening is prioritized. To reduce the number of door opening and closing cycles, and can moderately reduce To reduce gate dynamics; when the target fault type involves the braking device, priority should be given to reducing... and To reduce braking thermal load. To avoid abrupt changes in passenger experience, parameter adjustments employ a step-by-step activation and gradual rollback: while the warning is in effect, tightening can only be resumed after at least one full travel; upon warning removal, the parameters will be adjusted according to the set recovery step size. Gradual recovery to .

[0100] The maintenance notification is used to link the maintenance process: The system generates structured notification content based on the early warning information, including the early warning level, target fault type, trigger time interval, and the corresponding window identifier (consistent with the predicted data window in step S4), and includes the handling configuration for this operation. Whether the notification was sent, the first reading time, and the confirmation time are all recorded for future reference. For elevator stop control, the processing sequence is as follows: restrict new elevator call acceptance → leveling at the nearest floor → decelerating to zero speed and completing leveling → opening doors and guiding passengers out → putting the elevator into stop state; elevator stop cancellation is only executed when the cancellation conditions in step S5 are met and maintenance confirmation is completed, and the execution sequence is consistent with parameter restoration.

[0101] To suppress frequent switching caused by instantaneous fluctuations, step S6 maintains consistency and hysteresis with step S5: maintenance operations are initiated only when the warning trigger conditions are met consecutively; parameters are restored by gradient rollback after the release conditions are met consecutively. The entire process simultaneously archives the warning information, handling type, parameter differences before and after adjustment, and corresponding timestamps and window indices as input for subsequent fault tracing and model optimization. Through this process, step S6 reliably transforms "warning information" into "executable and traceable maintenance operations," achieving closed-loop control from identification to classification to handling to rollback without introducing new terminology.

[0102] Example 2 is the second embodiment of the present invention, which differs from the previous embodiment in that:

[0103] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art or the current technical solution, can be embodied in the form of a software product. This current computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0105] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0106] Example 3, an embodiment of the present invention, provides an elevator data acquisition and analysis system, including a data acquisition and preprocessing module, a sequence segmentation module, a feature extraction module, a multi-task analysis module, an early warning generation module, and a maintenance operation module.

[0107] Data acquisition and preprocessing module: Acquires multi-dimensional sensor data during elevator operation and preprocesses the sensor data to obtain a preprocessed data sequence;

[0108] Sequence segmentation module: Divides the preprocessed data sequence into data segments according to a preset time window, performs embedding encoding processing on each data segment, and obtains the corresponding sequence embedding representation;

[0109] Feature extraction module: Inputs the sequence embedding representation into the Transformer model based on self-attention mechanism, extracts the global dependencies between time segments through multi-head attention operation, and outputs the time series feature representation;

[0110] Multi-task analysis module: Performs multi-task analysis based on the time series feature representation and outputs predicted data;

[0111] Early warning generation module: compares the predicted data with a preset threshold and generates early warning information based on the comparison result;

[0112] Maintenance Operation Module: Executes corresponding elevator maintenance operations based on the warning information.

[0113] Example 4 is an embodiment of the present invention, which provides an elevator data acquisition and analysis method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation / comparative experiments.

[0114] The test subjects were seven passenger elevators (codes E-101, E-203, E-307, E-412, E-518, E-623, and E-707) within the same commercial complex. These elevators had similar models and load capacities, but varied in usage intensity, with average monthly trips ranging from 11,000 to 16,000. The data acquisition system used the controller clock as a unified time base and included sampling of motor-side current / voltage, a three-axis accelerometer on the car's bottom frame, door mechanism current and door position status, floor encoder position / speed, braking device temperature and operation status, and machine room ambient noise and temperature. Continuous data was collected at sampling rates covering the target frequency band (kHz for motors, 1–5 kHz for vibration, 8–16 kHz for noise, and 1–2 Hz for temperature). Discrete event data were recorded with top / bottom timestamps. To avoid subsequent misinterpretations, overflow and communication anomaly flags were recorded simultaneously at the acquisition end. The data acquisition period was a full calendar month, with data packaged daily to ensure traceability.

[0115] After entering the preprocessing pipeline, the data is first synchronized and resampled to a common 1 kHz time axis. Then, a channel-based denoising strategy (power frequency / switching frequency notch filtering, mechanical bandpass filtering, spike suppression, door position and braking debouncing) is implemented. Short-term missing data is interpolated or Kalman smoothed, while long-term interruptions are simply marked as missing. Each channel is independently standardized to eliminate scaling differences between different elevators and sensors. The segmentation strategy prioritizes event boundaries (start / stop, acceleration / deceleration, leveling, door opening / closing, braking actions), degenerating into a fixed-length sliding window (30 s, step size 2–5 s) when the event is unavailable. The preprocessing output forms a multi-channel sequence arranged by window on a unified time axis, carrying a channel dictionary, standardized parameters, missing / anomaly markers, and event indexes, directly usable for embedding encoding.

[0116] In the embedding encoding stage, the multi-channel sequences within each window are extracted and compressed using one-dimensional convolution, and then mapped into fixed-dimensional segment vectors via a fully connected layer. To preserve temporal and channel semantics, temporal position encoding and channel identifier encoding are superimposed on the segment vectors and concatenated in chronological order to form a sequence embedding representation. Subsequently, the sequence embedding representation is fed into a model consisting of multiple Transformer encoders, with each layer performing multi-head self-attention, feedforward network, residual, and layer normalization. The attention stage introduces padding masks and masks derived from missing / anomaly markers to avoid spurious correlations caused by sampling interruptions and tail padding from participating in aggregation. To balance short-term impacts (such as current peaks and gate position edges at the moment of door closing) and long-term degradation (such as slow RMS drift of motor current), the attention head adopts a differentiated design with both relative position bias and no bias. After multi-layer concatenation, learnable readout vectors are aggregated in the time dimension to output a fixed-dimensional time-series feature representation.

[0117] In the multi-task analysis phase, the same time-series feature representation generates three results: a fault probability vector for a preset set of fault types, a non-negative remaining life estimate, and a dimensionless anomaly index. At the result level, single-window extreme values ​​are smoothed in the time domain and consistency is checked to ensure numerical stability. In the threshold comparison and grading phase, the three results are compared with a preset threshold set to comprehensively determine the warning level and provide the target fault type, triggering a hysteresis strategy to avoid jitter. Regarding maintenance coordination, based on the warning information, the system selects to execute maintenance notifications, adjust operating parameters (speed, load, and floor range), or stop the elevator. Control actions take effect after leveling and door handling are completed to avoid safety risks caused by stopping the elevator at non-leveling locations. Simultaneously, warning information, handling types, and parameter changes are recorded and archived for subsequent traceability and model optimization. Specific data can be found in Table 1.

[0118] Table 1 Data Reference Table

[0119]

[0120] As shown in the "Detection Lead Time" row of Table 1, the traditional threshold scheme has an average monthly lead time of approximately 3.7 hours for seven elevators, while the method of this invention improves it to approximately 26.0 hours, an average increase of approximately 22.3 hours, demonstrating a 6–7 times improvement in early warning capability. This advantage stems from the unified modeling of sequence embedding and multi-head self-attention for long-term dependence and cross-channel coupling: for example, with a lead time of 31.1 hours for E-412, the maintenance team can schedule maintenance during the off-peak hours at night, directly reducing the impact on passengers during the daytime. In terms of false alarm rate, the traditional scheme averages approximately 8.9%, while this invention reduces it to approximately 2.3%, an absolute reduction of approximately 6.6 percentage points; the false alarm rate is reduced from approximately 20.5% to approximately 7.3%, an absolute reduction of approximately 13.2 percentage points. The significant reduction in false alarms is due to the mutual verification mechanism between multi-task outputs (consistency verification of classification probability, remaining lifetime, and anomaly degree in the temporal neighborhood) and the masking of missing / anomaly locations at the attention level, reducing spurious correlations caused by transient noise or sampling interruptions; the reduction in false negatives is related to the global attention capturing of the composite pattern of "slow drift + occasional shocks", which can identify degradation trends in the early stage before the threshold is crossed.

[0121] The overall indicators also reflect stable benefits: F1 improved from a mean of 0.59 to 0.80, and AUROC improved from 0.76 to 0.91, indicating a significant enhancement in separability within the threshold range. Warning stability (measured by window-level jitter rate) decreased from approximately 15.5% to approximately 5.6%, consistent with the hysteresis strategy and consistency constraints in step S5, avoiding frequent alarm opening and closing caused by short-term fluctuations. In terms of engineering benefits, the average elevator downtime decreased from 5.43 hours / month to 2.20 hours / month, a reduction of approximately 59%; the number of parameter adjustments decreased from 20.1 times / month to 10.3 times / month, indicating that operational strategy adjustments are more "fewer and more accurate," reducing unnecessary interference with the passenger experience. The maintenance work order hit rate increased from 58.5% to 81.0%, which is directly related to the improved consistency of the "target fault type" indicated by the early warning; the number of passenger complaints decreased from 8.29 per month to 3.43 per month, reflecting that under the same or stricter safety strategies, service availability and comfort have been improved simultaneously.

[0122] Compared to existing technologies, the innovation of this invention is not a single-dimensional model replacement, but a full-link optimization from data evidence chain construction (synchronization, segmentation, and label preservation), embedding representation (temporal location and channel identification), attention modeling (masking and head differentiation) to multi-task consistency and threshold grading. Data shows that this integrated "data-representation-discrimination-processing" design achieves consistent improvements across multiple independent objects (seven elevators) and multiple indicators, and the improvement is not a marginal adjustment but a systematic improvement that transcends statistical significance. Therefore, it can be concluded that this invention has significant inventiveness and novelty compared to existing solutions in terms of early warning capability, robustness, and engineering feasibility, and possesses a reproducible quantitative advantage.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for elevator data acquisition and analysis, characterized in that, include: Multidimensional sensor data during elevator operation is collected, and the sensor data is preprocessed to obtain a preprocessed data sequence. The preprocessed data sequence is divided into data segments according to a preset time window, and each data segment is subjected to embedding encoding processing to obtain the corresponding sequence embedding representation; The sequence embedding representation is input into a Transformer model based on a self-attention mechanism to extract the global dependencies between time segments and output a time series feature representation. Multi-task analysis is performed based on the time series feature representation to output predicted data; The predicted data is compared with a preset threshold, and an early warning message is generated based on the comparison result. Perform the corresponding elevator maintenance operation based on the warning information; The process of obtaining the corresponding sequence embedding representation includes dividing the preprocessed data sequence into data segments according to a preset time window. For each data segment, a fixed-dimensional segment vector is generated through one-dimensional convolutional mapping and fully connected mapping; Add time position encoding to the segment vector; Add channel identifier codes based on the sensor channels corresponding to the segment vectors; The sequence embedding representation is obtained by concatenating the segment vectors encoded by time position and channel identifier in chronological order. The output time series feature representation includes feeding the sequence embedding representation into a Transformer model composed of multiple Transformer encoders in chronological order as an input sequence. In each layer, a query vector, a key vector, and a value vector are generated from the input sequence, and multi-head self-attention operations are performed to obtain an attention output sequence. The attention output sequence is processed by a feedforward network, and the output sequence of the current layer is obtained by residual connection and normalization. The final encoded sequence is obtained after hierarchical concatenation. The final encoded sequence is subjected to feature aggregation operations in the time dimension to obtain the time series feature representation; In the multi-head self-attention operation, two types of masking are applied: an attention head with relative position bias and an attention head without distance penalty. The outputs of the two types of attention heads are merged in the same layer. The two types of masking are: one is the padding position masking generated by sequence length alignment, and the other is the masking matrix constructed based on missing and abnormal labels. Before normalization, the attention scores of the masked positions are processed by suppression and zeroing. The multi-task analysis includes fault classification analysis, remaining lifetime estimation analysis, and anomaly detection analysis based on the time series feature representation. The fault classification analysis outputs a fault probability vector for a preset set of fault types. The remaining lifetime estimation analysis outputs an estimated remaining lifetime value; The anomaly detection and analysis outputs an anomaly index. The predicted data includes a failure probability vector, an estimated remaining lifetime, and an anomaly index.

2. The elevator data acquisition and analysis method as described in claim 1, characterized in that, The multidimensional sensing data includes the motor's current signal and voltage signal; Acceleration and vibration signals of the elevator car; The current signal of the elevator door mechanism and the door opening / closing status signal; Position and speed signals from the floor encoder; Temperature and operating status signals of the elevator braking device; Noise and temperature signals in the elevator operating environment.

3. The elevator data acquisition and analysis method as described in claim 2, characterized in that, The data preprocessing includes time synchronization processing, signal denoising processing, missing value completion processing, normalization processing, and data segmentation processing.

4. The elevator data acquisition and analysis method as described in claim 1, characterized in that, The process of generating early warning information includes comparing the fault probability vector, remaining lifetime estimate, and anomaly index in the predicted data with their respective preset thresholds to determine the corresponding early warning level and target fault type, and generating early warning information containing the early warning level and target fault type.

5. The elevator data acquisition and analysis method as described in claim 4, characterized in that, The step of performing corresponding elevator maintenance operations based on the early warning information includes selecting and performing corresponding elevator maintenance operations based on the early warning level and target fault type in the early warning information. The elevator maintenance operations include: When the warning information indicates that maintenance is required, a maintenance prompt message containing the target fault type is sent to the maintenance personnel. When the warning information indicates that operation adjustment is required, the elevator operation parameters are adjusted, including the operating speed, maximum load and the range of floors to stop at. When the warning message indicates that the elevator needs to be stopped, elevator stop control will be executed; The warning information and corresponding operational data will be stored for subsequent fault tracing and model optimization.

6. An elevator data acquisition and analysis system, used to implement the elevator data acquisition and analysis method as described in any one of claims 1 to 5, characterized in that, include: Data acquisition and preprocessing module: Acquires multi-dimensional sensor data during elevator operation and preprocesses the sensor data to obtain a preprocessed data sequence; Sequence segmentation module: Divides the preprocessed data sequence into data segments according to a preset time window, performs embedding encoding processing on each data segment, and obtains the corresponding sequence embedding representation; Feature extraction module: Inputs the sequence embedding representation into the Transformer model based on self-attention mechanism, extracts the global dependencies between time segments through multi-head attention operation, and outputs the time series feature representation; Multi-task analysis module: Performs multi-task analysis based on the time series feature representation and outputs predicted data; Early warning generation module: compares the predicted data with a preset threshold and generates early warning information based on the comparison result; Maintenance Operation Module: Executes corresponding elevator maintenance operations based on the warning information.

Images