Method and system for predicting elevator health state based on voiceprint and vibration signal

Abstract

The invention discloses a voiceprint and vibration signal-based elevator health state prediction method and system, and the method comprises the steps: collecting a sound signal and a vibration signal through a sensor group; respectively extracting voiceprint features and vibration features of voiceprint and vibration by using a double-branch convolutional neural network, obtaining a first enhanced feature and a second enhanced feature through a cross attention mechanism, and splicing to obtain a multi-modal fusion vector; and finally, inputting the multi-modal fusion vector into a long short-term memory network to output a health score sequence and the remaining service life of a future time window. Through modal complementation realized by a cross attention mechanism and a physical-data dual verification mechanism, environmental interference and real faults are effectively distinguished; clear physical constraints (such as monotonicity constraints and fatigue crack propagation rules) are fused into the deep learning model, so that the model is prevented from being over-fitted, and a prediction result is more in line with a mechanical principle.

Description

A method and system for predicting elevator health status based on acoustic signature and vibration signals Technical Field

[0001] This invention relates to the field of elevator safety monitoring technology, and in particular to a method and system for predicting the health status of elevators based on acoustic signatures and vibration signals. Background Technology

[0002] Traditional elevator maintenance mainly relies on "regular inspections," such as once every 15 days. However, the frequency of elevator use varies under different circumstances. For example, elevators in hospitals and shopping malls may run more than 600 times a day and are often fully loaded. In contrast, elevators in residential buildings or freight elevators with low occupancy rates run less than 150 times a day. Therefore, a uniform maintenance cycle may cause high-frequency elevators to malfunction due to wear and tear during the maintenance interval, creating a safety blind spot. Meanwhile, low-frequency elevators face over-maintenance, resulting in a huge waste of manpower and component resources.

[0003] To address this, some existing intelligent elevator systems incorporate IoT monitoring technology to monitor the vibration spectrum or sound texture of load-bearing components such as bearing wear, guide shoe loosening, and broken wires in steel cables. However, in practical applications, this monitoring method often employs single-modal monitoring. Modal detection suffers from numerous problems, including but not limited to: while a single vibration sensor is sensitive to mechanical impact, it struggles to distinguish between the violent bouncing of passengers in the car, collisions during cargo handling, and actual abnormal vibrations of mechanical parts; while a single sound sensor can capture friction sounds, it is highly susceptible to interference from environmental noise such as construction noise, hallway noise, and even thunderstorms. Industry statistics indicate that the false alarm rate of single-modal systems is often as high as 30%-40%. In pursuit of high-precision fault diagnosis, traditional research-grade systems often upload raw high-frequency signals (such as 16kHz sound and 1kHz vibration) directly to the cloud without filtering. A single elevator can generate 2-3GB of data traffic per day. For an elevator IoT platform deploying tens of thousands of devices, this places unbearable pressure on network bandwidth and storage costs, severely limiting large-scale commercial deployment. An architecture that relies entirely on centralized cloud processing results in a delay of 3-5 seconds or even longer between the occurrence of a fault and the issuance of a system alarm (affected by network fluctuations), which cannot meet the millisecond-level handling requirements of sudden emergency faults such as people being trapped or overshooting. On the other hand, pure edge-based solutions are limited by computing power and can often only perform simple threshold judgments, unable to run complex deep learning models for trend prediction.

[0004] In short, existing elevator health monitoring methods based on vibration spectrum or sound texture suffer from high false alarm rates, large data transmission volumes, and prediction lag. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting the health status of elevators based on acoustic and vibration signals. This method and system can deeply integrate multimodal information, have edge intelligent processing capabilities, and effectively balance data transmission costs and diagnostic accuracy for predicting the health status of elevators.

[0006] The objective of this invention is achieved through the following technical solution: Firstly, this application discloses a method for predicting the health status of an elevator based on acoustic signatures and vibration signals, comprising: collecting acoustic and vibration signals during elevator operation using a sensor array deployed on the top of the elevator car and at the traction machine base; compensating for high-frequency components in the acoustic signals using a first-order high-pass filter and performing adaptive band-pass filtering based on the elevator operation phase; removing the gravitational acceleration component from the vibration signals and filtering them using a low-pass filter; and then constructing a dynamic model based on the elevator load and speed to obtain a theoretical vibration reference value V. base Then, normalization is performed to obtain the normalized vibration signal; the voiceprint features and vibration features are extracted using a dual-branch convolutional neural network, and the first and second enhancement features are obtained through a cross-attention mechanism. The first and second enhancement features are concatenated to obtain a multimodal fusion vector; finally, the multimodal fusion vector is input into a long short-term memory network containing a physical constraint loss function to output the health score sequence and remaining lifespan for future time windows.

[0007] Furthermore, the adaptive bandpass filtering includes: constructing a transfer function as follows: A variable parameter filter is used to identify the elevator's operating status: when the elevator is detected to be in the door operator's working phase, the filter is automatically set to a center frequency f. c =1200Hz, bandwidth B=800Hz, order n=2; when the elevator is detected to be in the traction machine operation stage, the filter automatically switches to: center frequency f c =5000Hz, bandwidth B=2000Hz, order n=4; at the same time, the ambient background noise intensity is monitored in real time. When the ambient noise is below 40dB, n=2 is set. Otherwise, the algorithm automatically increases the filter order to n=4 and narrows the passband.

[0008] Furthermore, it also includes the following steps executed after S400: S410, calculating at the edge: data-driven anomaly scores obtained based on local outlier factors, and physical anomaly scores based on elevator dynamics models in parallel; and weighted summing of the data-driven anomaly scores and physical anomaly scores to obtain a comprehensive anomaly score.

[0009] Furthermore, the physical anomaly score S physics and drive anomaly score S data Perform a weighted summation to obtain the sum-driven anomaly score S. dataSpecifically, the local reachability ratio (LOF) of the current sample point p relative to its k nearest neighbors is calculated based on the Local Outlier Factor (LOF) algorithm. Then, the LOF value (local outlier value) is normalized to the interval of 0-1 using the Sigmoid function and used as the data-driven anomaly score S. data Furthermore, the calculation is based on the theoretical vibration amplitude V. theory Calculate the physical anomaly score S physics : In the formula, V actual For the measured vibration amplitude, a and b are both coefficients; the theoretical vibration amplitude V theory The calculation formula is: Where W is the real-time load of the car, v is the elevator speed, and a0, a2, a3, a4, and a5 are coefficients.

[0010] Furthermore, it also includes S420, which calculates the compression ratio CR based on the comprehensive anomaly score to obtain different compression ratios CR and perform tiered transmission: the formula for calculating the compression ratio CR is: in, This represents a high compression ratio under normal conditions. For low compression ratio under abnormal conditions, λ is a coefficient that controls the compression ratio decay rate; the transmission strategy includes: normal state, only a few tens of bytes of statistical features are uploaded; suspected abnormal state, feature vectors that have been reduced by PCA and quantized by int8 are uploaded; confirmed abnormal state, complete feature vectors and LZ4 compressed raw waveform data for 5 seconds before and after the fault are uploaded.

[0011] Furthermore, the step of obtaining the first and second enhanced features through the cross-attention mechanism and then concatenating them to obtain the multimodal fusion vector includes: based on the voiceprint feature vector Vibration feature vector Construct an attention query for vibration based on voiceprint and calculate the first enhancement feature. : Construct a vibration-attention query for the voiceprint and calculate the second enhancement feature. : In the above two equations, Let be a learnable weight matrix, d be the feature dimension, and Softmax be the activation function; then and The output multimodal fusion vector after concatenation and passing through a fully connected layer .

[0012] Furthermore, the training loss function of the Long Short-Term Memory network Includes physical constraint terms, defined as: in, In the above two equations, , , These are the trend term weights, monotonicity weights, and physical constraint weights, respectively. These hyperparameters are fine-tuned based on the model's performance on the validation set. MSE It is the mean square error term. As a trend consistency penalty item, As a monotonic constraint, considering irreversible degradation such as bearing wear and cable fatigue, the predicted health curve should monotonically decrease without maintenance intervention.

[0013] Based on Paris's fatigue crack propagation law, the rate of change of the remaining life prediction is constrained to conform to the power-law decay characteristic. Where C and m are constants related to material properties, and RUL() is the remaining useful life prediction function at time t. It is the time difference; as in the above formula, RUL(t) is the predicted remaining service life at time t; for example, C=0.002, m=3.5 are empirical parameters based on the material properties of bearing steel, and this constraint forces the degradation trajectory learned by the neural network to conform to physical facts.

[0014] Secondly, this application discloses an elevator health status prediction system based on voiceprint and vibration signals. The system is characterized by comprising: an edge terminal system, including a data acquisition module, an edge processor, and a communication module; the data acquisition module includes an omnidirectional electret condenser microphone deployed at the center of the car top and a triaxial accelerometer rigidly fixed to the traction machine base; and a cloud subsystem, including a data storage cluster, a multimodal fusion module, a health status prediction module, and a model training platform; the edge processor is configured to perform signal preprocessing, feature extraction, anomaly detection, and data compression; and the cloud subsystem is configured to perform deep feature fusion, long-cycle time-series prediction, and incremental training and distribution of the model.

[0015] Furthermore, the edge processor has a built-in lightweight neural network model, which is generated by the cloud-based multi-task learning model through knowledge distillation and 8-bit weight quantization. The edge processor also supports receiving model update packages from the cloud via the MQTT protocol.

[0016] Its effectiveness lies in the following: modal complementarity (sound and vibration mutually verifying each other) achieved through cross-attention mechanism, and physical-data dual verification mechanism, effectively distinguishing environmental interference (such as decoration noise, passenger fighting) from real faults; adaptive compression strategy reduces the average daily traffic of a single elevator from GB to less than 100MB, significantly reducing 4G traffic costs and cloud storage costs, making large-scale IoT deployment based on public networks economically feasible; 50ms-level real-time anomaly detection is achieved at the edge, ensuring that sudden faults such as entrapment can trigger local alarms and rescue processes in an instant; long-term trend prediction in the cloud can detect potential hazards in advance; and incorporating explicit physical constraints (such as monotonicity constraints and fatigue crack propagation laws) into the deep learning model not only prevents model overfitting but also makes the prediction results more consistent with mechanical principles. Attached Figure Description

[0017] Figure 1 is a flowchart illustrating the elevator health status prediction method based on acoustic signature and vibration signal according to an embodiment of this application. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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] Before explaining the embodiments of this application, some technical terms will be explained: MFCC: Mel-frequency coefficients, an acoustic feature based on the characteristics of human hearing. By introducing a Mel-scale filter bank, the linear frequency spectrum is converted into a Mel spectrum that approximates the human ear's perception.

[0020] Filter order and bandwidth: The order refers to the number of poles in the filter's transfer function. A higher order results in a steeper transition between the passband and stopband, leading to higher filtering accuracy but also greater computational complexity and more noticeable phase delay. Bandwidth refers to the range of frequencies the filter can pass through; a narrower bandwidth indicates stronger frequency selectivity.

[0021] PCA dimensionality reduction: This method projects high-dimensional data into a low-dimensional linear space (principal components) through orthogonal transformation, maximizing the variance of the first principal component and decreasing the variance of subsequent components. Its core principle is to preserve the most prominent patterns of change in the data while removing redundancy and noise. It is commonly used for visualization, feature compression, and computational acceleration.

[0022] BatchNorm and ReLU: BatchNorm standardizes the input of each layer (subtracting the mean and dividing by the standard deviation), stabilizing the data distribution, accelerating model training convergence, and mitigating gradient vanishing. ReLU is an activation function.

[0023] The following describes the embodiments of this application. Considering the strict periodicity and phased nature of elevator operation, with drastically different physical characteristics at different stages, background noise would overwhelm fault features if not processed indiscriminately. Therefore, this application proposes an elevator health status prediction method based on acoustic signature and vibration signals. Referring to Figure 1, the method includes: S100: Simultaneously collecting sound and vibration signals throughout the entire elevator cycle using sensor arrays deployed on the top of the elevator car and at the traction machine base. S200: Compensating for high-frequency components in the sound signal using a first-order high-pass filter, and performing adaptive band-pass filtering based on elevator operation stages (e.g., starting, stabilizing, braking, door opening / closing). S300: Removing the gravitational acceleration component from the vibration signal and using a low-pass filter to prevent high-frequency signal folding from interfering with low-frequency features. Then, constructing a dynamic model based on load and speed, and normalizing it to obtain a normalized vibration signal. S400: Extracting features from the sound and vibration signals from S200 and S300 to obtain acoustic signature features and vibration features, respectively. Then, the S500 algorithm is used to extract high-dimensional latent features of voiceprint and vibration using a dual-branch convolutional neural network. The correlation weights between modalities are calculated through a cross-attention mechanism to generate a multimodal fusion feature vector.

[0024] In detail, high-frequency components of sound signals attenuate rapidly during airborne propagation and microphone pickup. To highlight high-frequency fault characteristics such as bearing wear and gear meshing (typically distributed in the 3-8kHz range), this embodiment uses a first-order high-pass filter for pre-emphasis. In the formula, Z is a complex variable, and H(Z) is the filter transfer function.

[0025] In this way, by suppressing the low-frequency part of the signal that is usually strong, while relatively boosting the high-frequency part that is relatively weak, the signal-to-noise ratio of the high-frequency band is improved, so that the high-frequency fault features (such as bearing wear and gear meshing signals) in subsequent analysis can be more clearly highlighted from the background noise.

[0026] The adaptive bandpass filter utilizes a dual-threshold method based on short-time energy and zero-crossing rate, combined with the vibration signal envelope, to divide the operating cycle into seven stages, as follows: The operating cycle is precisely divided into seven stages: door opening, door closing, start-up, speed stabilization, deceleration, leveling, and standby. To simplify filter parameter control, these seven stages are categorized into three typical operating conditions: During the door operator's working condition (corresponding to the door opening and closing stages), the main noise sources are the door motor, belt, and guide rail friction, with relatively low frequencies. When the elevator is detected to be in the door operator's working stage, the filter automatically sets the center frequency f. c=1200Hz, bandwidth B=800Hz, order n=2.

[0027] During traction machine operation (corresponding to the start-up, speed stabilization, and deceleration phases), focus on high-frequency faults in the main bearings and gearbox. When the elevator is detected to be in the traction machine operation phase, the filter automatically switches to: center frequency f c =5000Hz, bandwidth B=2000Hz, order n=4.

[0028] Environmental noise adaptive (standby mode): Real-time calculation of background noise sound pressure level in non-operating states (standby, single-level). When the ambient noise is below 40dB, n=2 is set. If background noise of 40dB is detected (e.g., the sound of a construction drill nearby), the algorithm automatically increases the filter order to n=4 and narrows the passband to enhance the suppression of out-of-band non-stationary noise.

[0029] Construct the transfer function as follows The system uses a variable parameter filter to identify the elevator's operating status: when the elevator is detected to be in the door operator's working phase (opening or closing), the main noise sources are the door motor, belt, and guide rail friction, which have low frequencies. The filter is automatically set to a center frequency f. c =1200Hz, bandwidth B=800Hz, order n=2 to focus on the friction noise between the door operator guide rail and the motor. When the elevator is detected to be in the traction machine operation phase (especially the steady speed phase), pay attention to high-frequency faults in the main bearing and gearbox; the filter automatically switches to: center frequency f c =5000Hz, bandwidth B=2000Hz, order n=4, to cover the high-frequency characteristics of bearing failure and gear meshing. Simultaneously, the ambient background noise intensity is monitored in real time to calculate the background noise sound pressure level in non-operational states under different operating conditions. When the ambient noise is below 40dB, n=2 is set. If background noise ≥40dB is detected (e.g., the sound of a nearby electric drill), the algorithm automatically increases the filter order to n=4 and narrows the passband to enhance the suppression of out-of-band non-stationary noise.

[0030] In S300, for the vibration signal, the gravitational acceleration component is first removed, and the calculation formula is as follows: Where V() is the amplitude of the original vibration signal acquired by the sensor, t is the current sampling time, V' is the amplitude of the vibration signal after removing the gravitational acceleration (DC component), and N is the number of sampling points used to calculate the mean. Subsequently, a third-order Butterworth low-pass filter with a cutoff frequency of 450Hz (0.45 times the sampling rate) is used to prevent high-frequency signal folding from interfering with low-frequency characteristics.

[0031] Then, considering the significant impact of the car's real-time load W and running speed v, the vibration baseline values ​​for fully loaded upward travel and unloaded downward travel differ greatly. In this embodiment, to extract the essential characteristics reflecting the equipment's health, it is necessary to eliminate baseline fluctuations caused by operating conditions and construct a dynamic baseline model: Among them, V base The theoretical vibration baseline values ​​are k1=0.015, k2=0.8, and k3=0.5, which are initial empirical parameters. These can be calibrated later through regression analysis of the elevator's historical normal data via the cloud platform. Then, the normalized vibration signal is: V norm The normalized vibration signal, as described above, makes the characteristics robust to load changes.

[0032] In S400, the extracted acoustic signature features are 39-dimensional MFCC, including 13-dimensional static coefficients. These retain the first 13 coefficients after the DCT (Discrete Cosine Transform), including DC components, cepstral features reflecting the channel shape, and 13-dimensional first-order and second-order difference coefficients. The frame length is 25ms, with a frame shift of 10ms. Vibration features are 36-dimensional, including 12-dimensional time-domain features such as mean, standard deviation, RMS (root mean square), peak value, kurtosis (extremely sensitive to early impulsive impacts, often used for bearing pitting diagnosis), skewness, waveform factor, and impulse factor; and 24-dimensional frequency-domain features, such as frequency bands divided based on elevator fault mechanisms. Low frequencies (0-10Hz) correspond to traction sheave imbalance and eccentricity; mid-frequency frequencies (10-100Hz) correspond to poor coupling alignment; and high frequencies (100-500Hz) correspond to bearing wear and wire rope breakage. The energy E for each frequency band is calculated. L E M E H And its proportion. Furthermore, a three-level decomposition was performed using db4 wavelet packets to calculate the energy entropy of eight frequency bands. , where p i As an energy probability distribution, this index can effectively quantify the complexity and disorder of a signal, and is a sensitive indicator of early faults.

[0033] In S500, voiceprint features are represented as vectors. Vibration characteristics are represented as vectors Construct an attention query for vibration based on voiceprint, and calculate the first enhancement feature. : Calculate the first enhanced feature Used to capture vibration patterns that accompany specific sounds; construct an attention query for vibrations on the voiceprint, and calculate the second enhancement feature. : Calculate the second enhancement feature Used to capture the acoustic signature features that accompany specific vibrations; in the above two equations... Let be a learnable weight matrix, and d be the feature dimension; then and The output multimodal fusion vector after concatenation and passing through a fully connected layer This enables complementary enhancement between modes.

[0034] For example, the voiceprint feature input is a 39×150 (feature dimension × time step) MFCC matrix. After three layers of one-dimensional convolution (Conv1D) with 64, 128, and 256 filters respectively, combined with BatchNorm and ReLU activation, the final output is a voiceprint feature sequence. The vibration feature input is a 36-dimensional statistical feature vector. After passing through three fully connected (Dense) layers with 128, 256, and 256 nodes, the output vibration feature vector is determined. .

[0035] Understandably, it is possible to construct a voiceprint focusing on vibrations, that is, vibration vectors. As a query, voiceprint sequence As key and value. Because It is a global feature, so it is first broadcast in the time dimension.

[0036] In this way, based on the current vibration pattern (e.g., "high-frequency strong vibration"), the most relevant segments in the voiceprint sequence (e.g., "a sharp howl at t=3.5s") are identified and given higher weights. Similarly, vibration-focused voiceprints use voiceprint features to weight vibration features. The enhanced feature vectors are concatenated and mapped to a 512-dimensional fusion vector E through a fully connected layer. This mechanism enables the model to capture definitive faults that "have both vibration and abnormal noise," while ignoring interference from "no vibration" (external noise).

[0037] Finally, the S600 algorithm is used to input the time-series-based multimodal fusion vector into a Long Short-Term Memory (LSTM) network containing a physical constraint loss function, and outputs a health score sequence and remaining useful life (RUL) for future time windows.

[0038] Specifically, the training loss function of the Long Short-Term Memory network Includes physical constraint terms, defined as: ,in Used to constrain the monotonic decline in health caused by degenerative faults (such as wear and fatigue); , , These are the trend term weights, monotonicity weights, and physical constraint weights, respectively. These hyperparameters are fine-tuned based on the model's performance on the validation set. MSE It is the mean square error term. As a trend consistency penalty item, As a monotonic constraint, considering irreversible degradation such as bearing wear and cable fatigue, the predicted health curve should monotonically decrease without maintenance intervention.

[0039] Based on Paris's fatigue crack propagation law, the rate of change of the remaining life prediction is constrained to conform to the power-law decay characteristic. Where C and m are constants related to material properties, and RUL() is the remaining useful life prediction function at time t. It is the time difference; according to the Paris formula in fracture mechanics The crack propagation rate increases power-lawly with size, and in RUL (Remaining Life) prediction, this means that the rate of RUL decay should accelerate non-linearly over time (slow initially, fast later). For example, C=0.002, m=3.5 are set as empirical parameters based on the material properties of bearing steel. This constraint forces the neural network to learn a degradation trajectory that conforms to physical reality, rather than merely fitting the data.

[0040] For example, the fused vector sequence of the past 72 hours is input into a two-layer Long Short-Term Memory (LSTM) neural network (256+128 units). The output is a health score H for the next 24 hours. t The loss function in a two-layer long short-term memory neural network is: Furthermore, for irreversible degradation such as bearing wear and cable fatigue, the predicted health curve should show a monotonically decreasing trend without maintenance intervention. The prediction of a "rebound" in health is penalized, eliminating non-physical oscillations in the prediction curve.

[0041] In some embodiments, in order to address the uncertainty of a single model, the embodiments of this application further include the following steps executed after S400: S410, computing in parallel at the edge the data-driven anomaly score based on the local outlier factor (LOF) and the physical anomaly score based on the elevator dynamics model, and then weighting and summing the data-driven anomaly score and the physical anomaly score to obtain a comprehensive anomaly score.

[0042] Specifically, the Local Outlier Factor (LOF) is mainly based on the Local Outlier Factor algorithm, which does not require a pre-defined data distribution type. For example, calculating the local reachability ratio of the current sample point p relative to its k nearest neighbors (k=20): In the formula, Nk Let (p) represent the set of k nearest neighbors of sample point p, lrd(o) be the local reachability density of sample point o, and lrd(p) be the local reachability density of sample point p. The score obtained by normalizing the LOF (Local Outlier) value to the interval [0,1] using the Sigmoid function is the data-driven anomaly score S. data Data-driven anomaly score S data The higher the value, the more isolated the current data point is in the statistical space, meaning it is more abnormal. It excels at capturing unknown and complex abnormal patterns.

[0043] Simultaneously, physical anomaly scores were obtained, and a theoretical vibration prediction model was constructed based on the elevator dynamics equations, yielding: In the formula, V theory The theoretical vibration amplitude is given by denoted as W, the real-time load of the car is given by denoted as v, and the elevator speed is given by denoted as v. The coefficients a0 to a5 are obtained by fitting historical data of the elevator under healthy conditions using the least squares method.

[0044] The coefficients [a0...a5] are fitted using least-squares data during the initial self-learning period of system deployment. The self-learning period refers to the parameters that the elevator needs to determine through "self-learning" during the initial system deployment phase to eliminate individual equipment differences. For example, the system is set to a "self-learning period" of 7 days (assuming the elevator is in a healthy state during this time), during which the average load W of each run is automatically collected. i Average velocity v i and vibration amplitude V i Construct a sample set {W} containing different operating conditions (no load / full load, uplink / downlink). i ,v i V i Subsequently, the least squares method was used to perform multiple linear regression fitting on the sample set, and a set of optimal coefficients a0 to a5 were obtained.

[0045] The physical anomaly score is calculated as a relative error: Among them, V actual For the measured vibration amplitude, 0.3 and 1.0 here are set so that a deviation of 30% from the theoretical value by the actual value is considered a full score anomaly (1.0). This physical model is good at identifying faults that violate mechanical common sense and is not sensitive to environmental noise.

[0046] Then, adaptive weight fusion is performed to calculate the comprehensive anomaly score S. anomaly : For example, the initial settings are α=0.6, β=0.4. The system maintains a historical false alarm counter.

[0047] In detail, the above equation satisfies α+β=1; establish a historical false alarm feedback mechanism based on a sliding window, if S data In recent If a preset number of false alarms occur during the detection, then decrease α and increase β, with a step size of Δw; if S physics If a false alarm occurs for a predetermined number of consecutive times, then α is increased and β is decreased; where S physics The calculation is based on the theoretical vibration amplitude V theory .

[0048] In other words, if S data If an anomaly is indicated but confirmed as a false positive by manual analysis or subsequent cloud-based in-depth analysis, it means the data model is overly sensitive to certain rare but normal operating conditions (such as continuous off-center load during relocation). The system automatically reduces α (in steps of 0.05) and increases β. If S physics If an anomaly is detected but confirmed as a false alarm, it indicates that the physical model parameters have drifted (e.g., guide shoe wear causing a natural rise in the normal vibration reference). The system automatically decreases β, increases α, and triggers a recalibration process for the physical parameters [a0...a5]. Constraints: α∈[0.4,0.8], β∈[0.2,0.6], ensuring that a single model does not completely dominate the decision.

[0049] Furthermore, in order to balance data value and bandwidth cost, the method also includes S420, calculating the compression ratio CR based on the comprehensive anomaly score to obtain different compression ratios CR and perform graded transmission.

[0050] The formula for calculating the compression ratio (CR) is: in, This represents a high compression ratio under normal conditions (only statistical summaries are uploaded). For low compression ratio (lossless compression) under abnormal conditions, λ is a coefficient that controls the compression ratio decay rate; the transmission strategy is divided into three levels: in normal state, only a few tens of bytes of statistical features are uploaded; in suspected abnormal state, feature vectors that have been reduced by PCA and quantized by int8 are uploaded; in confirmed abnormal state, the complete feature vector and LZ4 compressed raw waveform data for 5 seconds before and after the fault are uploaded.

[0051] For example, the compression ratio CR is calculated as follows: We can obtain: normal state ( ):CR ≈ 0.95. At this time, the data value is low, and only the statistical summaries of the features (such as mean, extreme values) are uploaded. The data volume for a single run is only about 50 bytes, and it is only necessary to prove that the device is online and healthy. Attention state / blue warning (0.3 < S ≤ 0.5): CR ≈ 0.70. At this time, there are suspected anomalies, and features need to be uploaded for cloud analysis. PCA dimensionality reduction (retaining 95% variance) is performed on the MFCC and vibration features, and the floating-point numbers are quantized into int8 format for uploading. Abnormal state / yellow-red warning (S > 0.5): CR ≈ 0.30. At this time, the data value is extremely high and must be completely retained. The system uploads the uncompressed complete feature vector and simultaneously intercepts the original waveform data (about 80KB) 5 seconds before and after the abnormal moment, and uploads it using LZ4 lossless compression. These original waveforms are the materials for cloud expert diagnosis and model iterative training.

[0052] The elevator health status prediction system based on voiceprint and vibration signals according to the embodiments of the present application is used to implement the elevator health status prediction method based on voiceprint and vibration signals in the above embodiments, and includes: an edge terminal system, including a data acquisition module, an edge processor, and a communication module; the data acquisition module includes an omnidirectional electret condenser microphone deployed at the center position of the car top and a three-axis MEMS acceleration sensor rigidly fixed to the traction machine base. Exemplarily, the sound sensor selects an industrial-grade omnidirectional electret condenser microphone with a sensitivity index of -38dB ± 3dB, a frequency response range as wide as 20Hz - 20kHz, and a signal-to-noise ratio > 65dB. The installation position of the sensor is optimized through sound field simulation, and it is fixed to the geometric center of the outside of the car top through a strong magnetic suction base, and is 20cm vertically away from the surface of the car roof. This distance setting constitutes an "acoustic buffer zone", which can effectively pick up the reverberation in the hoistway (reflecting the flatness of the guide rail) and the airborne sound conducted by the traction machine, and can also attenuate the low-frequency structure noise generated by the direct vibration of the car roof. The sampling rate is set to 16kHz to cover the audible range of the human ear and the high-frequency whistling of early bearing failures. The vibration sensor selects a three-axis MEMS accelerometer (ADXL series) with a measuring range of ±16g, a resolution of 0.001g, and a non-linearity < 0.5%. The sensor is rigidly bonded to the vicinity of the bearing seat of the traction machine base through high-strength epoxy resin glue, which is the most direct point for the vibration transmission of the mechanical transmission chain. To avoid aliasing, the cut-off frequency of the hardware low-pass filter is set to 500Hz, and the sampling rate is set to 1kHz. The edge computing gateway, as the core of data aggregation and processing, is equipped with an ARM Cortex-A53 quad-core processor with a main frequency of 1.5GHz, 2GB of LPDDR4 memory, and 16GB of eMMC storage. The gateway integrates a hardware watchdog circuit to prevent the program from running away. The power input end is designed with surge protection and backup capacitors to ensure that the system can still work for at least 10 seconds when the elevator suddenly loses power due to a fault, so as to complete the preservation and upload of the fault data at the last moment.

[0053] The cloud subsystem includes a data storage cluster, a multimodal fusion module, a health status prediction module, and a model training platform. The edge processor is configured to perform signal preprocessing, feature extraction, anomaly detection, and data compression. The cloud subsystem is configured to perform deep feature fusion, long-term time-series prediction, and incremental training and deployment of the model. For example, the cloud subsystem adopts a microservice architecture and is deployed on a Kubernetes cluster. It includes an MQTT message broker (EMQX), a time-series database (InfluxDB), object storage (MinIO), a relational database (PostgreSQL), and an AI inference service (TorchServe).

[0054] In some embodiments, the edge processor has a built-in lightweight neural network model, which is generated by a cloud-based multi-task learning model through knowledge distillation and 8-bit weight quantization. The edge processor also supports receiving model update packages from the cloud via the MQTT protocol and has a mechanism for performing grayscale verification before updating the model.

[0055] Furthermore, incremental training and model updates are performed monthly in the cloud.

[0056] The data loop is implemented by cleaning and expanding the dataset using work order feedback (truth labels) from maintenance engineers. Samples marked as "false alarms" are added to a negative sample set for focused training. Sample balancing is achieved by oversampling scarce fault samples (such as broken wire ropes) using the SMOTE algorithm to balance the distribution of the 12 fault sample classes. Model compression follows, with weight quantization (32-bit -> 8-bit) and knowledge distillation to generate a lightweight .tflite model. A canary deployment (ShadowMode) is implemented, distributing the model to edge devices via MQTT. Upon receiving the new model, the edge devices do not immediately replace the old model but instead activate "shadow mode": the new model runs in the background, its output does not trigger alarms, and is only uploaded to the cloud for comparison with the old model. Once the cloud verifies that the new model has no abnormal crashes for seven consecutive days and its accuracy is superior to the old model, a command is sent to officially switch over, thus eliminating the risk of widespread device failure due to model updates.

[0057] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting the health status of elevators based on acoustic signatures and vibration signals, characterized in that, include: The system collects sound and vibration signals during elevator operation by deploying sensor arrays on the top of the elevator car and at the traction machine base. For the sound signal, a first-order high-pass filter is used to compensate for high-frequency components, and adaptive band-pass filtering is performed based on the elevator's operating phase. For the vibration signal, the gravitational acceleration component is removed and filtered by a low-pass filter. Then, a dynamic model is constructed based on the elevator's load and speed to obtain the theoretical vibration reference value V. base Then, normalization is performed to obtain the normalized vibration signal; the voiceprint features and vibration features are extracted using a dual-branch convolutional neural network, and the first and second enhancement features are obtained through a cross-attention mechanism. The first and second enhancement features are concatenated to obtain a multimodal fusion vector; finally, the multimodal fusion vector is input into a long short-term memory network containing a physical constraint loss function to output the health score sequence and remaining lifespan for future time windows.

2. The elevator health status prediction method based on voiceprint and vibration signals according to claim 1, characterized in that, The adaptive bandpass filter includes identifying the elevator's operating status: when the elevator is detected to be in the door operator's working phase, the filter is automatically set to a center frequency f. c =1200Hz, bandwidth B=800Hz, order n=2; when the elevator is detected to be in the traction machine operation stage, the filter automatically switches to: center frequency f c =5000Hz, bandwidth B=2000Hz, order n=4; at the same time, the ambient background noise intensity is monitored in real time. When the ambient noise is below 40dB, n=2 is set. Otherwise, the algorithm automatically increases the filter order to n=4 and narrows the passband.

3. The elevator health status prediction method based on acoustic signature and vibration signal according to claim 1 or 2, characterized in that, It also includes S410, which is executed after S400: data-driven anomaly scores obtained based on local outlier factors and physical anomaly scores based on elevator dynamics models, computed at the edge. The comprehensive anomaly score is obtained by weighted summation of the data-driven anomaly score and the physical anomaly score.

4. The elevator health status prediction method based on acoustic signature and vibration signal according to claim 3, characterized in that, Score for physical anomalies physics and drive anomaly score S data Perform a weighted summation to obtain the sum-driven anomaly score S. data Specifically, the local reachability ratio of the current sample point p relative to its k nearest neighbors is calculated based on the local outlier factor algorithm. Then, the result of normalizing the local outlier values ​​to the interval of 0 to 1 is used as the data-driven anomaly score S. data Furthermore, the calculation is based on the theoretical vibration amplitude V. theory Calculate the physical anomaly score S physics : In the formula, V actual For the measured vibration amplitude, a and b are the deviation tolerance and saturation threshold, respectively; the theoretical vibration amplitude V theory The calculation formula is: Where W is the real-time load of the car, v is the elevator speed, and a0, a2, a3, a4, and a5 are coefficients.

5. The elevator health status prediction method based on acoustic signature and vibration signal according to claim 4, characterized in that, It also includes S420, which calculates the compression ratio CR based on the comprehensive anomaly score to obtain different compression ratios CR and perform hierarchical transmission: the transmission strategy includes: normal state, only a few tens of bytes of statistical features are uploaded; suspected anomaly state, the feature vector after dimensionality reduction and quantization is uploaded; confirmed anomaly state, the complete feature vector and compressed raw waveform data for 5 seconds before and after the fault are uploaded.

6. The elevator health status prediction method based on acoustic signature and vibration signal according to claim 1, characterized in that, The process of obtaining the first and second enhanced features through the cross-attention mechanism and then concatenating them to obtain the multimodal fusion vector includes: based on the voiceprint feature vector. Vibration characteristic vector Construct an attention query for vibration based on voiceprint and calculate the first enhancement feature. ： Construct a vibration-attention query for the voiceprint and compute the second enhancement feature. ： In the above two equations, Let be a learnable weight matrix, d be the feature dimension, and Softmax be the activation function; then and The output multimodal fusion vector after concatenation and passing through a fully connected layer 。 7. The elevator health status prediction method based on acoustic signature and vibration signal according to claim 1, characterized in that, The training loss function of the Long Short-Term Memory network Includes physical constraint terms, defined as: in, In the above two equations, 、 、 These are the trend term weights, monotonicity weights, and physical constraint weights, respectively. MSE It is the mean square error term. As a trend consistency penalty item, For monotonicity constraint terms, Based on the fatigue crack propagation law, the rate of change of the remaining life prediction is constrained to conform to the power-law decay characteristic. ,in and Let RUL() be a constant, and let RUL() be the remaining useful life prediction function at time t. It is the time difference.

8. An elevator health status prediction system based on voiceprint and vibration signals, characterized in that, The method for predicting elevator health status based on acoustic signature and vibration signals as described in any one of claims 1-7 comprises: an edge terminal system, including a data acquisition module, an edge processor, and a communication module; the data acquisition module includes an omnidirectional electret condenser microphone deployed at the center of the top of the car and a triaxial accelerometer rigidly fixed to the traction machine base; and a cloud subsystem, including a data storage cluster, a multimodal fusion module, a health status prediction module, and a model training platform; the edge processor is configured to perform signal preprocessing, feature extraction, anomaly detection, and data compression; and the cloud subsystem is configured to perform deep feature fusion, long-cycle time-series prediction, and incremental training and distribution of the model.

9. The elevator health status prediction system based on voiceprint and vibration signals according to claim 8, characterized in that, The edge processor has a built-in lightweight neural network model, which is generated by a cloud-based multi-task learning model through knowledge distillation and 8-bit weight quantization. The edge processor also receives model update packages from the cloud via a protocol.

Images