Method and system for diagnosing running state of elevator traction machine in real time based on high-frequency sampling

Abstract

The invention relates to the technical field of elevator equipment state monitoring and fault diagnosis, and discloses an elevator traction machine running state real-time diagnosis method and system based on high-frequency sampling. According to the method, vibration (larger than or equal to 20 kHz), current (larger than or equal to 10 kHz), sound / sound emission, temperature and rotating speed signals of a traction machine are synchronously collected through a high-frequency multi-mode sensor array; capturing early weak fault transient characteristics; the edge computing unit completes data preprocessing, time synchronization, feature extraction and anomaly detection, and uploads key data to a cloud end through cloud-edge collaboration; the cloud end adopts a working condition self-adaptive strategy and a multi-modal fusion model to carry out deep diagnosis, and outputs fault types, positions and grades; and combining incremental learning and a degradation model to realize health quantification and residual life prediction. Through fusion of high-frequency data capture and an intelligent algorithm, the early fault detection capability, variable working condition adaptability and diagnosis real-time performance of the traction machine are improved, and a solution is provided for predictive maintenance of an elevator.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of elevator equipment state monitoring and fault diagnosis, and more particularly to a real-time diagnosis method and system for the running state of an elevator hoisting machine based on high-frequency sampling. BACKGROUND

[0002] As an indispensable vertical transportation tool in modern buildings, the safe and stable operation of elevators is directly related to the safety of people's lives and property and travel efficiency. As the core power component of an elevator, a hoisting machine bears the key function of driving the car to ascend and descend, and the monitoring and diagnosis of its running state is a core link to ensure the reliability of the elevator. With the increase of the service life of the elevator and the increase of the running intensity, the hoisting machine is prone to failure due to mechanical wear (such as bearing peeling, gear pitting), electrical aging (such as motor inter-turn short circuit, winding insulation failure) or assembly defects (such as loose coupling), etc. If not discovered in time, it may lead to shutdown, trapping people or even more serious safety accidents.

[0003] The traditional hoisting machine state monitoring and diagnosis technology has many limitations and cannot meet the needs of modern elevator predictive maintenance, mainly in the following aspects:

[0004] Insufficient sampling rate, missing early weak fault features: Traditional methods mostly use lower sampling rates (usually ≤1kHz) to collect vibration, current and other signals. However, early weak faults (such as bearing rolling element peeling, gear tooth surface initial pitting, motor stator winding early insulation aging) are mainly characterized by high-frequency transient impact or modulated signals (frequencies are mostly above 2kHz), and low-frequency sampling will cause these key features to be filtered out or aliased, and only when the fault develops to the middle and late stages (the characteristic signal energy is strong enough) can it be detected, missing the early warning window.

[0005] Limited information from single modal data, poor diagnosis robustness: Existing technologies mostly rely on a single physical quantity (such as vibration or current) for monitoring. However, hoisting machine faults often have cross-physical quantity coupling characteristics: for example, mechanical looseness may cause abnormal vibration and distortion of sound signals, and motor electrical faults may affect current harmonics and vibration modulation characteristics. Single modal data cannot fully reflect the nature of the fault and is easily affected by environmental interference (such as car vibration transmission, power grid noise), resulting in high false alarm and missed alarm rates.

[0006] It is difficult to adapt to variable working condition interference, and the diagnosis accuracy is limited: during the operation of the elevator, the speed (start-stop, acceleration-deceleration, constant speed stage) and load (empty load, full load, partial load) change significantly. Under different working conditions, even if the equipment is in normal state, the signal characteristics (such as vibration amplitude, current effective value) will also have obvious differences. The traditional diagnosis method mainly uses fixed threshold or single model, and does not consider the influence of working condition dynamic change on signal characteristics. In the variable working condition scene, it is easy to misjudge as a fault due to "normal fluctuation", or miss the judgment due to "fault characteristics are covered by working condition noise".

[0007] Data processing and diagnosis real-time performance is insufficient: the traditional scheme mainly relies on the "edge collection-cloud centralized analysis" mode, and high-frequency data (if collected) is directly uploaded to the cloud, which will occupy a large amount of communication bandwidth, and the cloud centralized calculation has a delay (usually ≥10s), which is difficult to realize real-time alarm and rapid response. For elevators and other devices with high real-time requirements, the delay may cause the fault to continue to expand, increasing the difficulty of maintenance.

[0008] Lack of fault evolution tracking and life prediction capability: the existing technology is mainly limited to "fault / normal" binary classification diagnosis, which cannot quantify the severity of the fault, and it is more difficult to predict the remaining useful life (RUL) based on the dynamic evolution trend of the fault characteristics. This leads to the maintenance strategy still mainly relying on "after-maintenance" or "periodic overhaul", which is easy to cause the problems of over-maintenance (increasing cost) or insufficient maintenance (leaving risk).

[0009] Based on this, the present application provides a real-time diagnosis method and system for the running state of an elevator traction machine based on high-frequency sampling to solve the above problems. SUMMARY

[0010] In order to overcome the above-mentioned defects of the prior art, the present application provides a real-time diagnosis method and system for the running state of an elevator traction machine based on high-frequency sampling to solve the problems existing in the background art.

[0011] The present application provides the following technical scheme: a real-time diagnosis method for the running state of an elevator traction machine based on high-frequency sampling, comprising the following steps:

[0012] Step 1: high-frequency multi-modal data collection: through the sensor array deployed at the key parts of the traction machine, vibration signals, three-phase current signals, sound signals or acoustic emission signals, temperature signals and speed or encoder signals are synchronously collected to form an original high-frequency data set, wherein the vibration signal sampling frequency is ≥20kHz, the current signal sampling frequency is ≥10kHz, and the sound signal or acoustic emission signal sampling frequency is ≥44.1kHz;

[0013] Step 2: Edge-side real-time preprocessing: The edge computing unit performs anti-aliasing filtering, digital noise reduction, time synchronization, and working condition parameter extraction on the original high-frequency data, performs real-time anomaly detection based on preset thresholds, performs feature compression on normal data, marks the time stamp of abnormal data, and retains the key segments;

[0014] Step 3: High-frequency transient feature extraction: For the preprocessed high-frequency signal, the variational mode decomposition (VMD) and wavelet packet transform (WPT) are used to separate the transient impact components, the envelope demodulation is used to amplify the fault features, and the one-dimensional convolutional neural network (1D-CNN) is used to extract deep discriminative features.

[0015] Step 4: Cloud-edge collaborative diagnostic analysis: The edge side uploads the compressed features, anomaly markers, and working condition parameters to the cloud, the cloud identifies the current operating condition based on the speed / current signal, calls the corresponding working condition sub-diagnostic model, and outputs the fault type, location, and severity level through the attention mechanism to fuse multi-modal features.

[0016] Step 5: Incremental learning and life prediction: The cloud builds a health index (HI) based on historical high-frequency data, updates the diagnostic model using an incremental deep learning algorithm, and predicts the remaining useful life (RUL) of the traction machine by combining the fault feature evolution trend and the degradation model.

[0017] Through high-frequency multi-modal data acquisition to capture early weak fault transient features, combined with edge preprocessing to reduce data transmission pressure, intelligent feature extraction and cloud-edge collaborative diagnosis to achieve accurate fault identification, and finally through incremental learning and life prediction to support predictive maintenance, the problems of traditional methods such as "early fault difficult to find, lack of real-time performance, and lack of life prediction" are solved.

[0018] As a further scheme of the present application: the deployment strategy of the sensor array in step 1 is to arrange three-axis vibration acceleration sensors on the traction machine bearing seat, install wideband current transformers on the motor power supply harness, set up acoustic emission sensors on the gear box shell, and install temperature sensors on the stator winding and bearing outer ring, and connect the traction machine encoder signal.

[0019] The targeted deployment strategy of the sensor array ensures the effective collection of high-frequency signals such as vibration, current, and acoustic emission at key fault locations (bearings, gearboxes, and motors), avoids feature loss caused by signal propagation attenuation, provides high-quality raw data for subsequent multi-modal fusion, and improves fault location accuracy to the component level.

[0020] As a further scheme of the present application: the time synchronization in step 2 is achieved through hardware triggering or high-precision time stamping, with a time error ≤1ms; the working condition parameter extraction includes calculating the real-time speed based on the encoder signal and estimating the load torque based on the current signal.

[0021] High-precision time synchronization (error ≤1 ms) ensures the time correlation of multi-modal data, ensuring that the coupling characteristics of vibration and current, sound signals can be effectively mined; real-time working condition parameter extraction provides a basis for variable working condition adaptive diagnosis, reduces the interference of speed / load fluctuations on the diagnosis results, and reduces the false alarm rate.

[0022] As a further scheme of the application: the variational mode decomposition (VMD) in step 3 enhances the separation ability of weak impact signals by adaptively selecting the number of mode decompositions and the penalty factor; envelope demodulation is performed on the high-frequency resonance components obtained by VMD decomposition to extract fault modulation features.

[0023] The optimized VMD algorithm enhances the separation ability of weak impact signals, and the envelope demodulation effectively amplifies the fault modulation features in the high-frequency resonance band, solving the problem that early fault signals are overwhelmed by noise, and advancing the detection time of bearing micro-peeling, gear slight pitting and other faults.

[0024] As a further scheme of the application: the working condition adaptive diagnosis in step 4 includes: establishing a speed-load working condition dictionary, and realizing working condition recognition through a clustering algorithm; converting the vibration signal to the angle domain for order analysis to eliminate the speed influence; training a special sub-model or dynamically adjusting the model parameters for different working conditions.

[0025] The working condition adaptive diagnosis eliminates the influence of speed / load changes on signal features through working condition recognition, order analysis and multi-working condition sub-models, ensures the diagnosis accuracy under complex working conditions such as start-stop, acceleration-deceleration, light-heavy load, and significantly improves the engineering practicability of the method.

[0026] As a further scheme of the application: the health index (HI) in step 5 is constructed based on high-frequency features, including specific frequency band energy proportion, envelope spectrum peak factor, and fault probability value output by a deep learning model; the degradation model adopts a hybrid model combining an exponential model and a machine learning regression.

[0027] The health index (HI) constructed based on high-frequency features can quantify the equipment degradation state, and the degradation model combining an exponential model and a machine learning regression improves the prediction accuracy of the remaining useful life (RUL).

[0028] As a further scheme of the application: the elevator traction machine running state real-time diagnosis system based on high-frequency sampling includes:

[0029] Intelligent high-frequency data acquisition module: integrates three-axis vibration sensors, wideband current transformers, acoustic emission sensors, temperature sensors and signal conditioning circuits, supports vibration signal sampling frequency ≥20 kHz, current signal sampling frequency ≥10 kHz, sound / acoustic emission signal sampling frequency ≥44.1 kHz, and the sampling rate can be dynamically configured;

[0030] Edge computing unit: uses a processor with a DSP core, runs a real-time operating system, realizes data preprocessing, time synchronization, lightweight anomaly detection algorithm and edge feature calculation, supports industrial bus and wireless communication;

[0031] Cloud diagnosis platform: including data storage server, model training server and diagnosis engine, deploying multi-modal fusion model, working condition recognition model, incremental learning module and RUL prediction module, providing API interface and visual interface;

[0032] Human-computer interaction terminal: shows real-time diagnosis results, fault warning information, health index trend and residual life prediction, supports alarm notification, original signal spectrum viewing and maintenance work order generation.

[0033] The intelligent high-frequency acquisition module ensures the high-quality synchronous acquisition of multi-modal signals, the edge computing unit realizes real-time preprocessing and lightweight diagnosis, the cloud platform supports deep analysis and model optimization, the human-computer interaction terminal provides intuitive diagnosis result display and maintenance decision support, and the system modules work together to balance real-time, accuracy and ease of use, meeting the long-term online monitoring needs of elevator traction machines.

[0034] As a further scheme of the application: the intelligent high-frequency data acquisition module adopts modular design, built-in anti-aliasing filter and self-calibration circuit, supports POE power supply or low-power battery power supply, and the anti-aliasing filter cutoff frequency is configurable.

[0035] The modular design of the acquisition module and the configurable anti-aliasing filter adapt to the monitoring needs of different types of traction machines, the self-calibration circuit ensures long-term operation accuracy, the POE / low-power power supply mode adapts to the complex environment of the elevator shaft, and reduces the installation and maintenance cost.

[0036] As a further scheme of the application: the edge computing unit and the cloud diagnosis platform adopt an adaptive data uploading strategy: under normal conditions, compressed features and working condition parameters are uploaded once every 5-10 minutes; when an anomaly is detected, the abnormal segment raw data and detailed features are uploaded in real time.

[0037] The adaptive data uploading strategy reduces the data transmission volume under normal conditions, and uploads key data in priority when an anomaly is detected, balancing real-time and economy, and avoiding bandwidth congestion and delay caused by full uploading of high-frequency data.

[0038] As a further scheme of the application: the cloud diagnosis platform supports model hot update, continuously optimizes the diagnosis model through the incremental learning module, regularly issues the updated lightweight model to the edge computing unit, and realizes the co-evolution of edge and cloud models.

[0039] The cloud model hot update and edge-cloud co-evolution mechanism enable the diagnostic model to continuously adapt to equipment performance degradation and new fault modes, maintain long-term diagnostic accuracy without on-site hardware upgrades, extend the effective service life of the system, and reduce the total life cycle cost.

[0040] Technical effects and advantages of the present application:

[0041] The present application provides a systematic solution for the core requirements of elevator traction machine state diagnosis through the organic combination of high-frequency sampling, multi-modal fusion, cloud-edge collaboration and intelligent algorithms, which has the following beneficial effects:

[0042] Early weak fault capture capability is significantly improved: through high-frequency multi-modal data acquisition (vibration, current, sound / sound emission, etc.), transient fault features (such as high-frequency impact, modulation signals) that cannot be identified by traditional low-frequency sampling are accurately captured, and effective identification of early weak faults of core components such as traction machine bearings, gears and motors is achieved, breaking through the limitations of traditional methods that can only detect faults in the middle and late stages.

[0043] Diagnosis accuracy and robustness are enhanced: a multi-modal feature fusion strategy is adopted, combined with optimized time-frequency analysis and deep learning feature extraction technology, to fully utilize the complementary information of different physical quantity signals, reduce misjudgment caused by single signal interference, and improve the identification accuracy of fault type and location under complex working conditions.

[0044] Real-time and economic balance: through the cloud-edge collaboration architecture, the edge side realizes real-time preprocessing, lightweight diagnosis and data compression of high-frequency data, the cloud side focuses on deep analysis and model optimization, effectively solving the contradiction between large amount of high-frequency data and real-time requirements, reducing the data transmission bandwidth demand, and meeting the real-time monitoring requirements of elevator operation.

[0045] Strong adaptability to variable working conditions: through working condition recognition, order analysis and adaptive diagnosis strategy, the influence of speed, load changes on signal features is dynamically eliminated, ensuring the stability of the diagnosis results under complex operating conditions such as start-stop, acceleration-deceleration, light-heavy load, and improving the engineering practicability of the technical solution.

[0046] Supporting the construction of predictive maintenance system: through the quantification of equipment degradation state by health indicators, combined with incremental learning and life prediction technology, the fault evolution trend tracking and remaining useful life assessment are realized, providing technical support for the predictive maintenance of elevator traction machines, and promoting the transformation of maintenance mode from passive response to active prevention.

[0047] Excellent system scalability and applicability: modular hardware design, configurable sampling parameters and adaptive model update mechanism adapt to the monitoring needs of different types of traction machines (geared / non-geared, asynchronous / synchronous motors), and are easy to integrate with existing elevator monitoring systems, reducing deployment and upgrade costs. BRIEF DESCRIPTION OF DRAWINGS

[0048] The application will be further described below with reference to the drawings.

[0049] Figure 1 is the method flow chart of the real-time diagnosis method of the elevator hoisting machine operation state based on high-frequency sampling of the application;

[0050] Figure 2 is the system block diagram of the real-time diagnosis system of the elevator hoisting machine operation state based on high-frequency sampling of the application. DETAILED DESCRIPTION

[0051] The technical solutions of the application will be described in detail below with reference to the drawings and specific embodiments. The embodiments are only used to explain the application and do not limit the protection scope of the application.

[0052] Please refer to Figure 1 The application provides a real-time diagnosis method of the elevator hoisting machine operation state based on high-frequency sampling, which comprises the following steps:

[0053] Step 1: High-frequency multi-modal data acquisition

[0054] Synchronous acquisition of multi-physical quantity high-frequency signals at key positions of the hoisting machine through a sensor array to build a raw data base. The specific implementation mode is as follows:

[0055] Sensor selection and deployment:

[0056] Vibration sensor: a three-axis acceleration sensor (frequency response range 0.1 Hz-50 kHz, sensitivity 100 mV / g, resolution ≤0.001 g) is selected and arranged at the input end bearing seat of the hoisting machine (to collect the vibration of the motor side bearing), the output end bearing seat (to collect the vibration of the gear box side bearing) and the shell position away from the bearing of the gear box (to collect the gear meshing vibration), to ensure that the impact signals in the radial (horizontal X / Y) and axial (Z) directions are covered.

[0057] Current sensor: a wide-band current transformer (bandwidth DC-100 kHz, measurement range 0-50 A, accuracy 0.5 level) or a Rogowski coil (bandwidth DC-50 kHz, response time ≤1 μs) is selected and installed at the incoming end of the motor three-phase power supply line to collect the current transient change and harmonic characteristics.

[0058] Sound / acoustic emission sensor: an acoustic emission sensor (frequency response 10 kHz-1 MHz, sensitivity -75 dB±3 dB) or a microphone (sampling rate above 44.1 kHz, signal-to-noise ratio ≥60 dB) is selected and arranged at the gear box shell or the motor end cover to capture high-frequency sound radiation signals (such as abnormal gear meshing sound and bearing friction sound).

[0059] Temperature sensor: PT100 platinum resistance (measurement range -50℃ ~ 200℃, accuracy ± 0.1℃) or thermocouple is selected, installed in the motor stator winding (monitoring winding temperature) and bearing outer ring (monitoring bearing temperature), auxiliary judgment overheating failure.

[0060] Rotational speed / encoder signal: directly access the incremental encoder signal (resolution ≥ 1024 lines / revolution) of the traction machine, used for calculating real-time rotational speed and angular position.

[0061] Sampling parameter configuration:

[0062] Vibration signal sampling frequency ≥ 20kHz (preferably 25.6kHz, ensuring coverage of 5-20kHz fault impact frequency band); current signal sampling frequency ≥ 10kHz (preferably 12.8kHz, meeting the 5kHz harmonic analysis requirement); sound / acoustic emission signal sampling frequency ≥ 44.1kHz (preferably 48kHz, covering high-frequency sound characteristics); temperature signal sampling frequency 1-10Hz (satisfying temperature slow change monitoring); encoder signal acquisition frequency matches motor rotational speed (ensuring ≥ 1024 points per revolution).

[0063] Synchronous acquisition control: through hardware synchronization trigger (such as synchronization pulse signal) or high-precision timestamp (based on GPS / Beidou or local high-stability clock, time error ≤ 1ms) of intelligent acquisition module, realize time alignment of multi-sensor data, ensure correlation of vibration, current, sound and other signals in the same time dimension.

[0064] Step 2: real-time preprocessing on the edge side

[0065] The edge computing unit cleans, denoises and preliminarily analyzes the original high-frequency data, reduces data volume and realizes real-time abnormality filtering. The specific process is as follows:

[0066] Anti-aliasing and digital filtering:

[0067] The original signal is processed by the analog anti-aliasing filter built-in or in the acquisition module (vibration signal cutoff frequency is set to 0.4-0.5 times of the sampling rate, such as 25.6kHz sampling corresponding to 12kHz cutoff frequency; current signal cutoff frequency is set to 5-10kHz), to prevent high-frequency signal aliasing to the effective frequency band; the edge side further removes 50Hz power frequency interference and shaft environment noise through digital filtering (such as 8th order Butterworth low-pass filter, adaptive notch filter).

[0068] Denoising: for random noise in vibration and current signals, wavelet threshold denoising (selecting db4 or sym8 wavelet basis, 3-5 layers of decomposition, using soft threshold function) or ensemble empirical mode decomposition (EEMD) denoising is adopted, which improves the signal-to-noise ratio by 10-15dB, highlighting the effective signal components.

[0069] Time synchronization correction: Based on the synchronization trigger signal or timestamp in step 1, time alignment correction is performed on the multi-modal data (achieved by interpolation or resampling), ensuring that the vibration, current, and sound signals at the same time have a deviation of ≤1ms on the time axis, providing a basis for subsequent multi-modal fusion.

[0070] Working condition parameter extraction:

[0071] Speed calculation: Based on the pulse interval of the encoder signal, the real-time speed is calculated by the M / T method (formula: speed n = 60 x f_pulse / N, where f_pulse is the pulse frequency and N is the number of encoder lines), with a resolution of ≤0.1 rpm and a sampling period of 10ms.

[0072] Load estimation: Based on the effective value of the three-phase current signal (I = √(I_a 2 + I_b 2 + I_c 2 ) / √3), combined with motor parameters (rated current, rated load), a load estimation model is established (load T = k x I, k is the calibration coefficient).

[0073] Lightweight anomaly detection and data compression:

[0074] Edge-side running of anomaly detection algorithm based on statistical features: Calculate the kurtosis and crest factor of the vibration signal, the total harmonic distortion (THD) and negative sequence component of the current signal, and set a dynamic threshold (based on the 3σ principle under normal working conditions); when any feature exceeds the threshold, it is marked as an abnormal state, and the original data segment (including vibration, current, and sound signals) 5-30 seconds before and after the abnormal moment is retained; under normal conditions, only compressed features (such as vibration spectrum peak, current harmonic amplitude, temperature value, etc.) are calculated and saved, with a single channel data volume of ≤1MB per hour, greatly reducing the upload data volume.

[0075] Step 3: High-frequency transient feature extraction

[0076] For the pre-processed high-frequency signal, a combination of traditional signal processing and deep learning is used to extract transient features that can represent early faults. The specific implementation is:

[0077] Transient impact separation based on VMD and WPT:

[0078] Optimized VMD algorithm is used for vibration signal: the number of modes K = 5-8 (self-adaptive adjustment according to signal complexity), penalty factor a = 1000-3000 (larger value for high frequency signal), the center frequency and bandwidth of each modal component are minimized through iterative optimization, and the high frequency resonance component (usually 8-20 kHz frequency band) containing fault impact is separated; the resonance component is decomposed into 3-4 layers by wavelet packet transform (WPT), and the energy proportion of each sub-band is calculated to locate the fault sensitive frequency band.

[0079] Envelope demodulation and modulation feature extraction:

[0080] The high frequency resonance component separated by VMD is subjected to Hilbert transform to obtain the envelope curve of the analytical signal, and then Fourier transform is performed on the envelope curve to obtain the envelope spectrum, and the fault characteristic frequency in the envelope spectrum (such as bearing fault characteristic frequency f_b = 0.5 x n x (1-d / D) x Z, where n is the rotating speed, d is the rolling element diameter, D is the pitch diameter, and Z is the number of rolling elements; the gear fault characteristic frequency is the meshing frequency f_m = Z x n / 60 and its sideband) is extracted. For example, when the outer ring of the bearing fails, the envelope spectrum will have peak values of f_b and its harmonics; when the gear pitting occurs, the envelope spectrum will have sideband peak values of f_m ± k x n / 60 (k = 1, 2,...).

[0081] Deep feature learning based on 1D-CNN:

[0082] A one-dimensional convolutional neural network (1D-CNN) is constructed to learn discriminative features directly from the original high-frequency vibration / current signal: the network input is a 1024-2048 point signal segment (length corresponding to 40-80 ms data), which is sequentially processed through 3-5 convolution layers (convolution kernel size 3-7, step size 1-2, activation function ReLU), 2-3 max pooling layers (pooling kernel size 2-4), and 1 fully connected layer (64-128 neurons), and outputs a high-dimensional deep feature vector. The network parameters are optimized through back propagation to make the output features effectively distinguish between normal state and different fault types (such as bearing fault, gear fault, and electrical fault).

[0083] Step 4: Cloud-edge collaborative diagnosis analysis

[0084] Edge side and cloud end work together to realize self-adaptive multi-modal fusion diagnosis. The specific process is as follows:

[0085] Data upload strategy: The edge side uploads the following data to the cloud: compressed features (including vibration spectrum features, current harmonic features, working condition parameters, temperature values) are uploaded every 5-10 minutes under normal conditions; abnormal markers, 30-second raw data segments, detailed feature vectors, and working condition parameters are uploaded in real time under abnormal conditions to ensure that critical information is not lost.

[0086] Cloud working condition recognition: The cloud constructs a working condition dictionary based on the speed and load parameters uploaded by the edge through the K-means clustering algorithm: the speed is divided into low speed (0-500 rpm), medium speed (500-1000 rpm), and high speed (1000-1500 rpm), and the load is divided into light load (0-300 kg), medium load (300-700 kg), and heavy load (700-1000 kg), which are combined into 9 typical working conditions; the real-time working condition parameters are matched to determine the current working condition category (such as "high-speed heavy load").

[0087] Working condition adaptive signal processing: For vibration signals, the time domain signal is converted to the angle domain (with the rotation angle as the horizontal axis) through order analysis to eliminate the influence of speed fluctuations on characteristic frequencies (such as stabilizing the gear meshing frequency at the Z order in the order domain to avoid frequency domain drift caused by speed changes); for current signals, normalization processing (divided by the rated current) is used to eliminate the influence of load differences on characteristic amplitudes.

[0088] Multi-modal feature fusion and diagnosis: The cloud calls the corresponding working condition sub-diagnosis model (each working condition is pre-trained with a dedicated sub-model), and uses a graph neural network (GNN) with attention mechanism to fuse multi-modal features: vibration features (32 dimensions), current features (24 dimensions), and sound features (16 dimensions) are used as graph nodes, and the attention weight matrix is used to dynamically allocate the weights of each modality (the weight of the fault-sensitive modality is increased by 2-3 times), enhancing the model's attention to key features; the model output layer uses a softmax activation function to output the fault type (bearing outer ring / inner ring / rolling element fault, gear pitting / teeth breaking, motor inter-turn short circuit / rotor broken bar, etc.), fault location (accurate to the component level), and severity level (mild: feature amplitude < 2 times the normal mean; moderate: 2-5 times; severe: > 5 times).

[0089] Step 5: Incremental learning and life prediction

[0090] The cloud continuously optimizes the model based on historical data and predicts the remaining useful life. The specific implementation is as follows:

[0091] Health Indicator (HI) Construction: High-frequency features sensitive to faults are selected to construct health indicators, including: fault frequency band energy ratio of VMD decomposition (weight 0.3), envelope spectrum peak factor (weight 0.3), fault probability value output by 1D-CNN (weight 0.2), and temperature change rate (weight 0.2). HI is mapped to the [0,1] interval through normalization (HI = (x - x_min) / (x_max - x_min)), where HI = 0 represents a healthy state and HI = 1 represents a fault threshold.

[0092] Incremental learning model update: The Elastic Weight Consolidation (EWC) algorithm is used to update the diagnostic model online: Every 50 hours of accumulated running data (including normal and fault samples), the model parameters are fine-tuned to adapt to the slow degradation of equipment performance while retaining historical knowledge (by constraining key parameters through regularization terms); a full model retraining is performed every 3 months to ensure that the model's diagnostic accuracy is ≥95%.

[0093] Remaining useful life (RUL) prediction: RUL prediction is based on the evolution trend of HI and a hybrid degradation model. The Wiener process is used to describe the stochastic degradation trend of HI (μ(t) = a × t^b, where a and b are model parameters). Historical HI data is fitted by maximum likelihood estimation. The prediction results are corrected by integrating a random forest regression model (input is the current HI value, rotational speed, and load). Finally, the RUL and confidence interval are output.

[0094] Please see Figure 2 As shown, the present invention also provides a system for implementing the above method, comprising the following core modules:

[0095] Intelligent high-frequency data acquisition module

[0096] Hardware Integration: Employing a modular design, it includes multi-channel signal conditioning circuitry (vibration, current, sound / acoustic emission, temperature), a 24-bit high-precision ADC (sampling rate adjustable from 1kHz to 100kHz), an anti-aliasing filter (cutoff frequency configurable via software), and a self-calibration circuit (periodically calibrating gain and offset errors). The housing is made of aluminum alloy with an IP65 protection rating, suitable for humid and dusty environments in elevator shafts.

[0097] Power supply and communication: Supports PoE power supply (IEEE802.3af standard, power consumption ≤5W) or DC24V power supply, with built-in backup battery (battery life ≥4 hours); communication interfaces include EtherCAT (industrial real-time bus), CAN bus (connecting to elevator control cabinet), and 4G / NB-IoT wireless module (uploading data to the cloud) to ensure stable data transmission.

[0098] Edge computing unit

[0099] Hardware configuration: Industrial-grade processor (such as NXP i.MX8MPlus, containing 4-core ARM Cortex-A53 + 1-core Cortex-M7 + DSP core), main frequency ≥ 1.8GHz, memory ≥ 2GB LPDDR4, storage ≥ 16GB eMMC (expandable to 128GB), supporting -40℃ ~ 70℃ wide temperature working.

[0100] Software function: Running Linux RT_PREEMPT real-time operating system (real-time ≤ 10ms), deploying data preprocessing module (filtering, noise reduction, synchronization), lightweight algorithm module (anomaly detection, feature calculation), data compression module (normal data compression ratio ≥ 100:1), communication management module (adaptive upload strategy).

[0101] Cloud diagnosis platform

[0102] Architecture design: Adopting micro-service architecture deployed in cloud server (such as Aliyun ECS or private cloud), including:

[0103] Data storage service: InfluxDB time series database (store high-frequency raw data and features, single-node write rate ≥ 100,000 points / second), MySQL relational database (store device archives, diagnosis results, work order information);

[0104] Model training service: Based on TensorFlow / PyTorch framework, configure GPU acceleration card (such as NVIDIA Tesla V100), support training and optimization of multi-modal fusion model, working condition recognition model, RUL prediction model;

[0105] Diagnosis engine service: Deploy real-time diagnosis algorithm, response time ≤ 500ms, support 1000+ devices parallel diagnosis;

[0106] Incremental learning service: Regularly analyze new data, generate model update package (size ≤ 10MB), and distribute it to the edge unit through OTA.

[0107] Interface support: Provide RESTful API and MQTT protocol interface, can be seamlessly integrated with elevator remote monitoring system (such as elevator Internet of Things platform).

[0108] Human-computer interaction terminal

[0109] Function design: Including Web management platform and mobile APP:

[0110] Real-time monitoring: Show traction machine running status (speed, load, temperature, HI value), diagnosis results (fault type, location, level), support device list and map view;

[0111] Early warning notice: serious faults are notified to maintenance personnel through SMS, APP push, sound and light alarm, etc., with fault characteristic atlas (time domain waveform, frequency spectrum, envelope spectrum);

[0112] Historical analysis: provides HI trend curve, fault statistical report, characteristic parameter comparison analysis function, supports original data segment download and offline analysis;

[0113] Maintenance management: automatically generates maintenance work order (including fault description, recommended measures, spare parts list), tracks work order processing progress, and records maintenance history.

[0114] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for real-time diagnosis of elevator traction machine operating status based on high-frequency sampling, characterized in that, Includes the following steps: Step 1: High-frequency multimodal data acquisition: Through sensor arrays deployed in key parts of the traction machine, vibration signals, three-phase current signals, sound signals or acoustic emission signals, temperature signals, and speed or encoder signals are collected synchronously to form a raw high-frequency dataset, wherein the sampling frequency of vibration signals is ≥20kHz, the sampling frequency of current signals is ≥10kHz, and the sampling frequency of sound signals or acoustic emission signals is ≥44.1kHz. Step 2: Real-time preprocessing at the edge: The edge computing unit performs anti-aliasing filtering, digital noise reduction, time synchronization and operating parameter extraction on the original high-frequency data, performs real-time anomaly detection based on preset thresholds, performs feature compression on normal data, and marks the abnormal data with timestamps and retains key segments. Step 3: High-frequency transient feature extraction: For the preprocessed high-frequency signal, variational mode decomposition (VMD) and wavelet packet transform (WPT) are used to separate the transient impact components. The fault features are amplified by envelope demodulation and combined with a one-dimensional convolutional neural network (1D-CNN) to extract depth discrimination features. Step 4: Cloud-edge collaborative diagnostic analysis: The edge side uploads the compression features, anomaly markers and operating parameters to the cloud. The cloud identifies the current operating condition based on the speed / current signal, calls the corresponding sub-diagnostic model, and fuses multimodal features through the attention mechanism to output the fault type, location and severity level. Step 5: Incremental learning and life prediction: The cloud constructs a health indicator (HI) based on historical high-frequency data, updates the diagnostic model using an incremental deep learning algorithm, and predicts the remaining service life (RUL) of the traction machine by combining the failure feature evolution trend and degradation model.

2. The real-time diagnostic method for elevator traction machine operating status based on high-frequency sampling according to claim 1, characterized in that, The deployment strategy of the sensor array in step 1 is as follows: a triaxial vibration acceleration sensor is arranged in the traction machine bearing housing, a wideband current transformer is installed in the motor power supply harness, an acoustic emission sensor is set in the gearbox housing, and temperature sensors are installed in the stator windings and bearing outer rings, and connected to the traction machine encoder signal.

3. The real-time diagnostic method for elevator traction machine operating status based on high-frequency sampling according to claim 1, characterized in that, The time synchronization mentioned in step 2 is achieved through hardware triggering or high-precision timestamps, with a time error of ≤1ms; the extraction of operating parameters includes calculating the real-time speed based on encoder signals and estimating the load torque based on current signals.

4. The real-time diagnostic method for elevator traction machine operating status based on high-frequency sampling according to claim 1, characterized in that, The variational mode decomposition (VMD) described in step 3 enhances the ability to separate weak impulse signals by adaptively selecting the number of mode decompositions and the penalty factor; envelope demodulation performs Hilbert transform on the high-frequency resonance components obtained by VMD decomposition to extract fault modulation features.

5. The real-time diagnostic method for elevator traction machine operating status based on high-frequency sampling according to claim 1, characterized in that, The adaptive diagnostic process described in step 4 includes: establishing a speed-load condition dictionary and identifying the condition through a clustering algorithm; converting the vibration signal to the angular domain for order analysis to eliminate the influence of speed; and training a dedicated sub-model for different conditions or dynamically adjusting the model parameters.

6. The real-time diagnostic method for elevator traction machine operating status based on high-frequency sampling according to claim 1, characterized in that, The health indicator (HI) mentioned in step 5 is constructed based on high-frequency features, including the energy proportion of a specific frequency band, the peak factor of the envelope spectrum, and the failure probability value output by the deep learning model; the degradation model adopts a hybrid model of fusion index model and machine learning regression.

7. A real-time diagnostic system for the operating status of an elevator traction machine based on high-frequency sampling, characterized in that, include: Intelligent high-frequency data acquisition module: integrates a triaxial vibration sensor, a wideband current transformer, an acoustic emission sensor, a temperature sensor and a signal conditioning circuit, supports vibration signal sampling frequency ≥20kHz, current signal sampling frequency ≥10kHz, sound / acoustic emission signal sampling frequency ≥44.1kHz, and the sampling rate can be dynamically configured; Edge computing unit: It adopts a processor with a DSP core, runs a real-time operating system, and realizes data preprocessing, time synchronization, lightweight anomaly detection algorithm and edge feature calculation. It supports industrial bus and wireless communication. Cloud-based diagnostic platform: includes data storage server, model training server and diagnostic engine, deploys multimodal fusion model, working condition recognition model, incremental learning module and RUL prediction module, and provides API interface and visualization interface; Human-computer interaction terminal: Displays real-time diagnostic results, fault warning information, health indicator trends and remaining life prediction, and supports alarm notifications, viewing of raw signal graphs and generation of maintenance work orders.

8. The real-time diagnostic system for elevator traction machine operation status based on high-frequency sampling according to claim 7, characterized in that, The intelligent high-frequency data acquisition module adopts a modular design, with a built-in anti-aliasing filter and self-calibration circuit. It supports POE power supply or low-power battery power supply, and the cutoff frequency of the anti-aliasing filter is configurable.

9. The real-time diagnostic system for elevator traction machine operation status based on high-frequency sampling according to claim 7, characterized in that, The edge computing unit and the cloud diagnostic platform adopt an adaptive data upload strategy: under normal conditions, compressed features and operating parameters are uploaded every 5-10 minutes; when an anomaly is detected, the original data and detailed features of the anomaly fragment are uploaded in real time.

10. The real-time diagnostic system for elevator traction machine operation status based on high-frequency sampling according to claim 7, characterized in that, The cloud-based diagnostic platform supports hot model updates, continuously optimizes the diagnostic model through an incremental learning module, and periodically distributes the updated lightweight model to the edge computing unit to achieve collaborative evolution of edge and cloud models.

Images