Elevator internet-of-things abnormity diagnosis data management method and system

Abstract

The invention discloses an elevator Internet of Things abnormity diagnosis data management method and system. The method comprises the following steps: collecting heterogeneous time sequence data of an elevator and carrying out time alignment; performing real-time anomaly detection, data cleaning and adaptive compression on the edge computing node; uploading the processed data to a cloud, extracting and intelligently fusing multi-source features, and forming a unified feature vector; outputting an anomaly type, a severity level and a fault root cause chain by using the dynamic anomaly diagnosis model; and based on historical data and feature vectors, predicting the remaining service life of the key component through a fusion model, and generating a maintenance strategy by using a reinforcement learning optimization engine. The system comprises a physical sensing layer, a data acquisition and edge processing layer, a cloud data management and intelligent analysis layer and an application service layer. According to the elevator abnormal diagnosis system and method, full-process intelligent management from real-time monitoring and accurate diagnosis to predictive maintenance is achieved, and the accuracy, timeliness and operation and maintenance efficiency of elevator abnormal diagnosis are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of elevator Internet of Things and equipment health management technology, and specifically discloses an elevator IoT anomaly diagnosis data management method and system. Background Technology

[0002] With the acceleration of urbanization and the proliferation of high-rise buildings, elevators, as core equipment in vertical transportation, are receiving increasing attention for their operational safety and reliability. Traditional elevator maintenance mainly relies on periodic inspections and reactive repairs, which suffers from problems such as delayed response, high maintenance costs, and difficulty in preventing sudden failures. In recent years, the rise of Internet of Things (IoT) technology has made intelligent monitoring and predictive maintenance of elevators possible. By deploying various sensors on key elevator components, multimodal operating data such as vibration, noise, current, and temperature can be collected in real time, enabling online perception of equipment status.

[0003] However, in the actual process of elevator IoT anomaly diagnosis and data management, a series of technical challenges and limitations still exist: First, heterogeneous time-series data from sensors of different types and sampling rates are difficult to align and synchronize accurately, resulting in a weak foundation for subsequent analysis and affecting the accuracy of diagnosis; second, massive amounts of real-time monitoring data put enormous pressure on transmission bandwidth and cloud storage, and the raw data often contains a lot of noise and invalid information, so direct uploading leads to resource waste and low analysis efficiency, and existing systems usually lack the ability to perform real-time cleaning, compression, and preliminary screening at the data source (edge); third, at the anomaly diagnosis level, most existing... The methods rely on a single data source or simple rule thresholds, making it difficult to integrate multi-source features such as vibration, acoustics, and electrical characteristics for comprehensive judgment. They lack sensitivity in identifying early and complex faults and lack in-depth tracing of fault root causes and analysis of propagation paths. In addition, traditional life prediction models are mostly based on single statistical methods or physical models, failing to effectively combine real-time operating data with historical degradation information, resulting in limited prediction accuracy. Finally, in terms of maintenance decisions, existing strategies are often relatively static and fail to dynamically optimize based on the real-time health status of equipment, resource constraints, and multiple objectives (such as cost, safety, and availability), making it difficult to achieve a closed loop from "state perception" to "optimization decision-making".

[0004] Therefore, it is necessary to invent a method and system for managing elevator IoT anomaly diagnostic data to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies in the prior art, this invention provides an elevator IoT anomaly diagnosis data management method and system. This method involves collecting heterogeneous time-series data from elevators and performing time alignment; performing real-time anomaly detection, data cleaning, and adaptive compression at edge computing nodes; uploading the processed data to the cloud; extracting and intelligently fusing multi-source features to form a unified feature vector; using a dynamic anomaly diagnosis model to output the anomaly type, severity level, and root cause chain of the fault; and predicting the remaining service life of key components based on historical data and feature vectors through a fusion model, while using a reinforcement learning optimization engine to generate maintenance strategies. This effectively solves the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for managing elevator IoT anomaly diagnostic data, specifically including the following steps: S1. Collect heterogeneous time-series data through a multimodal sensor network deployed in the elevator, and use a millisecond-level data alignment mechanism based on device clock synchronization to align the heterogeneous time-series data in time sequence. In addition, adopt the "continuous acquisition + event trigger enhancement" mode to trigger the sensor to temporarily increase the sampling rate to capture instantaneous features when the elevator starts or stops or the load changes suddenly. S2. A lightweight anomaly detection model deployed at the edge is used to perform preliminary real-time anomaly detection on the aligned heterogeneous time series data, and the aligned heterogeneous time series data is cleaned, quality checked and adaptively compressed in real time at the edge computing node. S3. Upload the edge-processed data to the cloud, perform feature engineering to extract multi-source features, and use an attention-based feature fusion network to intelligently weight and fuse the multi-source features to construct a unified feature vector. S4. Input the unified feature vector into the dynamic anomaly diagnosis model to perform dynamic anomaly diagnosis, and output the anomaly type, severity level and fault root cause chain. The dynamic anomaly diagnosis model includes an early anomaly identification module based on a bidirectional long short-term memory network (Bi-LSTM) and a root cause analysis engine based on a fault propagation graph. S5. Based on historical anomaly data, equipment operating data, and unified feature vectors, the remaining useful life (RUL) and health score of key components are calculated by fusing physical models and data-driven remaining useful life prediction models. The fusion weights are dynamically allocated according to the prediction accuracy of the data-driven model. Furthermore, an optimization engine based on reinforcement learning is used to generate optimized maintenance strategies that include maintenance time windows, resource configuration, and task priorities.

[0007] Preferably, the heterogeneous time-series data includes vibration acceleration signals, acoustic feature signals, drive current signals, and temperature data of key components. The key components include the traction machine, guide rails, and door system. The multi-source features include time-domain, frequency-domain, and time-frequency domain combined features.

[0008] Preferably, the adaptive compression specifically involves dynamically adjusting the data sampling frequency and the amount of data transmitted based on the current operating status of the elevator and the preliminary anomaly detection results.

[0009] Preferably, the dynamic anomaly diagnosis further includes: dynamically adjusting the alarm thresholds for various anomalies based on the elevator's operating conditions.

[0010] Preferably, the maintenance decision optimization engine has optimization objectives including minimizing maintenance costs, maximizing equipment availability, and minimizing security risks.

[0011] Preferably, the early anomaly recognition module based on Bi-LSTM network adopts a 3-layer Bi-LSTM structure with 128 hidden layer units. It processes feature vectors through a sliding time window to achieve early anomaly recognition 3-5 minutes in advance and outputs a preliminary judgment of the anomaly type. The root cause analysis engine based on the fault propagation graph is used to construct the fault graph of elevator components. Combined with the early identification results, it traces back the source of the fault and finally outputs the accurate anomaly type, severity level 1-5, and fault root cause chain containing 3-5 nodes. Preferably, an elevator IoT anomaly diagnosis data management system includes: The physical sensing layer includes a multimodal sensor network deployed in the elevator to collect vibration, acoustic, current, and temperature data; The data acquisition and edge processing layer includes an edge computing node, which is equipped with a data alignment module, a preprocessing module and a lightweight anomaly detection model for synchronizing, cleaning, compressing and performing preliminary analysis on sensor data. The cloud-based data management and intelligent analysis layer includes: A time-series database is used to store time-series data from the edge processing layer; Feature engineering and fusion module, used to extract and fuse multi-source features; A model repository stores and manages anomaly diagnosis models, lifetime prediction models, and decision optimization models. The intelligent analytics engine is used to perform dynamic anomaly diagnosis and predictive maintenance decision calculations; The application service layer provides users with an interactive interface for real-time monitoring, fault alarms, health reports, and the generation and management of maintenance work orders.

[0012] Preferably, the health report provided by the application service layer includes a health score based on multi-dimensional indicators, visualization of the remaining service life of key components, and maintenance strategy recommendations.

[0013] The technical effects and advantages of this invention are as follows: 1. Achieving intelligent and automated diagnostic management throughout the entire process: This invention constructs a complete closed loop from data perception to maintenance decision-making. Through steps S1-S5, the system can automatically complete the entire process from raw data collection, real-time processing, cloud-based in-depth analysis, accurate diagnosis to the generation of intelligent maintenance suggestions, greatly reducing manual intervention and improving the intelligence level and efficiency of elevator operation and maintenance; 2. Achieve accurate, in-depth, and dynamic anomaly diagnosis: Employing an LSTM (Bi-LSTM) model to process time-series features enables the capture of contextual information, achieving early anomaly identification minutes in advance. The diagnostic model can dynamically adjust alarm thresholds based on operating conditions, reducing false alarms and adapting to different usage scenarios. By constructing a fault propagation graph for root cause tracing, it not only outputs the anomaly type but also locates the fault source, analyzes the propagation path (fault root cause chain), and assesses the severity level, providing direct evidence for precise maintenance and surpassing simple anomaly alarms. 3. Shift from post-failure repair to predictive maintenance: By integrating physical models and data-driven methods, the remaining service life of key components is predicted and a health score is calculated, enabling maintenance actions to be based on the actual health status of the equipment; by using a reinforcement learning engine, maintenance strategies that include time, resources, and priorities are generated with multiple objectives such as cost, availability, and security, achieving optimal allocation of maintenance resources and improving the economics of operation and maintenance and the availability of equipment. Attached Figure Description

[0014] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0016] Figure 2 This is a detailed flowchart of the dynamic anomaly diagnosis model of the present invention.

[0017] Figure 3 This is a detailed flowchart of the predictive maintenance of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides, for example Figure 1The elevator IoT anomaly diagnosis data management method shown includes the following steps: S1. Collect heterogeneous time-series data through a multimodal sensor network deployed in the elevator, and use a millisecond-level data alignment mechanism based on device clock synchronization to align the heterogeneous time-series data in time sequence. In addition, adopt the "continuous acquisition + event trigger enhancement" mode to trigger the sensor to temporarily increase the sampling rate to capture instantaneous features when the elevator starts or stops or the load changes suddenly. Furthermore, in the above technical solution, the heterogeneous time-series data includes vibration acceleration signals, acoustic feature signals, drive current signals, and temperature data of key components. The key components include traction machines, guide rails, and door systems. The multi-source features include time-domain, frequency-domain, and time-frequency domain combined features.

[0020] It should be further explained that a comprehensive, high-precision multimodal sensor network was constructed for key elevator components and core operating parameters. The deployment locations and monitoring targets are as follows: Vibration acceleration sensors: installed on the traction machine base, guide rail bracket (one at the top and one at the bottom), and door system transmission mechanism, are used to collect vibration acceleration signals and monitor component wear, imbalance, looseness and abnormality; Acoustic sensors: installed inside the car, in the traction machine room, and near the door system, are used to collect acoustic characteristic signals and monitor abnormal noise, impact sound, friction sound, etc. Current sensors: installed in the power supply circuit of the traction machine drive motor (one for each of the three phases) to collect drive current signals and monitor load abnormalities and motor faults; Temperature sensors: installed on the windings of the traction machine, the contact surface of the guide rail, the housing of the door system motor, and key components of the control cabinet, are used to collect temperature data of key components and monitor overheating, abnormal temperature fluctuations, etc. Furthermore, a "continuous collection + event-triggered enhancement" model is adopted for real-time data collection: Under normal operating conditions, each sensor continuously collects data at a preset sampling rate to generate a raw time-series data stream; When the elevator experiences special operating conditions such as start-stop switching, sudden load changes (e.g., the load changes from empty to full load), or speed adjustments, the edge computing node triggers the sensor to temporarily increase the sampling rate (e.g., the vibration sensor increases from 1000Hz to 2000Hz for 5 seconds) to capture the instantaneous data characteristics during sudden changes in operating conditions.

[0021] S2. A lightweight anomaly detection model deployed at the edge is used to perform preliminary real-time anomaly detection on the aligned heterogeneous time series data, and the aligned heterogeneous time series data is cleaned, quality checked and adaptively compressed in real time at the edge computing node. Furthermore, in the above technical solution, the adaptive compression specifically involves dynamically adjusting the data sampling frequency and the amount of data transmitted based on the current operating status of the elevator and the preliminary anomaly detection results.

[0022] It should be further explained that the lightweight anomaly detection model adopts a "lightweight CNN + logistic regression fusion model" to balance detection accuracy and edge computing power. The feature extraction layer uses a 3-layer lightweight CNN with a kernel size of 3×3 and channels of 16, 32 and 64 respectively, which can accurately capture local features of multimodal time series data; the classification layer abandons complex fully connected layers and uses logistic regression to reduce computational overhead, and finally outputs three categories of results: "normal (0), suspected anomaly (1), and anomaly (2)". In order to adapt to the computing power of edge nodes, the model size is controlled within 5MB after INT8 quantization and compression. Redundant parameters are pruned to ensure that the number of parameters is ≤100,000, ensuring that the inference latency is ≤50ms / frame on edge nodes with ARM Cortex-A53 architecture and a main frequency of 1.2GHz. When deploying the model, the quantized model is solidified to the eMMC flash memory of the edge computing node, and the TensorFlowLite edge inference framework is used to support real-time inference of streaming data without relying on cloud computing power.

[0023] S201, Detection Execution Process Input data preparation: Receive the aligned time series data output from step S1, divide it according to the "sliding time window" rule, set the window size to 10s and the step size to 5s (adapted to the window parameters of the Bi-LSTM model), and each window contains complete time series data of all modes within that time period.

[0024] Feature dimensionality reduction preprocessing: To avoid overloading the computing power on the edge side, lightweight feature extraction is performed on the multimodal data within each window. In the time domain dimension, four types of features are calculated for each modality: mean, variance, peak value, and kurtosis. In the frequency domain dimension, two simplified features, the dominant frequency and harmonic amplitude ratio, are extracted using Fast Fourier Transform (FFT). Finally, each window outputs a 24-dimensional lightweight feature vector (4 modalities × (4 time domain features + 2 frequency domain features)), which serves as the input data for the detection model.

[0025] Real-time inference and result output: Edge nodes perform streaming inference on the feature vectors of each window, synchronously outputting classification results and confidence scores in the 0-1 range. If the output is "abnormal" and the confidence score is ≥0.85, an edge-side local alarm is immediately triggered (indicator light flashing + local log recording), and the data in that window is marked as "high-priority transmission data"; if the output is "suspected abnormal" and the confidence score is between 0.6 and 0.85, it is marked as "medium-priority transmission data"; if the output is "normal" and the confidence score is <0.6, it is marked as "low-priority transmission data".

[0026] S202 Real-time Data Cleaning Implementation Plan (a) Vibration acceleration signal cleaning Vibration acceleration signals are susceptible to Gaussian noise and impulse noise (electromagnetic interference). A combined approach of "median filtering + adaptive threshold denoising" is adopted. First, impulse noise is removed by median filtering with a window size of 5. Then, the standard deviation σ of the signal is calculated, and extreme values ​​exceeding the range of "mean ± 3σ" are eliminated. For the missing points generated after elimination, linear interpolation is used to fill in the missing points and ensure data continuity.

[0027] (II) Acoustic Feature Signal Cleaning The main interference in the acoustic feature signal is environmental background noise, which is processed using "spectral subtraction + energy threshold screening". First, the background noise spectrum when the elevator is stopped is extracted. The interference of background noise on the effective signal is canceled by spectral subtraction. Then, the energy of the processed signal is calculated, and invalid data segments with energy below 10dB are removed, while valid data containing abnormal noise (such as friction sound and impact sound) are retained.

[0028] (III) Cleaning of drive current signal Drive current signals are prone to spike interference (power fluctuations) and zero-point drift. A "moving average filtering + baseline calibration" scheme is adopted. The spike interference is smoothed by a moving average filter with a window size of 3. The zero-point drift of the signal is corrected by using the current value when the elevator is stationary under no-load conditions as the baseline. At the same time, invalid values ​​that exceed the sensor range (0-50A) are eliminated to ensure that the current data reflects the actual load of the elevator and the operating status of the motor.

[0029] (iv) Temperature data cleaning Common interferences in temperature data include abrupt noise (due to poor sensor contact) and data jumps. These are addressed using a "first-order difference threshold method + trend consistency check." The absolute value of the difference between adjacent data points is calculated, and abrupt changes with a difference > 2℃ / ms are removed. If the trend of three consecutive data points is opposite to the overall trend (e.g., the overall trend is rising but consecutive cooling data appear), it is marked as abnormal data and corrected using linear interpolation to ensure that the temperature data conforms to the physical laws governing component operation.

[0030] S203, Data Quality Verification Implementation Plan (a) Integrity verification Using a 10-second sliding window as the unit, the missing rate of each modality data within a single window is calculated, with a threshold of ≤5%. If the missing rate is between 5% and 10%, data is supplemented using interpolation; if the missing rate is >10%, the data in that window is marked as "invalid," and only local logs are recorded, without uploading to the cloud to avoid consuming transmission bandwidth.

[0031] (ii) Validity verification The core metric for validity verification is the signal-to-noise ratio (SNR) of the cleaned data. Different thresholds are set for different modes: SNR ≥ 30dB for vibration and acoustic data, and SNR ≥ 40dB for current and temperature data. If the SNR does not meet the standard, the cleaning process for the corresponding mode is repeated. If the threshold requirement is still not met after the second cleaning, the data is marked as "low-quality data", compressed, uploaded, and a quality warning label is added.

[0032] (III) Consistency Verification Verify the logical consistency of multimodal data (e.g., whether the current changes synchronously when the temperature rises), focusing on the trend consistency of core modes (temperature-current, vibration-acoustics), with a threshold of ≥80%. If the consistency is <80%, mark it as "suspected data anomaly," and make a comprehensive judgment based on the lightweight anomaly detection results. If the detection result is "normal," retain the data and upload it; if the detection result is "suspected anomaly" or "anomaly," increase the data transmission priority.

[0033] (iv) Stability verification Calculate the fluctuation amplitude of continuous frames of single-modal data, and set a threshold of 10% of the sensor's full scale. If the fluctuation amplitude exceeds the limit, mark it as "unstable data", retain the original data and upload it to provide a complete sample of abnormal data for cloud-based in-depth analysis, and avoid losing key information due to excessive cleaning.

[0034] (v) Verification Execution Process After data cleaning is completed in each 10-second sliding window, quality verification is performed in parallel with lightweight anomaly detection to shorten the overall processing time. Four indicators—completeness, validity, consistency, and stability—are quickly calculated and compared with preset thresholds, outputting "verification passed," "verification warning," or "verification failed" results. Verification results are linked to detection results: Verification passed + normal detection → low-priority compressed upload; Verification passed + suspected / anomaly detected → medium / high-priority compressed upload; Verification warning → complete data retained, compressed upload, and marked; Verification failed → only local log recording, no upload.

[0035] S204, Adaptive Compression Implementation Scheme (a) Dynamic adjustment logic of compression strategy A three-dimensional dynamic compression strategy is constructed based on "elevator operating status" and "preliminary anomaly detection results" to achieve intelligent adaptation of sampling frequency, compression algorithm, and compression ratio.

[0036] Scenario 1 (Stationary / Unloaded Operation + Normal Detection Results): The elevator is in a low-load state and there are no abnormalities. The sampling frequency is reduced by 50%, and the LZ77+Huffman hybrid compression algorithm is used with a compression ratio controlled at 4:1 to minimize transmission and storage pressure.

[0037] Scenario 2 (Full load operation / frequent start-stop + normal detection results): The elevator is under high load and complete operating data needs to be retained. The sampling frequency remains unchanged. The LZ77 compression algorithm is used, and the compression ratio is controlled at 2:1 to balance data integrity and compression efficiency.

[0038] Scenario 3 (Any operating state + suspected anomaly detection results): It is necessary to accurately retain potential anomaly features, keep the sampling frequency unchanged, use the FLAC lossless compression algorithm, and control the compression ratio to 1.5:1 to avoid data distortion affecting subsequent diagnosis.

[0039] Scenario 4 (Arbitrary running state + anomaly detection results): It is necessary to capture complete anomaly data details, increase the sampling frequency by 50%, and use lossless compression (compression ratio 1:1, i.e., do not compress core data) to ensure that the cloud can obtain complete anomaly features.

[0040] Note: The elevator's operating status is identified in real time using the current data (load judgment: current ≥30A is full load) and speed signal (running / stationary judgment) collected in step S1.

[0041] (ii) Compressed execution process Scene recognition: Edge nodes analyze the cleaned data in real time, extract elevator operating status features (load, running / stationary), and combine them with lightweight anomaly detection results to automatically match the above four types of scenes.

[0042] Dynamic parameter configuration: Based on the matching scenario, the sampling frequency, compression algorithm and compression ratio parameters are automatically adjusted without manual intervention, ensuring that the parameters are adapted to the current operation and abnormal state.

[0043] Data compression and encapsulation: The verified data is resampled at the adjusted sampling frequency (downsampling / upsampling) and compressed using the corresponding compression algorithm. The compressed data frame includes a compression identifier, original sampling rate, compression ratio, timestamp range, data body, and CRC32 checksum (to ensure transmission integrity), facilitating cloud parsing and verification.

[0044] Compression effect monitoring: Edge nodes record the time and actual compression ratio of each compression. If the compression time exceeds the threshold or the compression ratio deviation is ≥20%, the system will automatically switch to a backup compression algorithm (such as switching from hybrid compression to LZ77) to ensure the stable operation of the compression process.

[0045] S3. Upload the edge-processed data to the cloud, perform feature engineering to extract multi-source features, and use an attention-based feature fusion network to intelligently weight and fuse the multi-source features to construct a unified feature vector. It should be further explained that the specific process for constructing a unified feature vector is as follows: S301, Multi-source Feature Engineering Extraction Feature extraction is performed using 10-second sliding window data output by S2 as units. For four types of modal data, namely vibration acceleration, acoustic features, drive current, and temperature of key components, time-domain, frequency-domain, and time-frequency joint features are extracted respectively to ensure that the feature dimensions fully cover the equipment operating status information. At the same time, redundant information is eliminated through feature screening.

[0046] (I) Feature extraction of vibration acceleration signal Time-domain features: Calculate eight basic features of the signal within the window, including peak value, mean, variance, kurtosis, skewness, impulse factor, peak factor, and waveform factor, to reflect the overall strength and distribution characteristics of the signal; additionally extract the peak-to-peak value, rising edge slope, and falling edge slope of the signal to capture vibrational abrupt changes.

[0047] Frequency domain features: After performing FFT on the signal within the window and converting it into frequency domain data, five types of features are extracted, including the main frequency (the frequency with the highest energy), the ratio of the amplitude of the first 5 harmonics to the main frequency, the energy proportion of the key frequency band from 50Hz to 500Hz, and the frequency domain entropy, which reflect the frequency distribution law of vibration.

[0048] Joint time-frequency domain features: The signal is decomposed into 5 layers using db4 wavelet packet transform, and the energy value, energy entropy, and the first 3 singular values ​​after singular value decomposition of each decomposed frequency band are extracted, for a total of 10 types of features. This captures the local features of the vibration signal in different time-frequency intervals, which is suitable for the identification of early weak anomalies.

[0049] (II) Acoustic Feature Signal Feature Extraction Temporal characteristics: Five types of features of the signal within the calculation window are short-time energy, zero-crossing rate, average amplitude, peak amplitude, and amplitude variance, which reflect the strength and fluctuation characteristics of the sound; the proportion of silent segments in the signal is extracted, and meaningless background noise interference is eliminated.

[0050] Frequency domain features: The frequency domain data after FFT transformation is mapped to Mel spectrum through Mel filter bank, and 12th order Mel frequency cepstral coefficients and their first and second order differences are extracted, totaling 36 types of features; additionally, the spectral center, spectral bandwidth, and spectral roll-off point are calculated to reflect the frequency concentration trend and distribution range of the sound.

[0051] Joint features in the time and frequency domain: The time and frequency matrix is ​​generated by short-time Fourier transform, and six types of features are extracted: row entropy (time dimension entropy), column entropy (frequency dimension entropy), time and frequency coordinates corresponding to the maximum energy point, and texture features (contrast, correlation) of the time and frequency matrix, to capture the time and frequency distribution features of abnormal sounds.

[0052] (III) Feature Extraction of Drive Current Signal Time-domain characteristics: Seven types of features are calculated for the three-phase current within the calculation window, including effective value, peak value, mean value, variance, peak factor, waveform factor, and distortion factor. The unbalance of the three-phase current (the ratio of the difference between the maximum and minimum effective values) and the number of current mutations (the number of deviations > mean ± twice the variance) are extracted to reflect the electrical stability of the motor operation.

[0053] Frequency domain characteristics: After performing FFT on the current signal, eight types of characteristics are extracted, including the amplitude of the fundamental frequency (50Hz), the ratio of the amplitude of the 2nd to 10th harmonics to the fundamental amplitude, the total harmonic distortion rate, and the phase deviation of each harmonic, which reflect the degree of harmonic pollution of the current.

[0054] Joint time-frequency domain features: The Hilbert-Huang transform is used to perform empirical mode decomposition on the current signal to obtain 5 intrinsic mode functions (IMFs). 15 types of features are extracted from each IMF, including energy, frequency centroid, and standard deviation of instantaneous frequency, to capture the nonlinear and non-stationary characteristics of the current signal.

[0055] (iv) Extraction of temperature data features from key components Time-domain features: The calculation window includes seven types of features: mean, maximum, minimum, range, fluctuation range (difference between maximum and minimum), trend slope (slope after linear fitting), and number of consecutive heating / cooling cycles, reflecting the overall temperature level and trend.

[0056] Frequency domain features: Perform FFT on temperature time series data to extract three types of features: the dominant frequency of temperature fluctuations (such as the fluctuation frequency caused by periodic heat dissipation), the amplitude of the dominant frequency, and the energy proportion of high-frequency components (abrupt fluctuations), which reflect the periodicity and abruptness of temperature changes.

[0057] Joint features in the time and frequency domains: Wavelet transform is used to decompose the temperature data into multiple scales, and six types of features are extracted at each scale, including the energy of detail coefficients, the energy of approximation coefficients, the time position corresponding to the abrupt change point and the temperature difference, to capture the time and frequency characteristics of abrupt temperature changes.

[0058] (v) Feature selection and dimensionality reduction Invalid feature removal: Calculate the variance of each type of feature using analysis of variance, and remove invalid features with a variance less than 0.01 (such as the mean temperature feature that does not change over a constant period of time) to ensure that the retained features have discriminative power.

[0059] Redundant feature removal: Calculate the mutual information value between the remaining features. If the mutual information value between two features is greater than 0.85, they are considered highly redundant. The feature that contributes more to the diagnosis of anomalies (feature importance score trained based on historical data) is retained to reduce the computational cost of subsequent fusion.

[0060] Final feature set determination: After screening, 30-dimensional, 40-dimensional, 35-dimensional, and 25-dimensional effective features were retained for the four modes of vibration, acoustics, current, and temperature, respectively, for a total of 130-dimensional multi-source features, providing comprehensive and concise input for subsequent fusion.

[0061] S302, Attention Mechanism Feature Fusion Network Deployment A two-layer attention architecture of "modal-level attention + feature-level attention" is adopted to achieve intelligent weighted fusion of multi-source features, ensuring that key modalities and key features receive higher weights and improving the recognizability of the fused vector.

[0062] (I) Network Structure Design and Adaptation Input layer: Receives 130-dimensional multi-source features after filtering, and divides them into 4 feature subsets according to mode (vibration 30-dimensional, acoustic 40-dimensional, current 35-dimensional, temperature 25-dimensional), which are then input into the feature mapping layer of the corresponding mode.

[0063] Feature mapping layer: Each modality is configured with an independent lightweight fully connected layer (the number of hidden layer units is twice the feature dimension of the corresponding modality, such as 60 hidden units in the vibration modality mapping layer). The activation function is ReLU, which maps the features of each modality to a high-dimensional space of the same dimension (all mapped to 64 dimensions), eliminating the feature scale differences between modalities.

[0064] Modal-level attention layer: Calculates the global variance of each modal mapping feature (reflecting the mode's sensitivity to anomalies), combines it with prior weights of modal importance obtained from training on historical data, and generates modal attention weights (4 weight values, summing to 1) through the Softmax function. For example, in abnormal states, the vibration mode variance is higher, and the corresponding weight may reach 0.4, the acoustic mode weight is 0.3, the current weight is 0.2, and the temperature weight is 0.1, ensuring a higher contribution from key modes.

[0065] Feature-level attention layer: For the 64-dimensional features mapped to each modality, a shared fully connected layer (128 hidden units) is used to calculate the feature importance score, and then the Sigmoid activation function is used to generate feature-level attention weights (64 weight values ​​for each modality, ranging from 0 to 1). Redundant features are suppressed and key features (such as the kurtosis of vibration and the harmonic distortion rate of current) are enhanced.

[0066] Fusion layer: The "weighted summation + residual connection" mechanism is adopted. First, the mapping features of each modality are multiplied with the corresponding modality weights and feature weights to obtain a weighted feature vector. Then, the weighted feature vectors of the four modalities are summed and a residual connection (0.1 times the sum of the original mapping features) is added to avoid gradient vanishing and enhance the stability of fusion.

[0067] Output layer: The fused feature vector is compressed to 64 dimensions through a fully connected layer (64 hidden units) and output as a unified feature vector, with the dimension adapted to the input requirements of the dynamic anomaly diagnosis model in step S4.

[0068] (II) Network Training and Deployment Training dataset construction: Collect historical elevator operation data (including normal state and 12 types of core fault state data, covering 32 fault propagation paths), generate labeled multi-source features and fusion labels according to the S1-S3 process, and construct a training set containing 100,000 samples (7:2:1 division of training set, validation set and test set).

[0069] Training parameter configuration: The Adam optimizer is used, with an initial learning rate of 0.001, which is decayed to 0.8 every 20 rounds; the loss function is cross-entropy loss (combined with the classification accuracy after feature fusion); the training rounds are set to 100, and the batch size is 32; overfitting is avoided by using an early stopping mechanism (training stops if the validation set loss does not decrease for 10 consecutive rounds).

[0070] Cloud Deployment: After training, the model is quantized and compressed using INT8 and then deployed to the model repository of the cloud-based intelligent analysis engine. The TensorFlow Serving framework is used to provide inference services, supporting real-time inference of streaming data. The latency of single-sample fusion inference is ≤100ms, meeting the needs of real-time diagnosis.

[0071] S303. Construction and Output of Unified Feature Vectors (a) Vector standardization Z-score standardization is performed on the 64-dimensional feature vector output by the fusion network. This involves subtracting the mean of the feature in the historical normal dataset from each feature value and then dividing by the standard deviation to ensure that all feature dimensions have a consistent scale (mean 0, variance 1), thus avoiding the impact of feature scale differences on the training and inference accuracy of subsequent diagnostic models.

[0072] (ii) Vector metadata binding Add core metadata to the standardized unified feature vector, including: the time window range corresponding to the vector (accurate to milliseconds, aligned with the S1-S2 timestamps), elevator ID, key component identifiers (traction machine, guide rail, door system), feature quality score (calculated based on the quality verification results of the input data and the feature selection pass rate, 0-10 points), and fusion network inference confidence (0-1 interval, reflecting the reliability of the fusion result), to ensure that the vector is traceable and evaluable.

[0073] (III) Vector Output and Storage Real-time output: The completed unified feature vector is pushed to the dynamic anomaly diagnosis module of the cloud-based intelligent analysis engine in real time to support real-time anomaly identification and root cause analysis; at the same time, it is stored in a dedicated feature vector database in chronological order, retaining the feature vector data of the most recent 90 days for model iteration and historical tracing.

[0074] S4. Input the unified feature vector into the dynamic anomaly diagnosis model to perform dynamic anomaly diagnosis, and output the anomaly type, severity level and fault root cause chain. Furthermore, in the above technical solution, the dynamic anomaly diagnosis also includes: dynamically adjusting the alarm thresholds for various anomalies based on the elevator's operating conditions.

[0075] Furthermore, in the above technical solution, the dynamic anomaly diagnosis model includes: An early anomaly detection module based on a bidirectional long short-term memory network (Bi-LSTM): It adopts a 3-layer Bi-LSTM structure with 128 hidden layer units. It processes feature vectors through a sliding time window to achieve early anomaly detection 3-5 minutes in advance and outputs a preliminary judgment of the anomaly type. The root cause analysis engine based on the fault propagation graph constructs a fault graph associated with elevator components (containing 12 types of core faults and 32 propagation paths), combines early identification results to trace the source of the fault in reverse, and finally outputs the accurate anomaly type, severity level 1-5, and fault root cause chain containing 3-5 nodes. It should be further explained that the specific process of dynamic anomaly diagnosis using the dynamic anomaly diagnosis model is as follows: Figure 2 As shown, it specifically includes: S401, Dynamic alarm threshold adjustment The alarm thresholds are dynamically optimized based on the real-time operating conditions of the elevator to avoid false alarms and missed alarms caused by fixed thresholds, and to adapt to the diagnostic needs of different operating scenarios.

[0076] (I) Extraction of working condition features From the bound metadata and time series database, three core operating condition features are extracted: load rate (calculated based on the effective value of the drive current, ≤10% is no load, 10%-60% is half load, and >60% is full load), operating frequency (number of elevator starts and stops per unit time, ≤5 times / hour is low frequency, 5-15 times / hour is medium frequency, and >15 times / hour is high frequency), and ambient temperature (based on the average temperature data of key components, ≤25℃ is normal temperature, 25-40℃ is medium temperature, and >40℃ is high temperature).

[0077] (ii) Threshold dynamic adjustment logic For 12 types of core faults, a preset basic alarm threshold is established (determined based on the boundary values ​​between historical normal data and fault data), and then dynamically adjusted according to real-time operating conditions. Load factor related adjustment: Under full load conditions, the alarm threshold for mechanical faults (such as traction machine bearing wear and guide rail deviation) is reduced by 20% (because heavy loads tend to exacerbate fault manifestation); under no-load conditions, the alarm threshold for electrical faults (such as excessive drive current harmonics) is increased by 15% (to avoid false triggering of electrical fluctuations under light loads).

[0078] Operating frequency correlation adjustment: When operating at high frequency, the alarm threshold for fatigue-related faults (such as aging of door system transmission components) is reduced by 15% (high frequency start-stop accelerates component wear); when operating at low frequency, the alarm threshold for static faults (such as loose components) is increased by 10% (fault characteristics are weaker under low activity).

[0079] Ambient temperature-related adjustment: Under high temperature conditions, the alarm threshold for heat-related faults (such as overheating of traction machine windings and aging of control cabinet components) is reduced by 25% (high temperature easily induces heat failure); under normal temperature conditions, the threshold remains at the base value; under low temperature conditions, the alarm threshold for lubrication-related faults (such as insufficient lubrication of guide rails) is increased by 20% (low temperature affects lubrication effect, and it is necessary to distinguish between normal friction and fault friction).

[0080] (III) Threshold Activation and Update Mechanism The dynamically adjusted threshold is linked to the corresponding elevator's diagnostic task in real time. The diagnostics for each time window use the threshold calculated under the current operating conditions. If the operating conditions do not change significantly (fluctuation ≤10%), the threshold remains unchanged. If the operating conditions change abruptly (e.g., switching from no load to full load), the threshold is immediately recalculated and takes effect. Threshold adjustment records are stored synchronously with diagnostic results for easy tracing of the causes of false alarms / missed alarms.

[0081] S402, Early Anomaly Identification Based on Bi-LSTM (a) Model structure and parameter adaptation The model structure strictly adheres to the document definition: the input layer receives a 64-dimensional uniform feature vector; the hidden layer consists of three Bi-LSTM layers with 128 hidden units per layer, employing a Dropout layer (dropout rate=0.2) to prevent overfitting; and the output layer uses the Softmax activation function, outputting the probability distributions of 12 core fault categories (such as traction machine bearing wear, guide rail parallelism deviation, and door system motor failure) and the "normal" state. The model is deployed in a cloud-based model repository after INT8 quantization and compression, and the inference framework uses TensorFlow Serving, supporting real-time streaming data processing.

[0082] (II) Sliding Window Reasoning Process Input sequence construction: Starting with the unified feature vector of the current time window as the endpoint, the feature vectors of the previous 5 consecutive time windows are concatenated forward (a total of 6 windows, covering 30 seconds of data) to construct a 6×64-dimensional temporal feature sequence, which is adapted to the Bi-LSTM's requirement for capturing contextual information.

[0083] Real-time inference execution: The time-series feature sequence is input into the Bi-LSTM model. The model extracts the correlation features between preceding and following time windows through a bidirectional propagation mechanism, and outputs the confidence scores for various faults and "normal" states. The inference latency is controlled to ≤50ms / sequence to ensure the timeliness of early warning.

[0084] Preliminary judgment output: Set the confidence threshold to 0.6. If the confidence of a certain type of fault is ≥0.6, output the fault type as "suspected anomaly type"; if the confidence of multiple fault types is ≥0.6, take the top 3 types with the highest confidence as the "suspected anomaly type set"; if the confidence of all fault types is <0.6 and the confidence of "normal" state is ≥0.7, output "currently no anomaly"; if the confidence of "normal" state is <0.7 and the confidence of no fault type is ≥0.6, output "suspected unknown anomaly" and mark it as high-priority data to be analyzed for root cause.

[0085] Early warning trigger: If the output is "suspected anomaly type" or "suspected unknown anomaly", the anomaly warning lead time is calculated by combining the time difference between the current time window and historical data (based on the fault development time series data during model training; for example, if it takes 3 minutes for a fault to go from feature appearance to failure, then the warning lead time is 3 minutes). The early warning information is pushed to the application service layer, including the warning time, suspected anomaly type, and confidence level.

[0086] (III) Model Training and Iteration The training dataset reuses the 100,000 sample sets constructed using S3, supplemented with time-series development data for 12 core fault categories (each fault category includes complete time-series samples from the initial appearance of features to complete failure). The training, validation, and test sets are divided in a 7:2:1 ratio. The Adam optimizer is used with an initial learning rate of 0.001, decaying to 0.8 times the original rate every 20 epochs. Cross-entropy loss is used as the loss function, and training lasts for 100 epochs. An early stopping mechanism (stopping training if the validation set loss does not decrease for 10 consecutive epochs) prevents overfitting. The model is incrementally trained monthly based on newly added fault data. Updated models are automatically deployed to the model repository, replacing older versions and ensuring continuous optimization of early identification accuracy.

[0087] S403, Root Cause Analysis Based on Fault Propagation Map By combining the suspected anomaly types output by Bi-LSTM, and constructing an elevator component association fault map, the source of the fault can be traced in reverse to clarify the anomaly type, severity level, and root cause chain of the fault.

[0088] (I) Construction and maintenance of fault propagation map The core structure of the fault map includes 12 types of core fault nodes, 32 fault propagation paths, and component-related edges. Core fault nodes cover typical faults in key components such as the traction machine, guide rails, and door systems (e.g., traction machine bearing wear, insufficient guide rail lubrication, and damaged door system transmission gears). Propagation paths are constructed based on historical fault data and the physical relationships between components (e.g., "traction machine bearing wear → increased traction machine vibration → drive current fluctuations → control cabinet component overheating"). Component-related edges indicate the probability of fault propagation (e.g., the probability of bearing wear causing increased vibration is 85%).

[0089] Dynamic map updates: After maintenance, fault verification data is collected through the effect evaluation and optimization closed-loop module. If a new fault propagation path is found (such as "door system motor failure → door switch jamming → car vibration" which is not recorded), a new path node and associated edge are added, and the propagation probability is updated. If the actual propagation probability of a certain path deviates from the map label by ≥30%, the labeled probability is corrected to ensure that the map is consistent with the actual fault patterns.

[0090] (II) Root Cause Tracing Implementation Process Suspected path screening: Using the “suspected anomaly type” output by Bi-LSTM as the endpoint node, screen all propagation paths ending at this node in the fault propagation map (e.g., if the suspected anomaly is “drive current fluctuation”, then screen the paths “traction machine bearing wear → traction machine vibration aggravation → drive current fluctuation”, “guide rail deviation → car tilting → drive current fluctuation”, etc.).

[0091] Path probability calculation: Combining key features in the unified feature vector (such as vibration kurtosis and current harmonic distortion rate) with operating condition information, the matching probability of each selected path is calculated. For example, if the confidence level of the feature "excessive vibration kurtosis" is high, the matching probability of the path "traction machine bearing wear → increased vibration → current fluctuation" increases; if the operating condition is full load, the matching probability of the path "guide rail deviation → car tilting → current fluctuation" increases.

[0092] Root cause node location: Select the path with the highest matching probability (probability ≥ 70%), and the starting node of this path is the root cause of the fault (e.g., "traction machine bearing wear"); if the highest probability path is < 70%, then combine the common nodes of the first 3 high probability paths and determine the root cause through feature correlation analysis (e.g., if the first 3 paths all contain "intensified traction machine vibration", and the vibration feature confidence is the highest, then the root cause is the source fault corresponding to "intensified traction machine vibration").

[0093] Fault root cause chain generation: From the root cause node to the suspected abnormal node, extract the key nodes in the path (remove redundant intermediate nodes) to form a fault root cause chain containing 3-5 nodes (such as "traction machine bearing wear → increased traction machine vibration → drive current fluctuation → control cabinet component overheating"). The propagation probability and propagation time between nodes are marked (based on the average propagation time statistically analyzed from historical data).

[0094] (III) Severity Level Assessment Based on the type of root cause of the failure, the speed of propagation, the scope of impact, and the safety risks, a five-level severity assessment standard was established: Level 1 (Minor Abnormality): The cause is minor wear, minor parameter deviation, etc., with no safety risk. The propagation speed is extremely slow (no significant deterioration after >72 hours), and it only affects the local performance of the equipment (such as slight jamming of door opening and closing). No emergency treatment is required.

[0095] Level 2 (General Abnormality): The root cause is moderate wear, local loosening, etc., with low safety risk and slow propagation speed (may worsen within 24-72 hours). It affects the normal operation of a single component (such as insufficient lubrication of the guide rail leading to increased local friction) and requires scheduled maintenance.

[0096] Level 3 (Severe Anomaly): Caused by severe wear, excessive electrical parameters, etc., with moderate safety risk and moderate propagation speed (may worsen within 6-24 hours). It affects multiple related components (such as increased vibration of the traction machine leading to uneven stress on the guide rail) and requires maintenance within 48 hours.

[0097] Level 4 (Severe Abnormality): The root cause is a precursor to component failure, a serious exceedance of key parameters, etc., with high safety risks and rapid propagation (may worsen within 1-6 hours), affecting the core operating functions of the elevator (such as excessive harmonic distortion rate of drive current leading to motor overheating), requiring emergency maintenance within 12 hours.

[0098] Level 5 (Emergency Anomaly): The cause is the partial failure of a critical component, a major safety hazard, an extremely high safety risk, and a very rapid spread (it may cause a shutdown or safety accident within 1 hour). For example, the traction machine bearing may be stuck or the door system may not be able to close. Immediate shutdown and maintenance are required.

[0099] During the assessment, the base level is first determined based on the root cause type, and then the level is adjusted based on the propagation speed (based on the propagation time estimate of the root cause chain) and the scope of impact (the number of components covered by the root cause chain). For example, if the base level is 3, it will be adjusted to level 4 if the propagation speed is fast. Finally, the severity level of 1-5 is output.

[0100] S5. Based on historical anomaly data, equipment operating data, and unified feature vectors, the remaining useful life (RUL) and health score of key components are calculated by fusing physical models and data-driven remaining useful life prediction models. The fusion weights are dynamically allocated according to the prediction accuracy of the data-driven model. Furthermore, an optimization engine based on reinforcement learning is used to generate optimized maintenance strategies that include maintenance time windows, resource configuration, and task priorities.

[0101] Furthermore, in the above technical solution, the maintenance decision optimization engine has optimization objectives including minimizing maintenance costs, maximizing equipment availability, and minimizing security risks.

[0102] It should be further explained that the specific process of step S5 is as follows: Figure 3 As shown, it specifically includes: S501 Input Data Preparation and Preprocessing (a) Integration of multi-source input data Core feature data: Extract the 64-dimensional unified feature vector output by S3, and focus on retaining features that are strongly correlated with the degradation of key components (traction machine, guide rail, door system) (such as vibration kurtosis, current harmonic distortion rate, temperature trend slope, etc.), and bind the corresponding time window, elevator ID, component identification metadata.

[0103] Diagnostic results data: Import the structured diagnostic results output by S4, including anomaly type, severity level 1-5, root cause chain of 3-5 nodes, and diagnostic confidence. The core components involved in the root cause chain are the focus of RUL prediction and maintenance strategies.

[0104] Historical basic data: Three types of historical data are extracted from the cloud time-series database and equipment archive: historical abnormal data (failure records and degradation trend data of similar components in the past 3 years), cumulative equipment operating condition data (cumulative running time, cumulative number of start-stops, average load rate, and historical ambient temperature distribution), and maintenance history data (past maintenance time, maintenance type, replacement component model, maintenance cost, and health recovery status after maintenance).

[0105] Basic component information: Collect the factory parameters (design service life, rated load, temperature range, wear threshold), installation time, and maintenance records of key components as the basic parameter inputs for the physical model.

[0106] (II) Data Preprocessing and Feature Enhancement Data alignment and fusion: Taking "component ID + time axis" as the core, real-time feature vectors, diagnostic results and historical operating condition data are aligned according to the time dimension (the smallest granularity is the cumulative duration corresponding to a 10-second sliding window) to form a three-dimensional data matrix of "real-time status - historical degradation - diagnostic conclusion".

[0107] Missing and outlier handling: For missing items in historical data (such as incomplete early maintenance records), fill them with the mean of similar components or use interpolation based on operating conditions; for outliers (such as fault data that exceeds physical limits), combine the anomaly judgment results of S4 to remove invalid data or mark them as "fault mutation points" for separate processing to avoid interfering with prediction accuracy.

[0108] Feature enhancement: Three new types of derived features have been added—degradation rate feature (calculated based on the feature vector change rate over 10 consecutive time windows), fault impact depth feature (generated based on the number of components covered by the root cause chain and the propagation probability), and maintenance gain feature (calculated based on the difference in health status before and after historical maintenance). The original feature vector has been expanded to 80 dimensions, improving the input recognition of the prediction model.

[0109] S502, RUL and Health Score Prediction Implementation Plan By adopting a weighted fusion approach of "physical model + data-driven model", the system balances prediction accuracy and physical rationality, while generating an intuitive health score to provide a quantitative basis for maintenance decisions.

[0110] (I) Physical Model Construction and Reasoning For the three core components—traction machine, guide rail, and door system—degradation physical models based on failure mechanisms are established respectively: Traction machine (focusing on bearings and windings): Construct a coupled degradation model of "wear-thermal aging". Based on the bearing wear rate formula (wear amount = k × load × running time / lubrication coefficient, where k is the material wear coefficient), and combined with the load rate and running time data collected by S1, the cumulative wear amount is calculated; at the same time, based on the winding thermal aging model (the exponential relationship between aging rate and temperature), and combined with historical temperature data, the insulation aging degree is calculated. Finally, through the coupling threshold of wear amount and aging degree (design wear limit and insulation failure threshold), the RUL is initially estimated.

[0111] Guide rail (focusing on parallelism and lubrication status): Establish a "friction-deformation" degradation model. Based on the linear relationship between guide rail friction loss and number of runs and load, calculate the friction loss by combining the cumulative number of start-stop cycles and average load rate; inversely infer the deformation by using guide rail deviation characteristics (such as vibration peak offset) in the vibration signal, and output the RUL prediction result at the physical model level by combining the friction loss threshold and deformation limit.

[0112] Door system (focusing on transmission components and motor): Construct a "fatigue-electrical degradation" model. Based on the fatigue life formula for transmission gears (fatigue life is inversely proportional to the cube of the number of start-stop cycles), fatigue damage is calculated by combining the cumulative number of start-stop cycles; the degree of degradation of electrical components is assessed by the harmonic distortion rate in the drive current and the motor temperature trend; and a preliminary predicted value of RUL is obtained by combining the fatigue failure threshold and the electrical degradation limit.

[0113] (II) Data-driven model construction and reasoning A hybrid model of "CNN-LSTM-Attention" was selected to capture the correlation between temporal degradation features and key influencing factors: Model structure design: The input layer receives an 80-dimensional enhanced feature vector, the CNN layer (3 convolutional kernels, size 3×3, number of channels 64, 128, 256) extracts local degradation features, the LSTM layer (2 layers, number of hidden units 256) captures temporal degradation trends, the Attention layer strengthens the feature weights related to the root cause of the fault (e.g., when the root cause chain is "traction machine bearing wear", the attention of features such as vibration kurtosis and temperature mean is increased), and the output layer outputs the RUL prediction value (unit: hour) through linear regression.

[0114] Model Training and Optimization: The training dataset contains 100,000 labeled "feature-degradation-lifetime" samples of the same type of component (including complete time-series data from normal state, minor anomalies, severe anomalies to failure), divided into training, validation, and test sets in a 7:2:1 ratio. The AdamW optimizer is used with an initial learning rate of 0.001, decaying to 0.7 times the original rate every 30 epochs. The loss function is mean squared error (MSE), combined with an early stopping mechanism (stopping if the validation set loss does not decrease for 15 consecutive epochs) to avoid overfitting. The model is deployed in a cloud-based model repository, supporting real-time streaming inference with a single-sample inference latency of ≤100ms.

[0115] (III) Fusion Model Prediction and Health Score Calculation Weighted fusion strategy: Dynamically assign weights based on the prediction accuracy of the physical model and the data-driven model. If the validation set R of the data-driven model... 2 If the prediction accuracy is ≥0.9 (high prediction accuracy), then the weight allocation is 0.7 for the data-driven model and 0.3 for the physical model; if the data-driven model R... 2 For values ​​between 0.7 and 0.9, equal-weighted fusion (0.5:0.5) is used; if R... 2 If the weight is less than 0.7 (if the sample size is insufficient), the physical model will be the primary model (weight 0.7), with data-driven model correction as an auxiliary. The fusion formula is: final RUL = ω1 × RUL physical + ω2 × RUL data (ω1 + ω2 = 1).

[0116] Health rating system: A 100-point scale is used, calculated based on three core indicators: Remaining Service Life (RUL) percentage (weight 0.4, RUL / Design Life × 40), Current Condition Level (weight 0.3, S4 Severity Levels 1-5 correspond to 30, 24, 18, 12, and 6 points respectively), and Degradation Rate (weight 0.3, ≤0.1% / day degradation rate earns 30 points, deducting 5 points for each 0.1% increase, with a minimum of 0 points). A score ≥85 is "Excellent," 70-84 is "Good," 50-69 is "Average," and <50 is "Poor," directly reflecting the component's health status.

[0117] S503, Implementation Plan for a Reinforcement Learning-Based Maintenance Decision Optimization Engine With the three-dimensional optimization goals of "minimizing maintenance costs, maximizing equipment availability, and minimizing safety risks", a dynamic and adaptive maintenance strategy is generated through reinforcement learning algorithms, taking into account both scientific rigor and practicality.

[0118] (a) Definition of core elements of reinforcement learning State space: includes 6 key state parameters - component health score (0-100 points), RUL (0-design life hours), current operating conditions (load rate, operating frequency, ambient temperature), safety risk level (based on S4 severity level mapping: level 1-2 is low risk, level 3 is medium risk, level 4-5 is high risk), resource constraint status (maintenance personnel status, spare parts inventory status, maintenance tool configuration), and elevator operating period (peak hours / off-peak hours / low-peak hours).

[0119] Action Space: Define three types of executable actions to form action combinations—maintenance time window (A1: immediate maintenance, A2: within 12 hours, A3: within 24 hours, A4: within 48 hours, A5: within 72 hours), resource allocation scheme (B1: senior technician + original spare parts, B2: intermediate technician + compatible spare parts, B3: junior technician + repair spare parts), and task priority (C1: highest priority, priority to occupy resources, C2: intermediate priority, arranged in sequence, C3: low priority, maintenance when appropriate).

[0120] Reward function: Based on the comprehensive three-dimensional optimization objective design, the formula is: Reward = α × R_availability + β × R_security - γ × R_cost, where: R Availability: Rewards for improving equipment availability. 100 points are awarded for availability ≥ 99% after maintenance, and 20 points are deducted for every 1% decrease. R Security: Rewards for reducing security risks: 80 points for high risk → low risk, 50 points for medium risk → low risk, and 30 points for maintaining low risk. R Cost: Maintenance cost penalty, 0 points are awarded if the cost is ≤ 80% of the budget, and 20 points are deducted for every 10% overspending; α, β, and γ are weighting coefficients that are dynamically adjusted (β=0.5, α=0.3, γ=0.2 for high-risk scenarios; β=0.3, α=0.4, γ=0.3 for normal scenarios).

[0121] (II) Training and Deployment of Reinforcement Learning Algorithms Algorithm Selection and Training: The proximal policy optimization (PPO) algorithm was chosen to meet the combined optimization requirements of continuous state space and discrete action space. The training environment was built based on historical maintenance data and digital twin simulations, simulating equipment state changes after performing maintenance actions under different states (such as the improvement in health, changes in availability, and cost consumption after maintenance), generating 100,000 simulated interaction samples for training. The training iterated for 1000 rounds, updating the policy network parameters in each round to ensure that the policy converges to the optimal level (the mean of the reward function is stable above 80 points).

[0122] Cloud Deployment and Real-Time Inference: The trained PPO model is deployed to the model repository of the cloud-based intelligent analysis engine and works in conjunction with the RUL prediction model. After receiving the RUL and health score, it reads the current state space parameters in real time, infers and outputs the optimal action combination (maintenance time window, resource configuration, task priority), with an inference latency of ≤150ms, meeting the needs of real-time decision-making.

[0123] (III) Constraints and Dynamic Adjustment Mechanism Key constraints: Maintenance time windows must avoid peak elevator operating hours, unless it is a level 4-5 emergency; if spare parts inventory is insufficient, automatically switch to alternative spare parts configuration scheme, or extend the maintenance time window until spare parts arrive; if maintenance personnel are insufficient, prioritize tasks according to their priority, with the highest priority task occupying resources first.

[0124] Dynamic adjustment logic: If the S4 diagnostic results are updated during maintenance (e.g., severity level is upgraded), reinforcement learning inference is immediately retried to adjust the maintenance strategy (e.g., time window is brought forward, priority is increased); if the resource status changes (e.g. spare parts arrive ahead of schedule, new maintenance personnel are added), the state space is updated in real time and the optimal action combination is recalculated.

[0125] S504, Maintenance Strategy Generation and Output (a) Strategy structured encapsulation The optimized maintenance strategy is output in a structured format, and its core consists of five modules: Key maintenance information includes: target components (such as traction machine bearings), explanation of maintenance necessity (based on health score, RUL, and safety risk level), recommended maintenance time window and time periods to avoid; Resource allocation plan: specify maintenance personnel level, required spare parts models and quantities, tool configuration list, and estimated labor and material costs; Task execution requirements: task priority, suggested execution time, key points of maintenance process (based on S4 fault root cause chain, such as "replace the bearing first, then calibrate the traction machine vibration parameters"). Risk prevention and control tips: Key parameters to be monitored during maintenance (such as vibration acceleration and temperature changes after maintenance), potential secondary faults and corresponding countermeasures; Expected results: Post-maintenance health score prediction, RUL extension range, and equipment availability improvement target.

[0126] (ii) Multi-channel output and interaction Application service layer push: The structured maintenance strategy is pushed to the application service layer in real time to generate standardized maintenance work orders, which include work order number, execution period, responsible person assignment, process nodes, and support online reception, confirmation and feedback by operation and maintenance personnel.

[0127] Visualization: The health report at the application service layer simultaneously displays maintenance strategy recommendations, combined with RUL visualization charts (such as remaining lifespan countdown and health status change curves) to help users intuitively understand the necessity of maintenance.

[0128] Emergency Warning: For the highest priority maintenance strategy corresponding to severe anomalies of level 4-5, in addition to work order push, dual alarms will be sent via SMS and APP to notify the operation and maintenance manager and elevator management to ensure that emergency maintenance is initiated in a timely manner.

[0129] This invention provides an elevator IoT anomaly diagnosis data management system, comprising: The physical sensing layer includes a multimodal sensor network deployed in the elevator to collect vibration, acoustic, current, and temperature data; The data acquisition and edge processing layer includes an edge computing node, which is equipped with a data alignment module, a preprocessing module and a lightweight anomaly detection model for synchronizing, cleaning, compressing and performing preliminary analysis on sensor data. The cloud-based data management and intelligent analysis layer includes: A time-series database is used to store time-series data from the edge processing layer; Feature engineering and fusion module, used to extract and fuse multi-source features; A model repository stores and manages anomaly diagnosis models, lifetime prediction models, and decision optimization models. The intelligent analytics engine is used to perform dynamic anomaly diagnosis and predictive maintenance decision calculations; The application service layer provides users with an interactive interface for real-time monitoring, fault alarms, health reports, and the generation and management of maintenance work orders.

[0130] Furthermore, in the above technical solution, the health report provided by the application service layer includes a health score based on multi-dimensional indicators, visualization of the remaining service life of key components, and maintenance strategy recommendations.

[0131] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for managing elevator IoT anomaly diagnostic data, characterized in that, Specifically, the following steps are included: S1. Collect heterogeneous time-series data through a multimodal sensor network deployed in the elevator, and use a millisecond-level data alignment mechanism based on device clock synchronization to align the heterogeneous time-series data in time sequence. In addition, adopt the "continuous acquisition + event trigger enhancement" mode to trigger the sensor to temporarily increase the sampling rate to capture instantaneous features when the elevator starts or stops or the load changes suddenly. S2. A lightweight anomaly detection model deployed at the edge is used to perform preliminary real-time anomaly detection on the aligned heterogeneous time series data, and the aligned heterogeneous time series data is cleaned, quality checked and adaptively compressed in real time at the edge computing node. S3. Upload the edge-processed data to the cloud, perform feature engineering to extract multi-source features, and use an attention-based feature fusion network to intelligently weight and fuse the multi-source features to construct a unified feature vector. S4. Input the unified feature vector into the dynamic anomaly diagnosis model to perform dynamic anomaly diagnosis, and output the anomaly type, severity level and fault root cause chain. The dynamic anomaly diagnosis model includes an early anomaly identification module based on a bidirectional long short-term memory network (Bi-LSTM) and a root cause analysis engine based on a fault propagation graph. S5. Based on historical anomaly data, equipment operating data, and unified feature vectors, the remaining useful life (RUL) and health score of key components are calculated by fusing physical models and data-driven remaining useful life prediction models. The fusion weights are dynamically allocated according to the prediction accuracy of the data-driven model. Furthermore, an optimization engine based on reinforcement learning is used to generate optimized maintenance strategies that include maintenance time windows, resource configuration, and task priorities.

2. The elevator IoT anomaly diagnosis data management method according to claim 1, characterized in that: The heterogeneous time-series data includes vibration acceleration signals, acoustic feature signals, drive current signals, and temperature data of key components. The key components include traction machines, guide rails, and door systems. The multi-source features include time-domain, frequency-domain, and time-frequency domain combined features.

3. The elevator IoT anomaly diagnosis data management method according to claim 1, characterized in that: The dynamic anomaly diagnosis also includes: dynamically adjusting the alarm thresholds for various anomalies based on the elevator's operating conditions.

4. The elevator IoT anomaly diagnosis data management method according to claim 1, characterized in that: The adaptive compression specifically involves dynamically adjusting the data sampling frequency and the amount of data transmitted based on the current operating status of the elevator and the preliminary anomaly detection results.

5. The elevator IoT anomaly diagnosis data management method according to claim 1, characterized in that: The maintenance decision optimization engine has optimization objectives including minimizing maintenance costs, maximizing equipment availability, and minimizing security risks.

6. The elevator IoT anomaly diagnosis data management method according to claim 1, characterized in that: The early anomaly recognition module based on Bi-LSTM network adopts a 3-layer Bi-LSTM structure with 128 hidden layer units. It processes feature vectors through a sliding time window (window size of 10s) to achieve early anomaly recognition 3-5 minutes in advance and outputs a preliminary judgment of the anomaly type. The root cause analysis engine based on the fault propagation graph is used to construct the fault map associated with elevator components (containing 12 types of core faults and 32 propagation paths). Combined with the early identification results, it traces back the source of the fault and finally outputs the accurate anomaly type, severity level 1-5, and fault root cause chain containing 3-5 nodes.

7. An elevator IoT anomaly diagnosis data management system, applied to the elevator IoT anomaly diagnosis data management method according to any one of claims 1-6, characterized in that, include: The physical sensing layer includes a multimodal sensor network deployed in the elevator to collect vibration, acoustic, current, and temperature data; The data acquisition and edge processing layer includes an edge computing node, which is equipped with a data alignment module, a preprocessing module and a lightweight anomaly detection model for synchronizing, cleaning, compressing and performing preliminary analysis on sensor data. The cloud-based data management and intelligent analysis layer includes: A time-series database is used to store time-series data from the edge processing layer; Feature engineering and fusion module, used to extract and fuse multi-source features; A model repository stores and manages anomaly diagnosis models, lifetime prediction models, and decision optimization models. The intelligent analytics engine is used to perform dynamic anomaly diagnosis and predictive maintenance decision calculations; The application service layer provides users with an interactive interface for real-time monitoring, fault alarms, health reports, and the generation and management of maintenance work orders.

8. The elevator IoT anomaly diagnosis data management system according to claim 7, characterized in that: The health report provided by the application service layer includes a health score based on multi-dimensional indicators, visualization of the remaining service life of key components, and maintenance strategy recommendations.

Images