Millimeter wave radar-based elevator hoisting machine vibration reconstruction method and system

Abstract

The present application relates to the elevator technical field, especially a kind of elevator hoisting machine vibration reconstruction method and system based on millimeter wave radar.The method includes: in the process of elevator operation, the echo signal of millimeter wave radar is collected;First, continuous phase sequence is obtained by echo signal;Then, according to the phase-displacement mapping relationship, the rough vibration signal of radar line-of-sight direction is generated, then the physical confidence is calculated and the double-channel physical priori tensor is constructed, and the three-axis reconstructed vibration signal tensor is synthesized by PGGN physical gate generation network, finally, the three-axis reconstructed vibration signal tensor is evaluated for physical rationality, and the gate parameter of anisotropic physical gate unit is adjusted according to the evaluation result.The present application realizes the physical consistency reconstruction of three-axis vibration signal of elevator hoisting machine without long-term installation of contact sensor on the surface of hoisting machine.

Description

Technical Field

[0001] This invention relates to the field of elevator technology, and in particular to a vibration reconstruction method and system for elevator traction machines based on millimeter-wave radar. Background Technology

[0002] As elevator equipment develops towards higher reliability, intelligence, and less maintenance, higher demands are placed on online monitoring of elevator traction machine operation and vibration information acquisition. Current elevator traction machine vibration monitoring mainly relies on contact-type vibration sensors, which are complex to install, costly to maintain, and susceptible to sensor detachment, aging, and environmental interference during long-term operation. Therefore, they cannot meet the engineering requirements for long-term, stable, and non-intrusive monitoring of elevators in use.

[0003] Millimeter-wave radar, as a non-contact sensing method, has advantages such as being non-attached and adaptable to complex environments, and theoretically can sense minute vibrations of targets through phase changes. However, in the complex industrial scenario of elevator traction machines, millimeter-wave radar vibration measurement still faces several key technical challenges: On the one hand, due to multipath effects, echo fading, and the influence of environmental noise, radar phase vibration measurement results have significant non-stationary noise and waveform distortion, making it difficult to directly obtain physically consistent vibration signals; on the other hand, limited by the physical measurement mechanism of millimeter-wave radar, a single radar can usually only acquire one-dimensional vibration information along the radar's line of sight, and cannot directly sense the vibration components of the traction machine in other spatial directions, resulting in a lack of vibration information dimensions, making it difficult to directly use for fault analysis.

[0004] Furthermore, existing deep learning-based vibration reconstruction methods mostly adopt an end-to-end black-box modeling approach, directly mapping vibration signals from radar signals. For example, the transformer vibration anomaly detection method based on millimeter-wave radar disclosed in CN119619658A lacks explicit modeling of radar physical measurement mechanisms and signal quality differences. This can easily lead to "signal illusions" that do not conform to physical laws when the signal quality is poor, thereby reducing the credibility of monitoring results in elevator safety assessment and fault diagnosis.

[0005] Therefore, how to fully utilize the advantages of millimeter-wave radar for non-contact measurement while achieving physical consistency reconstruction of the triaxial vibration signal of the elevator traction machine under complex working conditions has become a key technical problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to propose a vibration reconstruction method and system for elevator traction machines based on millimeter-wave radar. This method achieves physical consistency reconstruction of the three-axis vibration signals of the elevator traction machine without the need for long-term installation of contact sensors on the surface of the traction machine, providing a reliable data foundation for evaluating the operating status of the elevator traction machine, detecting anomalies, and providing safety warnings.

[0007] To achieve this objective, the present invention adopts the following technical solution:

[0008] A vibration reconstruction method for elevator traction machines based on millimeter-wave radar includes the following steps:

[0009] S1: Aim the main beam of the millimeter-wave radar at the vibrating part of the elevator traction machine and sample the echo signal during elevator operation.

[0010] S2: Determine the target range unit through the echo signal and extract the corresponding phase information to obtain a continuous phase sequence;

[0011] S3: The continuous phase sequence is drift compensated using a dynamic phase reference, and a rough vibration signal in the radar line-of-sight direction is generated based on the phase-displacement mapping relationship;

[0012] S4: Calculate the physical confidence level based on the statistical characteristics of the echo amplitude of the target distance unit, and construct a dual-channel physical prior tensor after aligning the rough vibration signal and the physical confidence level in the time dimension;

[0013] S5: The dual-channel physical prior tensor is input into the PGGN physical gated generation network, which is composed of a generator and anisotropic physical gated units cascaded together. The generator outputs the original triaxial residual features, which are then gated and modulated by the anisotropic physical gated units to generate gated triaxial residual features. The gated triaxial residual features are then superimposed with the rough vibration signal to synthesize a triaxial reconstructed vibration signal tensor.

[0014] S6: Perform a physical rationality assessment on the triaxial reconstructed vibration signal tensor, and adjust the gating parameters of the anisotropic physical gating unit based on the assessment results.

[0015] Furthermore, step S2 includes:

[0016] S21: Using a single Chirp as the processing unit, a window function is applied in the fast time dimension to suppress spectral sidelobe leakage. Then, the time-domain discrete signal is converted into complex echo data arranged in distance units through a distance-to-fast Fourier transform.

[0017] S22: Arrange the complex echo data of the same distance cell under multiple Chirps along the slow time dimension to form a complex echo sequence. Extract the zero frequency component of the complex echo sequence to obtain the corresponding complex value. Calculate the amplitude square of the complex value to determine the energy of each distance cell. Select the distance cell with the largest energy as the target distance cell within a preset distance window.

[0018] S23: Extract the complex echo sequence of the target distance cell, calculate the instantaneous phase to obtain the wrapped phase sequence; calculate the wrapped phase difference between the current sampling time and the previous sampling time in the slow time dimension. When the wrapped phase difference is greater than +π, subtract 2π from the wrapped phase at the current sampling time; when the wrapped phase difference is less than -π, add 2π to the wrapped phase at the current sampling time; otherwise, keep the wrapped phase at the current sampling time unchanged, complete the phase unwrapping, and obtain a continuous phase sequence.

[0019] Furthermore, step S3 includes:

[0020] S31: The dynamic phase reference of continuous phase is calculated by the time averaging method within a sliding time window to achieve drift compensation;

[0021] The formula for the dynamic phase reference is as follows:

[0022] ,

[0023] in, Indicates the length of the sliding time window. Indicates the real-time sampling time. Indicates the start time of the sliding time window. Represents the integral variable. Indicates a historical moment within the sliding time window. Continuous phase after unwinding;

[0024] S32: Based on the phase-displacement mapping relationship, the change in continuous phase relative to the dynamic phase reference is converted into a rough vibration signal;

[0025] The formula for the phase-displacement mapping relationship is as follows:

[0026] ,

[0027] in, Represents the moment The output rough vibration displacement, For millimeter wave wavelength, Indicates the real-time sampling time. For a moment Continuous phase after unwinding It serves as a dynamic phase reference.

[0028] Furthermore, step S4 includes:

[0029] S41: Calculate the physical confidence level based on the statistical characteristics of the echo amplitude of the target distance cell. The formula is as follows:

[0030] ;

[0031] in, For a moment The echo amplitude corresponding to the target distance cell, ; and They are time points The in-phase and quadrature components of the complex echo data corresponding to the target range cell; and These are the mean and standard deviation of the echo amplitude within the slow-time sliding window, respectively. The Sigmoid normalization function maps the physical confidence level to the (0,1) interval;

[0032] S42: For rough vibration signals within the sampling time period With physical confidence Discrete sampling is performed on each sample to obtain a rough vibration signal sequence. and physical confidence sequence Align and stack them along the time dimension to construct a two-channel physical prior tensor. .

[0033] Furthermore, step S5 includes:

[0034] S51: The PGGN physical gating generation network is pre-trained under the supervision of triaxial real vibration signals. During the training process, anisotropic physical gating units participate in the triaxial residual feature modulation, so that the physical confidence level constrains the triaxial residual feature generation process.

[0035] S52: Transfer the dual-channel physical prior tensor The trained PGGN physical gating generator network is input, and the generator outputs the original triaxial residual features. The anisotropic physical gating unit modulates the original triaxial residual features through physical confidence to generate gated triaxial residual features.

[0036] The coarse vibration signal sequence is dimensionally expanded to construct a triaxial reference tensor. The gated triaxial residual features and the triaxial reference tensor are superimposed to synthesize a triaxial reconstructed vibration signal tensor.

[0037] Furthermore, the generator includes an encoder, a bottleneck layer, a decoder, and an output projection layer;

[0038] The encoder includes... A series of cascaded downsampled coding blocks, the input of the first downsampled coding block receiving a dual-channel physical prior tensor. The encoder extracts multi-scale temporal features, compressed temporal resolution, and expanded feature channel dimensions of the input signal layer by layer through multi-level downsampling coding blocks;

[0039] The bottleneck layer contains several dilated residual blocks stacked in series according to the signal flow direction. The bottleneck layer captures global context information and long temporal dependency features through the dilated residual blocks to generate deep semantic features.

[0040] The decoder contains components symmetrical to the encoder. The decoder uses a multi-level upsampling decoding block and the corresponding level of the encoder to concatenate features, gradually recovering the signal from deep semantic features to the original time resolution and reconstructing high-dimensional vibration waveform features.

[0041] The output projection layer maps the decoder's output from the abstract feature space to the physical signal space, generating the original triaxial residual features. .

[0042] Furthermore, the method for generating gated triaxial residual features by the anisotropic physical gating unit includes:

[0043] The original triaxial residual features Decoupled according to channel dimension: radar line-of-sight direction component Longitudinal component and vertical components ;

[0044] Constructing time-varying suppression coefficients using physical confidence levels : ,in, As a physical sensitivity factor, the value range is set to [value range]. ;

[0045] Define the time-varying suppression coefficient The time-varying suppression coefficient sequence within the sampling period is as follows: Radar line-of-sight direction residual characteristics after physical gating constraints for: Where ⊙ represents the Hadamarda complex;

[0046] Longitudinal component and vertical components Data-driven mapping is used as the longitudinal residual feature. and vertical residual characteristics : , ;

[0047] Will , and Perform channel splicing to generate gated triaxial residual features. .

[0048] Furthermore, the method for synthesizing a triaxial reconstructed vibration signal tensor includes:

[0049] For rough vibration signal sequences Expand the dimensions to construct a three-axis reference tensor , ,in, Represents the set of real numbers. This represents the number of sampling points within the sampling time period.

[0050] Gated three-axis residual characteristics As a correction term, it is superimposed onto the triaxial reference tensor. Generate a triaxial reconstructed vibration signal tensor , , , This is the global residual scaling factor.

[0051] Furthermore, the method for training the PGGN physical gating generative network includes:

[0052] Constructing a composite objective loss function : ,in, For temporal consistency loss The weighting coefficients, Frequency domain envelope feature loss The weighting coefficients, To combat generation loss Weighting coefficients;

[0053] ,in, It is a triaxial reconstructed vibration signal tensor. It is a triaxial true vibration signal tensor. This represents the Frobenius norm square operation;

[0054] ,in, The envelope spectrum tensor is obtained by extracting the envelope spectrum of the triaxial reconstructed vibration signal through Hilbert transform and frequency domain transform. The envelope spectrum tensor is obtained by extracting the envelope spectrum of a real triaxial vibration signal through Hilbert transform and frequency domain transform. This represents the element-wise modulo operation. Represents logarithmic transformation, It is the minimum constant. This represents the Frobenius norm square operation;

[0055] ,in, It is a two-channel physical prior tensor. Represents a generator network. This represents the discriminator network. This represents the discriminator's judgment result on the generated samples, where 1 represents the target label of the real sample. This represents the Frobenius norm square operation;

[0056] A contact-type triaxial vibration sensor is installed in the elevator traction machine to collect real triaxial vibration signals. Using the real triaxial vibration signals as the supervised truth, an adversarial training strategy is adopted to alternately optimize the generator and discriminator. During the training process, the PGGN physical gated generation network is optimized with a composite objective loss function so that the reconstructed triaxial vibration signal approaches the real triaxial vibration signal in both time and frequency domain characteristics.

[0057] A vibration reconstruction system for an elevator traction machine based on millimeter-wave radar, the system being used to implement the above-mentioned method;

[0058] The system includes:

[0059] The data acquisition module includes a millimeter-wave radar, used to collect echo signals during elevator operation;

[0060] The signal preprocessing module is used to determine the target range unit through the echo signal and extract the corresponding phase information to obtain a continuous phase sequence;

[0061] The rough vibration estimation module uses a dynamic phase reference to compensate for the drift of the continuous phase sequence and generates a rough vibration signal in the radar line-of-sight direction based on the phase-displacement mapping relationship.

[0062] The prior vibration estimation module calculates the physical confidence level based on the statistical characteristics of the echo amplitude of the target distance unit, and constructs a dual-channel physical prior tensor after aligning the rough vibration signal and the physical confidence level in the time dimension.

[0063] The signal reconstruction module is equipped with a PGGN physical gated generation network consisting of a generator and anisotropic physical gated units cascaded together. After the dual-channel physical prior tensor is input into the PGGN physical gated generation network, the generator outputs the original triaxial residual features. After being gated and modulated by the anisotropic physical gated units, gated triaxial residual features are generated. The gated triaxial residual features are superimposed with the rough vibration signal to synthesize a triaxial reconstructed vibration signal tensor.

[0064] The evaluation module is used to evaluate the physical rationality of the triaxial reconstructed vibration signal and adjust the gating parameters of the anisotropic physical gating unit based on the evaluation results.

[0065] The technical solution provided by this invention may include the following beneficial effects:

[0066] 1. A physical prior vibration estimation mechanism based on the radar phase interferometry principle is introduced to extract coarse vibration signals with clear physical meaning from millimeter-wave radar echoes at the signal processing front end, providing a stable and reliable physical basis for the subsequent reconstruction process;

[0067] 2. Construct a vibration reconstruction model that integrates physical confidence (PGGN physical gated generation network). By jointly modeling the radar echo intensity and phase estimation reliability, physical measurement results are given priority when the vibration signal quality is high, and data-driven correction is introduced when the signal quality deteriorates, thereby avoiding the deep model from producing reconstruction results that do not conform to physical laws.

[0068] 3. To achieve the completion and reconstruction of three-axis vibration information under the condition of a single millimeter-wave radar, by utilizing the inherent coupling relationship between the vibration modes of the elevator traction machine, while preserving the physical authenticity of the radar line of sight, the vibration components in other spatial directions are completed to obtain a more complete vibration characterization.

[0069] 4. Enhance the diagnostic value of reconstructed signals by using frequency domain feature constraints. Introduce frequency domain constraints targeting the modulation characteristics of rotating machinery during model training so that the reconstructed triaxial vibration signals are not only realistic in the time domain, but also accurately reflect the operating status and potential fault characteristics of the traction machine in the frequency domain.

[0070] Through the above-mentioned technical means, the present invention can stably output elevator traction machine vibration signals with physical consistency and diagnostic significance in complex industrial environments, providing a reliable data foundation for subsequent elevator traction machine operation status assessment, anomaly detection and safety early warning, and has good engineering practical value and promotion prospects. Attached Figure Description

[0071] Figure 1 This is a schematic diagram showing the installation of millimeter-wave radar and triaxial vibration sensors in an elevator traction machine;

[0072] Figure 2 This is a flowchart illustrating the vibration reconstruction method for elevator traction machines based on millimeter-wave radar according to the present invention.

[0073] Figure 3 This is a schematic diagram of step S2 determining the target distance cell;

[0074] Figure 4 This is a flowchart illustrating the process of obtaining a continuous phase sequence in step S2.

[0075] Figure 5 This is a schematic diagram of calculating the dynamic phase reference within a sliding time window;

[0076] Figure 6This is a schematic diagram of the process for constructing a dual-channel physical prior tensor and synthesizing a triaxial reconstructed vibration signal tensor;

[0077] Figure 7 This is a schematic diagram of the training process for the PGGN (Physical Gated Generative Network).

[0078] Figure 8 This is a time-domain comparison diagram of the reconstructed triaxial vibration signal and the actual triaxial vibration signal.

[0079] Figure 9 This is a comparison diagram of the frequency domain between the reconstructed triaxial vibration signal and the actual triaxial vibration signal.

[0080] Among them: elevator traction machine 1, traction sheave 11, millimeter-wave radar 2, and triaxial vibration sensor 3. Detailed Implementation

[0081] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the present invention.

[0082] Reference Figure 1 and Figure 2 An elevator traction machine vibration reconstruction method based on millimeter-wave radar according to an embodiment of the present invention includes the following steps:

[0083] S1: Aim the main beam of the millimeter-wave radar at the vibrating part of the elevator traction machine and sample the echo signal during elevator operation.

[0084] S2: Determine the target range unit through the echo signal and extract the corresponding phase information to obtain a continuous phase sequence;

[0085] S3: The continuous phase sequence is drift compensated using a dynamic phase reference, and a rough vibration signal in the radar line-of-sight direction is generated based on the phase-displacement mapping relationship;

[0086] S4: Calculate the physical confidence level based on the statistical characteristics of the echo amplitude of the target distance unit, and construct a dual-channel physical prior tensor after aligning the rough vibration signal and the physical confidence level in the time dimension;

[0087] S5: The dual-channel physical prior tensor is input into the PGGN physical gated generation network, which is composed of a generator and anisotropic physical gated units cascaded together. The generator outputs the original triaxial residual features, which are then gated and modulated by the anisotropic physical gated units to generate gated triaxial residual features. The gated triaxial residual features are then superimposed with the rough vibration signal to synthesize a triaxial reconstructed vibration signal tensor.

[0088] S6: Perform a physical rationality assessment on the triaxial reconstructed vibration signal tensor, and adjust the gating parameters of the anisotropic physical gating unit based on the assessment results.

[0089] This invention introduces a physical prior vibration estimation mechanism based on the principle of radar phase interferometry. At the signal processing front end, it extracts coarse vibration signals with clear physical meaning from millimeter-wave radar echoes, providing a stable and reliable physical basis for subsequent reconstruction. A vibration reconstruction model incorporating physical confidence is constructed. Physical confidence is used to quantify the reliability of the current observation value. Physical measurement results are prioritized when the vibration signal quality is high, while data-driven correction is introduced when the signal quality deteriorates, thus avoiding reconstruction results from depth models that do not conform to physical laws. This invention achieves triaxial vibration information completion and reconstruction under single millimeter-wave radar conditions. Utilizing the inherent coupling relationship between the vibration modes of an elevator traction machine, it completes the vibration components in other spatial directions while preserving the physical authenticity of the radar line-of-sight direction, obtaining a more complete vibration characterization.

[0090] As an example, a millimeter-wave radar is deployed in the elevator machine room, with the radar's main beam aimed at the critical vibration points of the elevator traction machine (see reference). Figure 1 For example, the upper part of the rigid housing near the mounting shaft of the traction sheave on the base of an elevator traction machine. During elevator operation, millimeter-wave radar continuously and non-contactly detects key vibration areas, acquiring corresponding radar echo signals. These radar echo signals include amplitude components reflecting target distance information and phase components reflecting minute displacement changes. In practical applications, the start time and sampling duration of millimeter-wave radar signal acquisition can be determined based on the elevator traction machine's service life or historical fault records.

[0091] The continuous phase sequence obtained in step S2 of this invention can accurately reflect the minute displacement changes of the target under test in the radar line-of-sight direction, providing basic input data for subsequent vibration displacement reconstruction and vibration characteristic analysis. Specifically, refer to... Figure 3 and Figure 4 Step S2 includes:

[0092] S21: Using a single Chirp as the processing unit, a window function is applied in the fast time dimension to suppress spectral sidelobe leakage. Then, the time-domain discrete signal is converted into complex echo data arranged in distance units through a distance-to-fast Fourier transform.

[0093] S22: Arrange the complex echo data of the same distance cell under multiple Chirps along the slow time dimension to form a complex echo sequence. Extract the zero frequency component of the complex echo sequence to obtain the corresponding complex value. Calculate the amplitude square of the complex value to determine the energy of each distance cell. Select the distance cell with the largest energy as the target distance cell within a preset distance window.

[0094] S23: Extract the complex echo sequence of the target distance cell, calculate the instantaneous phase to obtain the wrapped phase sequence; calculate the wrapped phase difference between the current sampling time and the previous sampling time in the slow time dimension. When the wrapped phase difference is greater than +π, subtract 2π from the wrapped phase at the current sampling time; when the wrapped phase difference is less than -π, add 2π to the wrapped phase at the current sampling time; otherwise, keep the wrapped phase at the current sampling time unchanged, complete the phase unwrapping, and obtain a continuous phase sequence.

[0095] As an example, the analog-to-digital converter (ADC) module built into the millimeter-wave radar samples and quantizes the analog vibration signal generated by the traction machine, resulting in a digital data sequence that is the raw ADC sampled data. First, the sampled data is preprocessed to account for the temporal structure characteristics of the frequency-modulated continuous wave (FMCW) radar signal. In step S21, the raw ADC sampled data of the millimeter-wave radar is organized by frame, with each frame containing multiple linear frequency-modulated pulses (Chirps). Each Chirp contains a fixed number of discrete sampling points in the fast time dimension. Using a single Chirp as a processing unit, a window function is applied to its discrete sampling sequence in the fast time dimension. The window function is preferably any one of the Hanning window, Hamming window, or Blackman window to suppress the impact of spectral sidelobe leakage on the range resolution results.

[0096] After window function processing, a range-to-fast Fourier transform (Range-FFT) is performed on the sampling sequence of each chirp along the fast time dimension, thereby converting the time-domain discrete signal into complex echo data in the range dimension (arranged by range cells). In the range-dimensional complex echo data, different spectral cells correspond to different target range intervals, allowing signals from different reflectors in the radar echo to be distinguished in the range domain.

[0097] After obtaining the complex echo data in the range dimension, it is necessary to determine the target range cell corresponding to the measured vibration target from multiple range cells. Step S22 processes and analyzes the complex echo sequences corresponding to each range cell in the slow time dimension to distinguish between stationary or nearly stationary targets and targets with obvious radial motion characteristics. As an example, by performing Doppler spectrum analysis on the complex echo sequences in the slow time dimension, echo components located in the zero or near-zero Doppler frequency range are extracted, and the echo energy of the corresponding range cell is obtained by accumulating the squared amplitude of the echo components in the slow time dimension; or equivalently, slow-time low-pass filtering or time averaging is used to extract the zero or near-zero Doppler components in the slow-time complex echo sequences, and the corresponding echo energy is calculated by accumulating or averaging the squared amplitude of the processed echo components in the slow time dimension. The echo energy of each range cell is compared within a preset range window, and the range cell with the highest energy is selected as the target range cell. The preset distance window can be defined based on the relative position of the millimeter-wave radar and the measured vibration part in the radar's line-of-sight direction. Specifically, firstly, the estimated distance interval corresponding to the measured vibration part is determined based on the relative position of the millimeter-wave radar and the measured vibration part in the radar's line-of-sight direction. Then, continuous distance units corresponding to the estimated distance interval are selected from the range-dimensional complex echo data to form the preset distance window. Subsequently, the echo energy of each distance unit within the preset distance window is compared, and the distance unit with the highest energy is selected as the target distance unit. In this way, the relative position of the millimeter-wave radar and the measured vibration part in the radar's line-of-sight direction can be transformed into a limitation of the preset distance window in the range-dimensional complex echo data, thereby avoiding interference from irrelevant reflectors on the target distance unit selection result. Step S22 achieves automatic locking of the target distance unit corresponding to the measured vibration target, providing a reliable range gate basis for the stable extraction of subsequent phase signals.

[0098] The instantaneous phase in step S23 is preferably obtained by performing a four-quadrant arctangent operation on the in-phase and quadrature components of the complex echo sequence, with its value range limited to the interval (-π, π]. Since the calculated instantaneous phase sequence is a wrapped phase sequence, and the millimeter-wave carrier wavelength is short, minute displacement changes on the target surface may cause phase changes exceeding π between adjacent sampling moments, resulting in discontinuous phase jumps in the wrapped phase sequence. To restore the true physical process of continuous phase change over time, phase unwrapping processing is performed on the wrapped phase sequence. As an example, the wrapped phase difference between the current sampling moment and the previous sampling moment in the slow time dimension is calculated. When a large wrapped phase difference is detected... When the phase difference is +π, subtract 2π from the current sampling phase; when the phase difference is detected to be less than -π, add 2π to the current sampling phase; otherwise, keep the current sampling phase unchanged. By performing point-by-point detection and compensation for phase jumps, the original wrapped phase sequence limited to the (-π, π] interval is converted into an unwound phase sequence that changes continuously with time, thereby ensuring the continuity and physical consistency of the subsequent displacement measurement process based on phase changes in the time dimension. The continuous phase sequence obtained in step S23 can accurately reflect the small displacement changes of the measured target in the radar line-of-sight direction, providing basic input data for subsequent vibration displacement reconstruction and vibration characteristic analysis.

[0099] This invention addresses the problems of micron-level mechanical vibration, ambient temperature drift, and device phase noise encountered by elevator traction machines during industrial operation. The phase-displacement mapping relationship of this invention is used to extract coarse vibration signals with clear physical meaning from radar echoes. The obtained coarse vibration signals have clear physical meaning and can provide accurate physical prior constraints for the PGGN (Physical Generation Network), ensuring the physical authenticity of the monitoring results. Specifically, step S3 includes S31-S32.

[0100] Reference Figure 5 Step S31: Calculate the dynamic phase reference of the continuous phase using the time averaging method within the sliding time window to achieve drift compensation;

[0101] The formula for the dynamic phase reference is as follows:

[0102] ,

[0103] in, Indicates the length of the sliding time window. Indicates the real-time sampling time. Indicates the start time of the sliding time window. This represents the integration variable, used to iterate through each historical moment within the sliding time window; Indicates a historical moment within the sliding time window. The continuous phase after unwinding.

[0104] Step S32: Based on the phase-displacement mapping relationship, the change in continuous phase relative to the dynamic phase reference is converted into a rough vibration signal.

[0105] The formula for the phase-displacement mapping relationship is as follows:

[0106] ,

[0107] in, Represents the moment The output rough vibration displacement reflects the real-time radial micro-motion of the traction machine surface; For millimeter wave wavelength, Indicates the real-time sampling time. For a moment Continuous phase after unwinding It serves as a dynamic phase reference, used to characterize the target's static distance and the phase offset caused by slowly changing environmental factors.

[0108] After phase-displacement mapping and dynamic reference elimination, the obtained rough vibration signal can accurately reflect the true vibration characteristics of the traction machine in the frequency dimension. Although the signal may still be affected by multipath effects and amplitude distortion caused by non-ideal installation conditions, its dominant frequency components have clear physical meanings and can be used as physical prior inputs to the PGGN (Physically Gated Generative Network) to constrain the learning space of the generative model, thereby improving the authenticity, stability, and interpretability of the overall vibration reconstruction results of the system.

[0109] Unlike traditional end-to-end deep learning solutions that directly process high-dimensional raw data, this invention employs the FMCW phase interferometry principle as a means of acquiring prior knowledge to extract coarse vibration signals containing real physical frequency characteristics. This step not only overcomes the limitations of radar hardware range resolution, achieving micron-level micro-motion sensing, but more importantly, it provides strong physical gating constraints for subsequent generative adversarial networks, fundamentally eliminating the signal illusion risk commonly found in deep generative models in industrial fault diagnosis, and ensuring the interpretability and reliability of diagnostic results.

[0110] To address the two major technical bottlenecks of existing millimeter-wave radar in elevator traction machine vibration monitoring: firstly, the non-stationary distribution of phase noise due to multipath effects and echo fading; and secondly, the physical limitation that a single radar can only measure one-dimensional vibration along the radar's line-of-sight. This invention proposes a hybrid generation architecture driven by both physical and data aspects: the PGGN (Physical Gated Generation Network). Instead of using the common end-to-end deep learning paradigm, this network innovatively constructs a cascaded structure of a 1DU-Net-based backbone network and anisotropic physical gating units. This structure utilizes deep neural networks to learn a high-dimensional nonlinear mapping from the one-dimensional radar observation domain to the three-axis mechanical vibration domain, while explicitly transforming the signal-to-noise ratio prior in radar signal processing into deterministic physical boundary constraints in the network topology, thereby achieving non-contact three-axis vibration signal reconstruction using a single radar.

[0111] Reference Figure 6 This scheme first constructs a dual-channel physical prior tensor as input data for the subsequent generator network. Specifically, step S4 includes:

[0112] S41: Calculate the physical confidence level based on the statistical characteristics of the echo amplitude of the target distance unit in the slow time dimension. The statistical characteristics of the echo amplitude include the mean and standard deviation of the echo amplitude calculated within a slow time sliding window composed of multiple adjacent Chirps.

[0113] The physical confidence level The formula is as follows:

[0114] ;

[0115] in, For a moment The echo amplitude corresponding to the target distance cell, ; and They are time points The in-phase and quadrature components of the complex echo data corresponding to the target range cell; and These are the mean and standard deviation of the echo amplitude within the slow-time sliding window, respectively. The Sigmoid normalization function maps the physical confidence level to the (0,1) interval; The closer the value is to 1, the stronger the radar echo signal at that moment, the smaller the phase noise of the physical measurement, and the higher the reliability; conversely, the lower the value is, the lower the reliability is, and network correction is required.

[0116] S42: For rough vibration signals within the sampling time period With physical confidence Discrete sampling is performed on each sample to obtain a rough vibration signal sequence. and physical confidence sequence Align and stack them along the time dimension to construct a two-channel physical prior tensor. .

[0117] Dual-channel physical prior tensor It only contains one-dimensional physical information about the radar line of sight, of which The number of sampling points within the input time window refers to the number of time sampling points processed during a single inference process of the neural network. Represents the space of the set of real numbers.

[0118] The dual-channel physical prior tensor is then input into the PGGN physical gating generation network. Step S5 includes:

[0119] S51: The PGGN physical gating generation network is pre-trained under the supervision of triaxial real vibration signals. During the training process, anisotropic physical gating units participate in the triaxial residual feature modulation, so that the physical confidence level constrains the triaxial residual feature generation process.

[0120] S52: Transfer the dual-channel physical prior tensor The trained PGGN physical gating generator network is input, and the generator outputs the original triaxial residual features. The anisotropic physical gating unit modulates the original triaxial residual features through physical confidence to generate gated triaxial residual features.

[0121] The coarse vibration signal sequence is dimensionally expanded to construct a triaxial reference tensor. The gated triaxial residual features and the triaxial reference tensor are superimposed to synthesize a triaxial reconstructed vibration signal tensor.

[0122] The generator of the PGGN physical gating generation network includes an encoder, a bottleneck layer, a decoder, and an output projection layer;

[0123] The encoder includes... A series of cascaded downsampled coding blocks, the input of the first downsampled coding block receiving a dual-channel physical prior tensor. The encoder extracts multi-scale temporal features, compressed temporal resolution, and expanded feature channel dimensions of the input signal layer by layer through multi-level downsampling coding blocks;

[0124] The bottleneck layer contains several dilated residual blocks stacked in series according to the signal flow direction. The bottleneck layer captures global context information and long temporal dependency features through the dilated residual blocks to generate deep semantic features.

[0125] The decoder contains components symmetrical to the encoder. The decoder uses a multi-level upsampling decoding block and the corresponding level of the encoder to concatenate features, gradually recovering the signal from deep semantic features to the original time resolution and reconstructing high-dimensional vibration waveform features.

[0126] The output projection layer maps the decoder's output from the abstract feature space to the physical signal space, generating the original triaxial residual features. .

[0127] As an example, the backbone network of the generator, built upon a deep convolutional neural network, has the core technical task of establishing a low-dimensional radar observation domain (containing coarse vibration signals and physical confidence, with a dimension of...). From the high-dimensional mechanical vibration domain (containing spatial triaxial vibration components, with dimensions of...) to the high-dimensional mechanical vibration domain (containing spatial triaxial vibration components, with dimensions of...) The nonlinear mapping of the generator is used to capture long-term temporal dependencies while preserving high-frequency phase details of micrometer-level vibrations. The generator's backbone network employs an asymmetric U-Net (1 DU-Net) architecture based on one-dimensional temporal signals. This architecture mainly consists of an encoder, a bottleneck layer, a decoder, and an output projection layer.

[0128] The encoder contains (preferred) The cascaded downsampled coded blocks form a contraction path. The input of the first coded block receives the dual-channel physical prior tensor constructed in the preceding steps. Each downsampling encoding block consists of a one-dimensional convolutional layer, an instance normalization layer, and a LeakyReLU activation function cascaded sequentially according to the signal flow direction. The one-dimensional convolutional layer is configured to perform downsampling operations with a stride greater than 1 (e.g., 2) and a preset kernel size (e.g., 4 or 5). The instance normalization layer is used to standardize the feature channels of a single sample to eliminate the difference in amplitude dimensions of millimeter-wave radar echo signals under different ranging distances and reflection intensities, accelerating network convergence. The LeakyReLU activation function is configured with a preset negative half-axis slope coefficient (e.g., 0.2) to introduce nonlinear mapping and retain weak gradient information on the negative half-axis, thereby enhancing the model's ability to express non-stationary vibration signals and preventing the gradient vanishing problem in deep network training.

[0129] The bottleneck layer comprises several (preferably 2 to 3) dilated residual blocks stacked in series according to the signal flow direction. Located between the output of the contraction path and the input of the expansion path, the bottleneck layer maintains the same number of feature channels in each dilated residual block (e.g., consistent with the number of channels at the end of the contraction path). No further temporal downsampling or upsampling operations are performed to focus on capturing long-term temporally dependent features at the lowest possible resolution. Each dilated residual block contains a dilated one-dimensional convolutional layer, an instance normalization layer, and a LeakyReLU activation function connected in the signal flow direction, and is configured with residual skip connections (i.e., feature splicing layers within the decoding block that skip connections) spanning these components to facilitate gradient propagation. The dilated one-dimensional convolutional layer is configured with a dilation rate greater than 1 (e.g., 2, 4, or 8 increasing exponentially by the layer level) and a stride of 1. Its core function is to significantly expand the effective receptive field of the convolutional kernel through dilated sampling while maintaining the temporal resolution and the number of network parameters, enabling it to capture long-term low-frequency fault trends covering the entire rotation cycle of the elevator traction machine. The instance normalization layer is used to standardize the deep feature channels of a single sample, ensuring the numerical stability of the data transmission in the bottleneck layer. The LeakyReLU activation function is configured with a preset negative half-axis slope coefficient (e.g., 0.2) to introduce nonlinear mapping and retain the weak gradient information of the negative half-axis, preventing neuron death in the deep network structure. The residual skip connection is configured to perform identity mapping, directly superimposing the input of the module to the output of the main transform branch, constructing a high-speed backpropagation path for gradients, effectively solving the gradient degradation problem in the training process of deep convolutional networks.

[0130] The decoder contains elements symmetrical to the encoder. Each upsampled decoding block forms an expansion path. Specifically, each upsampled decoding block consists of a one-dimensional transposed convolutional layer, a feature concatenation layer, a one-dimensional convolutional layer, an instance normalization layer, and a ReLU activation function, cascaded sequentially according to the signal flow direction. The system employs a 1D transposed convolutional layer to enhance the temporal resolution of the input feature map. The feature concatenation layer is configured to perform cascaded operations along the channel dimension, receiving two inputs: one from the upsampled output of the 1D transposed convolutional layer, and the other from the output of the corresponding layer (upsampled coding block) with the same temporal resolution in the shrinking path. A 1D convolutional layer performs channel dimensionality reduction and semantic fusion on the concatenated mixed features, ensuring the reconstructed signal retains both low-frequency trends and high-frequency details. An instance normalization layer standardizes the fused feature map, unifying the distribution differences between features from deep semantics and shallow details, preventing covariate shifts within the network, and ensuring numerical stability during the generation process. The ReLU activation function introduces a nonlinear mapping, further eliminating redundant noise during the fusion process through sparsification suppression (zeroing operation) of negative features, assisting the network in accurately reconstructing nonlinear mechanical vibration waveforms.

[0131] The output projection layer is located at the end of the generator backbone network, receiving the output feature map from the end of the expansion path. The output projection layer consists of a one-dimensional convolutional layer with a kernel size of 1 and a stride of 1. This one-dimensional convolutional layer is strictly configured with three output channels, corresponding to the X-axis (horizontal, defined as the radar line-of-sight direction), Y-axis (horizontal, vertical), and Z-axis (vertical) in the Cartesian coordinate system of the elevator traction machine space, respectively. This layer... The convolution operation performs a linear weighted combination between channels, aiming to project the high-dimensional semantic feature vector output by the decoder onto a low-dimensional three-axis physical vibration space. The data output by the one-dimensional convolutional layer is defined as the original three-axis residual features. This feature represents the fully quantified inference suggestion made by the neural network on the current triaxial vibration state based solely on the modal coupling probability distribution between the radar line-of-sight one-dimensional vibration and the triaxial physical vibration of the elevator traction machine, which is learned from the training data, before the deterministic constraints of the physical confidence gating unit are applied.

[0132] The method for generating gated triaxial residual features by the anisotropic physical gating units of the PGGN physical gating generation network includes:

[0133] The original triaxial residual features Decoupled according to channel dimension: radar line-of-sight direction component Longitudinal component and vertical components In this embodiment, a spatial Cartesian coordinate system is established with the vibrating part of the elevator traction machine as a reference. The X-axis, Y-axis, and Z-axis of the spatial Cartesian coordinate system are all orthogonal to each other. The X-axis is defined as the horizontal direction and is also defined as the radar line of sight direction; the Y-axis is defined as the horizontal direction; and the Z-axis is defined as the vertical direction. Therefore, the radar line of sight direction component corresponds to the X-axis direction component, the vertical component corresponds to the Y-axis direction component, and the vertical component corresponds to the Z-axis direction component.

[0134] Constructing time-varying suppression coefficients using physical confidence levels : ,in, As a physical sensitivity factor, the value range is set to [value range]. ;

[0135] Define the time-varying suppression coefficient The time-varying suppression coefficient sequence within the sampling period is as follows: Radar line-of-sight direction residual characteristics after physical gating constraints for: Where ⊙ represents the Hadamarda complex;

[0136] Longitudinal component and vertical components Data-driven mapping is used as the longitudinal residual feature. and vertical residual characteristics : , Wherein, the data-driven mapping refers to the vertical component. and vertical components The generator, learned from the training data, employs an identity mapping on these two directions in the anisotropic physical gating process, without introducing additional physical gating constraints, thereby obtaining the longitudinal residual features respectively. and Therefore, the formula and This indicates that the generator output remains unchanged during the gating phase;

[0137] Will , and Perform channel splicing to generate gated triaxial residual features. .

[0138] Physical sensitivity factor of time-varying suppression coefficient The value range is set to Under this configuration, when the radar echo signal-to-noise ratio is extremely high (physical confidence level)... When it approaches 1), the time-varying suppression coefficient As the value approaches zero, the residuals generated by the neural network are forced to zero, and the system output completely degenerates into the physical measurement value, thus eliminating the risk of the model exhibiting hallucinations in strong signal regions. In another implementation, It can be configured as a learnable scalar parameter, automatically finding the optimal physical-to-data tradeoff as the network trains.

[0139] Based on the step of calculating the radar line-of-sight direction residual characteristics (X-axis residual characteristics) after physical gating constraints based on the time-varying suppression coefficient sequence, a dynamic complementary mechanism is constructed: when multipath interference is severe or signal fading leads to a decrease in physical confidence, the suppression coefficient increases, the system automatically takes over and allows the neural network to use context information to repair the missing signal, otherwise it gives way to physical measurement.

[0140] Because the principle of FMCW radar dictates that it cannot acquire Doppler phase information in the longitudinal (Y-axis) and vertical (Z-axis) directions, there is no corresponding physical confidence index. Therefore, this element configures these two channels as an all-pass mode or an identity mapping, without introducing physical confidence for weighting, or considers the physical confidence in this direction to be always 0. Therefore, the longitudinal component... and vertical components Data-driven mapping is used as the longitudinal residual feature. (Y-axis residual characteristics) and vertical residual characteristics (Z-axis residual characteristics): , ;Will , and By splicing, gated triaxial residual features are generated. This processing logic makes the outputs of the Y and Z axes completely dependent on the mechanical vibration mode coupling laws learned by the backbone network during the training phase through a large amount of paired data.

[0141] The methods for synthesizing triaxial reconstructed vibration signal tensors using PGGN (Physically Gated Generative Network) synthesis units include:

[0142] For rough vibration signal sequences By expanding the dimensions, a three-axis reference tensor is constructed. , ,in, Represents the set of real numbers. This represents the number of sampling points within the sampling time period.

[0143] Gated three-axis residual characteristics As a correction term, it is superimposed onto the triaxial reference tensor. Generate a triaxial reconstructed vibration signal tensor , , , is the global residual scaling factor, which is an adjustable hyperparameter. The value range is usually [0,1], preferably set to 1, indicating that the residual correction of the network is fully accepted; in scenarios where conservative generation is required, the value can be appropriately reduced. This value is used to reduce the fluctuation amplitude of the generated signal.

[0144] Understandably, due to the physical limitations of a single millimeter-wave radar, it can only directly observe displacement changes along the radar's line-of-sight (X-axis), and cannot observe displacement changes along the longitudinal (Y-axis) and vertical (Z-axis) directions. Therefore, this embodiment employs a sparse-filling strategy to construct the reference observation tensor, directly assigning a rough vibration signal sequence to the radar's line-of-sight (X-axis) direction. The signal retains the true low-frequency vibration trend (such as the traction machine rotation frequency component); the non-radar line-of-sight components (Y-axis and Z-axis) are initialized to all-zero tensors due to the lack of physical measurement values, representing the lack of physical observation data.

[0145] The PGGN (Physical Gated Generation Network) of this invention needs to be trained before it can be put into use. In order to ensure that the generated reconstructed vibration signal is not only realistic in the time domain waveform, but also accurately retains the key features for fault diagnosis (such as the impact frequency of bearing failure) in the frequency domain, this invention constructs a physical + perception composite objective loss function and uses a supervised adversarial training framework for model optimization.

[0146] Reference Figure 1 and Figure 7 A contact-type triaxial vibration sensor is installed on the traction sheave of the elevator traction machine to collect real triaxial vibration signals. Using the real triaxial vibration signals as the supervised truth, an adversarial training strategy is adopted to alternately optimize the generator and discriminator. During the training process, the PGGN physical gating generation network is optimized with a composite objective loss function so that the reconstructed triaxial vibration signal approaches the real triaxial vibration signal in both time and frequency domain characteristics.

[0147] During the training phase, a dual-channel physical prior tensor acquired by millimeter-wave radar is loaded. As input, the synchronously acquired triaxial real vibration signals are used to construct a triaxial real vibration signal tensor. ,and To improve training stability and avoid gradient instability caused by excessive dimensional differences, numerical normalization is performed on the triaxial true vibration signal tensor: preferably, the Z-Score normalization strategy is first used to eliminate scale differences under different working conditions, and then linear scaling / clipping is used to map it to a numerical range consistent with the output range of the generator's terminal activation function (such as Tanh), thereby accelerating model convergence and improving training stability.

[0148] Composite objective loss function : ,in, For temporal consistency loss The weighting coefficients, Frequency domain envelope feature loss The weighting coefficients, To combat generation loss The weighting coefficients.

[0149] Temporal consistency loss: ,in, It is a triaxial reconstructed vibration signal tensor. It is a triaxial true vibration signal tensor. This represents the Frobenius norm squaring operation. The constrained triaxial reconstructed vibration signal exhibits pixel-level consistency with the actual triaxial vibration signal in terms of time-series waveforms.

[0150] Frequency domain envelope feature loss: ,in, The envelope spectrum tensor is obtained by extracting the envelope spectrum of the triaxial reconstructed vibration signal through Hilbert transform and frequency domain transform. The envelope spectrum tensor is obtained by extracting the envelope spectrum of a real triaxial vibration signal through Hilbert transform and frequency domain transform. This represents the element-wise modulo operation. Represents logarithmic transformation, It is the minimum constant. This represents the Frobenius norm square operation. The frequency domain envelope feature loss ensures that the generated triaxial reconstructed vibration signal accurately preserves the amplitude-modulated fault characteristics unique to rotating machinery in the frequency domain, such as periodic impacts and their modulation sidebands caused by bearing defects. This loss forces the generator to focus on the signal's modulation sidebands and fault impact frequencies, preventing the generated signal from losing crucial diagnostic information despite having a similar waveform.

[0151] Adversarial generation loss: ,in, It is a two-channel physical prior tensor. Represents a generator network. This represents the discriminator network. This represents the discriminator's judgment result on the generated samples, where 1 represents the target label of the real sample. This represents the Frobenius norm square operation. The adversarial generation loss stems from the game between the discriminator and the generator. Discriminator An attempt was made to distinguish the generated triaxial reconstructed vibration signal from the actual contact vibrometer signal, while the generator... It attempts to deceive the discriminator. The geometric meaning of this loss function is to minimize the mean square distance between the discriminator score and the target value of 1, thereby improving the overall realism and naturalness of the generated signal's details and textures.

[0152] A contact-type triaxial vibration sensor is installed in the elevator traction machine to collect real triaxial vibration signals. Using the real triaxial vibration signals as the supervised truth, an adversarial training strategy is adopted to alternately optimize the generator and discriminator. During the training process, the PGGN physical gated generation network is optimized with a composite objective loss function so that the reconstructed triaxial vibration signal approaches the real triaxial vibration signal in both time and frequency domain characteristics.

[0153] As an example: A unified hardware trigger signal is provided to the millimeter-wave radar and triaxial vibration sensor through an external synchronization signal source, enabling both types of sensors to synchronously start sampling under the same time reference. The raw echo data output by the millimeter-wave radar and the synchronously acquired triaxial vibration sensor data are parsed and paired, and a multimodal dataset is constructed in chronological order. The multimodal dataset comes from no fewer than 12 in-use elevator traction machines, covering different traction machine models, rated speeds, and load conditions. Hardware synchronous data acquisition is performed on each elevator under normal operating conditions, with an effective acquisition time of no less than 30 minutes for each device.

[0154] The synchronously acquired echo signals and triaxial real vibration signals were time-aligned and processed using a sliding window slicing method, with a window length of 2048 points and a step size of 512 points, resulting in at least 18,000 samples. The samples were grouped according to elevator numbers, with the training set, validation set, and test set accounting for 80%, 10%, and 10% respectively, to avoid data from the same device appearing in different subsets.

[0155] During the model training phase, the Adam optimization algorithm is preferably used for parameter updates of both the generator and discriminator. The initial learning rates for the generator and discriminator can be set to... momentum coefficient and The values ​​are 0.5 and 0.999 respectively. The above parameters are not the only options in this invention; those skilled in the art can use them in practice. arrive The learning rate is adjusted within a reasonable range and a reasonable combination of momentum coefficients. The batch size for model training can be set to 32, and equivalent batch training can be achieved through gradient accumulation when memory is limited. To ensure the physical accuracy of the generated vibration signal, it is preferable to make the weight of the time-domain reconstruction loss higher than that of other loss terms; for example, the weight of the time-domain reconstruction loss can be set to 10, the weight of the frequency-domain feature loss can be set to 5, and the weight of the adversarial loss can be set to 1. Through the above weight configuration, the model training process prioritizes ensuring the overall waveform consistency of the generated signal in the time domain, while strengthening the frequency domain constraints on key fault feature frequency bands, and on this basis, taking into account the statistical fidelity of the generated signal, thereby improving the overall performance of the vibration signal reconstruction result while ensuring training stability.

[0156] Under the operating condition of the traction machine at a speed of 153 r / min (corresponding to a fundamental frequency of approximately 2.55 Hz), the echo signal from a millimeter-wave radar was collected. The echo signal was processed using the method of this invention and input into the trained PGGN physical gated generation network to synthesize a triaxial reconstructed vibration signal. The triaxial reconstructed vibration signal and the actual triaxial vibration signal were compared in the time and frequency domains as follows: Figure 8 and Figure 9As shown in the figure, the blue curve represents the actual triaxial vibration signal acquired by the contact triaxial vibration sensor, and the red curve represents the triaxial reconstructed vibration signal generated by the method of this invention. Figure 8 It is evident that, in the three time domain directions, the reconstructed triaxial vibration signal is basically consistent with the actual triaxial vibration signal in terms of period variation trend, dominant frequency position, and amplitude variation range. The peak-valley position correspondence is clear, and no obvious DC drift or non-physical jump phenomenon is observed. Figure 9 It can be seen that, in the three directions of the frequency domain, the triaxial reconstructed vibration signal is basically consistent with the triaxial real vibration signal in terms of spectral distribution, dominant frequency components and amplitude energy distribution. The correspondence between the frequency components of each order is clear, and there are no false frequency components, abnormal spectral peaks or non-physical energy abrupt changes.

[0157] In step S6 of the present invention, the main frequency distribution, vibration energy change trend and inter-axis synchronization characteristics are extracted from the triaxial reconstructed vibration signal to evaluate the physical rationality of the triaxial reconstructed vibration signal, and the physical gating constraints are adaptively adjusted according to the evaluation results.

[0158] As an example: Within a preset time window, the dominant frequency distribution of the triaxial reconstructed vibration signal is analyzed. Specifically, when the traction machine speed is 153 r / min, its corresponding rotational frequency is approximately 2.55 Hz; let the preset time window length be... Spectral analysis is performed on the triaxial reconstructed vibration signal within each preset time window. Since the spectral resolution is determined by the time window length, it can be calculated as the reciprocal of the preset time window length. ,in Indicates spectral resolution; when When, it can be calculated That is, the smallest frequency interval that the spectral analysis can distinguish is approximately 0.05 Hz; furthermore, the threshold for determining the dominant oscillation frequency is preferably set to twice the spectral resolution, i.e., 0.10 Hz, to balance the resolution error of the spectral analysis with the allowable range of actual vibration fluctuations; if in the... The principal frequencies of the X, Y, and Z axes extracted within the first time window were 2.54 Hz, 2.57 Hz, and 2.56 Hz, respectively. Within the first three time windows, the frequencies are 2.56Hz, 2.58Hz, and 2.57Hz respectively. Therefore, the changes in the principal oscillation frequency of each axis between adjacent time windows are 0.02Hz, 0.01Hz, and 0.01Hz respectively, all not exceeding 0.10Hz. Meanwhile, the... Within a given time window, the deviations of the main oscillation frequencies of the X, Y, and Z axes relative to the rotational frequency of 2.55 Hz are 0.01 Hz, 0.03 Hz, and 0.02 Hz, respectively. The difference in main oscillation frequencies between the X and Y axes is 0.02 Hz, between the Y and Z axes is 0.01 Hz, and between the X and Z axes is 0.01 Hz, all not exceeding 0.10 Hz. Based on this, it can be determined that the current main oscillation frequency distribution remains stable in the time dimension and is basically consistent with the traction machine rotational frequency, showing frequency consistency among the three axes, thus possessing physical rationality. Conversely, if the change in the main oscillation frequency of any axis between adjacent time windows, or the deviation of the main oscillation frequency of any axis relative to the rotational frequency, or the difference in main oscillation frequencies between any two axes is greater than 0.10 Hz, it can be determined that the current main oscillation frequency distribution exhibits drift or mismatch.

[0159] As an example: Within a preset time window, the vibration energy variation trend of the triaxial reconstructed vibration signal is analyzed. Specifically, within the preset time window, the squared amplitudes of the X-axis, Y-axis, and Z-axis reconstructed vibration signals at each sampling point are cumulatively averaged to obtain the corresponding vibration energy parameters. , and This is used to characterize the overall intensity of the reconstructed vibration signal in each axis within the time window; the change in vibration energy can be determined by the ratio of the absolute value of the difference between the vibration energy parameters of the current time window and the previous time window to the vibration energy parameters of the previous time window; the threshold for judging the change in vibration energy can be pre-calibrated based on the statistical results of the change in vibration energy under normal operating conditions. When the mean value of the change in vibration energy under normal operating conditions is... Standard deviation is The threshold for determining a steady change is preferably set to [value]. The threshold for determining mutations is preferably set to... This setting is based on the statistical distribution of normal operating conditions, using "mean + different multiples of standard deviation" to distinguish between normal fluctuation ranges and abnormal change ranges, thereby achieving a stratified determination of vibration energy change trends; when , At that time, one can obtain , If in the first The vibration energy parameters of the X-axis, Y-axis, and Z-axis within the first time window are 1.00, 0.95, and 1.05, respectively. Within a time window, the vibration energy parameters of the X-axis, Y-axis, and Z-axis are 1.04, 0.99, and 1.08, respectively. The corresponding vibration energy changes are 0.040, 0.042, and 0.029, all of which are no greater than 0.05, indicating that the current vibration energy change trend is stable. If the vibration energy change in any axis is greater than 0.05 but not greater than 0.07, it is considered a slow change. If the vibration energy change in any axis is greater than 0.07, it is considered a sudden change.

[0160] As an example: Within a preset time window, the inter-axis synchronization characteristics of the reconstructed vibration signals of the three axes are analyzed. Specifically, within the preset time window, the reconstructed vibration signal sequences of the X-axis, Y-axis, and Z-axis are acquired and time-aligned. Further, the normalized cross-correlation coefficients of any two axis vibration signals within this time window and the peak delay corresponding to their maximum values ​​are calculated to characterize the degree of synchronization of different axial vibration signals in terms of their changing trends, main vibration components, and time correspondence. The inter-axis synchronization judgment thresholds include the inter-axis synchronization correlation judgment threshold and the inter-axis synchronization delay judgment threshold. Both can be pre-calibrated based on the correlation and peak delay statistics of different axial vibration signals within multiple consecutive time windows under normal operating conditions of the traction machine. This setting is based on the statistical distribution under normal operating conditions, using "lower correlation limit + upper delay limit" to distinguish between normal coupling relationships and abnormal mismatch relationships. When the mean of the normalized cross-correlation coefficients of any two axis vibration signals under normal operating conditions is... Standard deviation is The threshold for determining inter-axis synchronicity correlation is preferably set as follows: Meanwhile, the mean peak delay corresponding to the maximum value of the statistically normalized cross-correlation coefficient is... Standard deviation is The threshold for determining inter-axis synchronization delay is preferably set as follows: When the mean of the normalized cross-correlation coefficient is 0.92 and the standard deviation is 0.04, and the mean of the peak delay is 0.008s and the standard deviation is 0.004s, under normal operating conditions, the following can be obtained: , If in the first Within a given time window, the normalized cross-correlation coefficients of the X-axis and Y-axis, Y-axis and Z-axis, and X-axis and Z-axis vibration signals are 0.90, 0.88, and 0.86, respectively, with peak delays of 0.010s, 0.012s, and 0.009s, respectively. Therefore, the normalized cross-correlation coefficient between any two axes is not less than 0.84, and the peak delay is not greater than 0.016s, indicating a stable synchronization relationship between the vibration signals of different axes. If the normalized cross-correlation coefficient between any two axes is less than 0.84, or the peak delay is greater than 0.016s, then the synchronization between the axes is weakened or mismatched. Thus, the synchronization characteristics of the three-axis reconstructed vibration signal can be evaluated based on the changes in the synchronization degree of vibration signals of different axes within a continuous time window.

[0161] By comparing the correlation between the vibration components in the radar line-of-sight direction and the vibration components in the non-line-of-sight direction along the aforementioned characteristic dimensions, the rationality of the current triaxial reconstructed vibration signal under physical structural constraints is evaluated.

[0162] ① When the three-axis reconstructed vibration signal is detected to maintain a stable correlation in terms of main frequency distribution, vibration energy change trend and inter-axis synchronization characteristics, the current reconstruction result is determined to be reliable, and the corresponding three-axis reconstructed vibration signal is output and stored as effective input data for subsequent elevator traction machine operation status analysis or trend monitoring.

[0163] ② When a mismatch is detected in any of the above-mentioned characteristic dimensions of the triaxial reconstructed vibration signal, indicating that the current reconstruction result may be affected by changes in radar measurement conditions or environmental interference, the evaluation result is fed back to the PGGN physical gating generation network module in step S5 for adaptive adjustment of the physical confidence gating strategy. The adjustment method includes: classifying the degree of mismatch of the current triaxial reconstructed vibration signal according to the evaluation results of the dominant frequency distribution, vibration energy change trend, and inter-axis synchronization characteristics; when all three evaluation results meet the preset judgment conditions, the current reconstruction result is judged to meet the physical rationality requirements, and the physical sensitivity factor is maintained. and global residual scaling factor Unchanged; when only one evaluation result fails to meet the preset judgment condition, and that evaluation result only deviates slightly from the corresponding judgment threshold, while all other evaluation results meet the preset judgment condition, it is judged as a slight mismatch, and the physical sensitivity factor is slightly increased. and / or slightly reduce the global residual scaling factor If two or more evaluation results fail to meet the preset judgment criteria, or if any evaluation result deviates significantly from the corresponding judgment threshold, it is judged as a significant mismatch, and the physical sensitivity factor is further increased. and further reduce the global residual scaling factor This enhances the suppression of radar line-of-sight residuals and reduces the overall correction magnitude of the triaxial generation residuals to the final output.

[0164] By employing method ② described above, the generated triaxial reconstructed vibration signal is used to achieve online evaluation and feedback control of the reconstruction quality, forming a closed-loop structure for the vibration signal reconstruction process, encompassing physical measurement, generation and completion, result evaluation, and parameter adjustment. This improves the stability and engineering reliability of non-contact vibration reconstruction results under complex working conditions without introducing additional sensors or external reference signals.

[0165] Accordingly, the present invention also provides an elevator traction machine vibration reconstruction system based on millimeter-wave radar to implement the above-mentioned method;

[0166] The system includes:

[0167] The data acquisition module includes a millimeter-wave radar, used to collect echo signals during elevator operation;

[0168] The signal preprocessing module is used to determine the target range unit through the echo signal and extract the corresponding phase information to obtain a continuous phase sequence;

[0169] The rough vibration estimation module uses a dynamic phase reference to compensate for the drift of the continuous phase sequence and generates a rough vibration signal in the radar line-of-sight direction based on the phase-displacement mapping relationship.

[0170] The prior vibration estimation module calculates the physical confidence level based on the statistical characteristics of the echo amplitude of the target distance unit, and constructs a dual-channel physical prior tensor after aligning the rough vibration signal and the physical confidence level in the time dimension.

[0171] The signal reconstruction module is equipped with a PGGN physical gated generation network consisting of a generator and anisotropic physical gated units cascaded together. After the dual-channel physical prior tensor is input into the PGGN physical gated generation network, the generator outputs the original triaxial residual features. After being gated and modulated by the anisotropic physical gated units, gated triaxial residual features are generated. The gated triaxial residual features are then superimposed with the rough vibration signal to synthesize a triaxial reconstructed vibration signal.

[0172] The evaluation module is used to evaluate the physical rationality of the triaxial reconstructed vibration signal and adjust the gating parameters of the anisotropic physical gating unit based on the evaluation results.

[0173] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0174] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0175] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a multimedia terminal device (which may be a mobile phone, computer, television receiver, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0176] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A vibration reconstruction method for elevator traction machines based on millimeter-wave radar, characterized in that, Includes the following steps: S1: Aim the main beam of the millimeter-wave radar at the vibrating part of the elevator traction machine and sample the echo signal during elevator operation. S2: Determine the target range unit through the echo signal and extract the corresponding phase information to obtain a continuous phase sequence; S3: The continuous phase sequence is drift compensated using a dynamic phase reference, and a rough vibration signal in the radar line-of-sight direction is generated based on the phase-displacement mapping relationship; S4: Calculate the physical confidence level based on the statistical characteristics of the echo amplitude of the target distance unit, and construct a dual-channel physical prior tensor after aligning the rough vibration signal and the physical confidence level in the time dimension; S5: The dual-channel physical prior tensor is input into the PGGN physical gated generation network, which is composed of a generator and anisotropic physical gated units cascaded together. The generator outputs the original triaxial residual features, which are then gated and modulated by the anisotropic physical gated units to generate gated triaxial residual features. The gated triaxial residual features are then superimposed with the rough vibration signal to synthesize a triaxial reconstructed vibration signal tensor. The generator includes an encoder, a bottleneck layer, a decoder, and an output projection layer. S6: Perform a physical rationality assessment on the triaxial reconstructed vibration signal tensor, and adjust the gating parameters of the anisotropic physical gating unit based on the assessment results.

2. The method according to claim 1, characterized in that, Step S2 includes: S21: Using a single linear frequency modulated pulse as the processing unit, a window function is applied in the fast time dimension to suppress spectral sidelobe leakage. Then, through a range-to-fast Fourier transform, the time-domain discrete signal is converted into complex echo data arranged in range units. S22: Arrange the complex echo data of the same distance unit under multiple linear frequency modulation pulses along the slow time dimension to form a complex echo sequence. Extract the zero frequency component of the complex echo sequence to obtain the corresponding complex value. Calculate the amplitude square of the complex value to determine the energy of each distance unit. Select the distance unit with the largest energy as the target distance unit within the preset distance window. S23: Extract the complex echo sequence of the target distance cell, calculate the instantaneous phase to obtain the wrapped phase sequence; calculate the wrapped phase difference between the current sampling time and the previous sampling time in the slow time dimension. When the wrapped phase difference is greater than +π, subtract 2π from the wrapped phase at the current sampling time; when the wrapped phase difference is less than -π, add 2π to the wrapped phase at the current sampling time; otherwise, keep the wrapped phase at the current sampling time unchanged, complete the phase unwrapping, and obtain a continuous phase sequence.

3. The method according to claim 2, characterized in that, Step S3 includes: S31: The dynamic phase reference of continuous phase is calculated by the time averaging method within a sliding time window to achieve drift compensation; The formula for the dynamic phase reference is as follows: ， in, Indicates the length of the sliding time window. Indicates the real-time sampling time. Indicates the start time of the sliding time window. Represents the integral variable. Indicates a historical moment within the sliding time window. Continuous phase after unwinding; S32: Based on the phase-displacement mapping relationship, the change in continuous phase relative to the dynamic phase reference is converted into a rough vibration signal; The formula for the phase-displacement mapping relationship is as follows: ， in, Represents the moment The output rough vibration displacement, For millimeter wave wavelength, Indicates the real-time sampling time. For a moment Continuous phase after unwinding It serves as a dynamic phase reference.

4. The method according to claim 3, characterized in that, Step S4 includes: S41: Calculate the physical confidence level based on the statistical characteristics of the echo amplitude of the target distance cell, the physical confidence level... The formula is as follows: ； in, For a moment The echo amplitude corresponding to the target distance cell, ; and They are time points The in-phase and quadrature components of the complex echo data corresponding to the target range cell; and These are the mean and standard deviation of the echo amplitude within the slow-time sliding window, respectively. The Sigmoid normalization function maps the physical confidence level to the (0,1) interval; S42: For rough vibration signals within the sampling time period With physical confidence Discrete sampling is performed on each sample to obtain a rough vibration signal sequence. and physical confidence sequence Align and stack them along the time dimension to construct a two-channel physical prior tensor. .

5. The method according to claim 1, characterized in that, Step S5 includes: S51: The PGGN physical gating generation network is pre-trained under the supervision of triaxial real vibration signals. During the training process, anisotropic physical gating units participate in the triaxial residual feature modulation, so that the physical confidence level constrains the triaxial residual feature generation process. S52: Transfer the dual-channel physical prior tensor The trained PGGN physical gating generator network is input, and the generator outputs the original triaxial residual features. The anisotropic physical gating unit modulates the original triaxial residual features through physical confidence to generate gated triaxial residual features. The coarse vibration signal sequence is dimensionally expanded to construct a triaxial reference tensor. The gated triaxial residual features and the triaxial reference tensor are superimposed to synthesize a triaxial reconstructed vibration signal tensor.

6. The method according to claim 5, characterized in that, The generator includes an encoder, a bottleneck layer, a decoder, and an output projection layer; The encoder includes... A series of cascaded downsampled coding blocks, the input of the first downsampled coding block receiving a dual-channel physical prior tensor. The encoder extracts multi-scale temporal features, compressed temporal resolution, and expanded feature channel dimensions of the input signal layer by layer through multi-level downsampling coding blocks; The bottleneck layer contains several dilated residual blocks stacked in series according to the signal flow direction. The bottleneck layer captures global context information and long temporal dependency features through the dilated residual blocks to generate deep semantic features. The decoder contains components symmetrical to the encoder. The decoder uses a multi-level upsampling decoding block and the corresponding level of the encoder to concatenate features, gradually recovering the signal from deep semantic features to the original time resolution and reconstructing high-dimensional vibration waveform features. The output projection layer maps the decoder's output from the abstract feature space to the physical signal space, generating the original triaxial residual features. .

7. The method according to claim 5, characterized in that, The method for generating gated triaxial residual features using the anisotropic physical gating unit includes: The original triaxial residual features Decoupled according to channel dimension: radar line-of-sight direction component Longitudinal component and vertical components ; Constructing time-varying suppression coefficients using physical confidence levels : ,in, As a physical sensitivity factor, the value range is set to [value range]. ; Define the time-varying suppression coefficient The time-varying suppression coefficient sequence within the sampling period is as follows: Radar line-of-sight direction residual characteristics after physical gating constraints for: Where ⊙ represents the Hadamarda product; Longitudinal component and vertical components Data-driven mapping is used as the longitudinal residual feature. and vertical residual characteristics : , ; Will , and Perform channel splicing to generate gated triaxial residual features. .

8. The method according to claim 5, characterized in that, The method for synthesizing triaxial reconstructed vibration signal tensors includes: For rough vibration signal sequences Expand the dimensions to construct a three-axis reference tensor , ,in, Represents the set of real numbers. This represents the number of sampling points within the sampling time period. Gated three-axis residual characteristics As a correction term, it is superimposed onto the triaxial reference tensor. Generate a triaxial reconstructed vibration signal tensor , , , This is the global residual scaling factor.

9. The method according to claim 5, characterized in that, The method for training the PGGN physical gating generative network includes: Constructing a composite objective loss function : ,in, For temporal consistency loss The weighting coefficients, Frequency domain envelope feature loss The weighting coefficients, To combat generation loss Weighting coefficients; ,in, It is a triaxial reconstructed vibration signal tensor. It is a triaxial true vibration signal tensor. This represents the Frobenius norm square operation; ,in, The envelope spectrum tensor is obtained by extracting the envelope spectrum of the triaxial reconstructed vibration signal through Hilbert transform and frequency domain transform. The envelope spectrum tensor is obtained by extracting the envelope spectrum of a real triaxial vibration signal through Hilbert transform and frequency domain transform. This represents the element-wise modulo operation. Represents logarithmic transformation, It is the minimum constant. This represents the Frobenius norm square operation; ,in, For a two-channel physical prior tensor Represents a generator network. This represents the discriminator network. This represents the discriminator's judgment result on the generated samples, where 1 represents the target label of the real sample. This represents the Frobenius norm square operation; A contact-type triaxial vibration sensor is installed in the elevator traction machine to collect real triaxial vibration signals. Using the real triaxial vibration signals as the supervised truth, an adversarial training strategy is adopted to alternately optimize the generator and discriminator. During the training process, the PGGN physical gated generation network is optimized with a composite objective loss function so that the reconstructed triaxial vibration signal approaches the real triaxial vibration signal in both time and frequency domain characteristics.

10. A vibration reconfiguration system for elevator traction machines based on millimeter-wave radar, characterized in that, The system is used to implement the method according to any one of claims 1-9; The system includes: The data acquisition module includes a millimeter-wave radar, used to collect echo signals during elevator operation; The signal preprocessing module is used to determine the target range unit through the echo signal and extract the corresponding phase information to obtain a continuous phase sequence; The rough vibration estimation module uses a dynamic phase reference to compensate for the drift of the continuous phase sequence and generates a rough vibration signal in the radar line-of-sight direction based on the phase-displacement mapping relationship. The prior vibration estimation module calculates the physical confidence level based on the statistical characteristics of the echo amplitude of the target distance unit, and constructs a dual-channel physical prior tensor after aligning the rough vibration signal and the physical confidence level in the time dimension. The signal reconstruction module is equipped with a PGGN physical gated generation network consisting of a generator and anisotropic physical gated units cascaded together. After the dual-channel physical prior tensor is input into the PGGN physical gated generation network, the generator outputs the original triaxial residual features. After being gated and modulated by the anisotropic physical gated units, gated triaxial residual features are generated. The gated triaxial residual features are then superimposed with the rough vibration signal to synthesize a triaxial reconstructed vibration signal. The evaluation module is used to evaluate the physical rationality of the triaxial reconstructed vibration signal and adjust the gating parameters of the anisotropic physical gating unit based on the evaluation results.

Images