A method for monitoring elevator car vibration anomalies

Abstract

This invention provides a method for monitoring abnormal vibration in elevator cars, belonging to the field of elevator monitoring technology. This invention utilizes sparse representation theory to construct an overcomplete dictionary and uses a matching pursuit algorithm to sparsely reconstruct the impact component from mixed vibration signals. It then combines a spectral kurtosis map to locate the optimal filtering frequency band, converts the impact component into a spatial domain signal based on the car speed integral result, constructs a multi-dimensional vibration feature matrix sample covariance matrix, and uses random matrix theory to derive a collaborative monitoring algorithm for elevator group vibration status to locate abnormal elevators. Finally, it inputs the impact component characteristics, spatial location coordinates, and degradation rate assessment value into a physical constraint, embedding them into a temporal inference artificial intelligence model to output fault category and severity score. Finally, it generates a maintenance decision report based on a comprehensive risk index. This invention solves the technical problem of accurately locating the spatial location of elevator guide rail defects and assessing their degradation trend in real time under conditions of multiple concurrent impact sources and insufficient signal-to-noise ratio.

Description

Technical Field

[0001] This invention belongs to the field of elevator monitoring technology, and more specifically, relates to a method for monitoring abnormal vibration of an elevator car. Background Technology

[0002] Elevator car vibration anomaly monitoring is a crucial technical means to ensure elevator operational safety. Current technologies typically employ methods such as time-domain statistical feature extraction, frequency-domain power spectrum analysis, or short-time Fourier transform to extract features from vibration signals collected by accelerometers for abnormal car vibration caused by elevator guide rail defects. These features are then combined with threshold judgment or shallow machine learning models to achieve fault detection. These methods demonstrate certain detection capabilities under laboratory conditions with a single excitation source and high signal-to-noise ratio, and have already been applied in some practical engineering scenarios.

[0003] However, in actual elevator operation, the guide rail has multiple installation joints. When the elevator runs at different speeds, multiple impact excitation sources simultaneously generate vibration signals, and the impact component and noise component in the signal are highly coupled. Traditional frequency domain methods cannot distinguish the contribution of each excitation source when multiple impact sources are present simultaneously, leading to feature aliasing. Time domain statistical methods are not sensitive enough to early weak impacts and are difficult to achieve effective detection under low signal-to-noise ratio conditions. Furthermore, the frequency modulation effect caused by variable speed operation makes it impossible to directly reflect the spatial periodicity of the impact signal in the time domain, making it difficult for existing methods to map the detection results back to the specific spatial location of the guide rail.

[0004] In existing technologies, multiple guide rail joints simultaneously generate impact excitation during elevator operation, and the variable speed condition causes the intervals of impact events in the time domain to be unpredictable. This makes it impossible for traditional time-domain or frequency-domain methods to separate the impact components from the aliased signal or eliminate the variable speed modulation effect to restore the spatial periodicity of the impact. Consequently, the spatial coordinates of multiple impact sources cannot be accurately estimated, and the degradation trend of early defects cannot be continuously quantified and tracked. In other words, existing technologies suffer from the technical problem of being unable to accurately locate the spatial position of elevator guide rail defects and assess their degradation trend in real time under conditions of multiple concurrent impact sources and insufficient signal-to-noise ratio. Summary of the Invention

[0005] In view of this, the present invention provides a method for monitoring abnormal vibration of elevator car, which can solve the technical problem in the prior art that it is impossible to accurately locate the spatial location of elevator guide rail defects and assess their degradation trend in real time under the condition of multiple impact sources occurring simultaneously and insufficient signal-to-noise ratio.

[0006] This invention is implemented as follows: This invention provides a method for monitoring abnormal vibration in an elevator car, comprising the following steps:

[0007] An overcomplete dictionary of impact signals is constructed using sparse representation theory. The impact components are sparsely reconstructed from the mixed vibration signals using a matching pursuit algorithm. The optimal filtering frequency band is located by combining the spectral kurtosis diagram.

[0008] The impact components are converted into spatial domain signals based on the car speed integral result. The spatial position coordinates of each impact excitation source are estimated using the guide rail spatial spectrum estimation periodic impact positioning algorithm, and the degradation rate evaluation value is output.

[0009] For all online elevators in the elevator group, a sample covariance matrix of the multidimensional vibration feature matrix is ​​constructed within the same time window. The elevator group vibration state collaborative monitoring algorithm is derived using random matrix theory to calculate the overall abnormal index of the elevator group and locate the abnormal elevator number.

[0010] The impact component characteristics, spatial location coordinates, and degradation rate assessment values ​​are input into the physical constraints and embedded into the time-series inference model, which outputs the fault category and fault severity score.

[0011] A lock-free circular buffer queue and a task directed acyclic graph asynchronous parallel scheduling framework are adopted to dynamically allocate feature extraction tasks and model inference tasks to multi-core processors;

[0012] Based on the fault severity score, spatial location coordinates, and degradation rate assessment value, a comprehensive risk index is calculated using an adaptive adjustment function for the fault severity score, and a maintenance decision report is generated.

[0013] The sparse representation theory refers to the signal decomposition theory that represents a signal as a linear combination of a few atoms in an overcomplete dictionary. The number of atoms in the overcomplete dictionary is much greater than the signal dimension, so that the impact signal has sparse coefficients in the dictionary domain.

[0014] In this context, each atom in the overcomplete dictionary corresponds to an impact waveform template. The atomic shape is designed based on the physical response characteristics of the impact on the elevator guide rail joint. The coverage range of the atomic center frequency and the coverage range of the atomic time width are obtained by performing time-frequency analysis on the measured impact signal and statistically analyzing the distribution range of waveform parameters.

[0015] The matching pursuit algorithm refers to selecting the atom with the largest inner product with the current residual signal from the overcomplete dictionary in each iteration, estimating its coefficient and subtracting it from the residual, and iterating cyclically until the residual energy is lower than a preset energy threshold, which is a multiple of the average noise energy.

[0016] The spatial domain signal refers to the acceleration signal with the car travel distance as the independent variable. The travel distance is obtained by integrating the car speed over time, and the frequency modulation effect caused by the variable speed operation is eliminated after conversion.

[0017] Specifically, the guide rail spatial spectrum estimation periodic impact positioning algorithm involves constructing a spatial domain signal observation matrix with multiple spatial windows, calculating the sample covariance matrix of the observation matrix and performing eigenvalue decomposition, constructing a multi-signal classification spatial spectrum function based on the orthogonality principle of the signal subspace and noise subspace, scanning spatial frequencies and searching for spectral peak positions.

[0018] The degradation rate assessment value is obtained by normalizing the slope of the spatial spectrum peak amplitude sequence of each abnormal impact source in multiple consecutive operating cycles through linear regression. The normalization method is to divide the original slope by the maximum value of the slope of the historical normal sample.

[0019] Specifically, the random matrix theory-derived elevator group vibration state collaborative monitoring algorithm determines the theoretical boundary of the sample covariance matrix eigenvalues ​​based on the Marchinko-Pastur law, uses the ratio of the deviation between the measured maximum eigenvalue and the theoretical boundary as the overall abnormality index of the elevator group, and uses the component with the highest amplitude in the eigenvector corresponding to the maximum eigenvalue exceeding the boundary to locate the abnormal elevator number.

[0020] The physical constraint embedded temporal reasoning model includes a bidirectional long short-term memory layer for extracting short-term local temporal dependencies, a sparse self-attention converter module for processing long-term temporal dependencies, and a physical constraint embedding mechanism that encodes prior physical parameters such as guide rail installation spacing into position bias matrices and injects them into attention calculations. It also includes an internal iterative refinement module that performs multiple self-iterative corrections to the preliminary diagnostic results.

[0021] Specifically, the position offset matrix is ​​constructed by calculating the time interval corresponding to the impact event based on the guide rail installation spacing and the current car speed, then calculating the cosine value using the ratio of the time step index difference to the time interval as the independent variable, and using the obtained value as the element of the position offset matrix.

[0022] The lock-free circular buffer queue refers to a fixed-capacity circular queue that is thread-safe using atomic operations instead of mutexes. Multiple acquisition threads concurrently write vibration data by atomically incrementing the write pointer, and consumption threads read data by atomically incrementing the read pointer.

[0023] The task directed acyclic graph asynchronous parallel scheduling framework refers to constructing an acyclic directed graph with computing tasks as nodes and data dependencies as directed edges. The scheduling framework determines the task execution order based on topological sorting and simultaneously distributes parallel nodes without data dependencies to multi-core processors for execution.

[0024] The comprehensive risk index is calculated by weighting and summing the fault severity score, the deterioration rate assessment value and the normalized value of the overall abnormality index of the elevator group. The sum of the three weight coefficients is 1. The weight coefficients are determined by normalizing each indicator with the Pearson correlation coefficient of the actual fault occurrence probability.

[0025] Specifically, the fault severity score adaptive adjustment function lowers the decision threshold of the physical constraint embedded in the time-series inference model to improve the fault recall rate when the comprehensive risk index is not lower than the high-risk threshold, raises the decision threshold to reduce the false alarm rate when the comprehensive risk index is lower than the low-risk threshold, and keeps the decision threshold unchanged when it is between the two.

[0026] Among them, the preset energy threshold multiple ranges from 1.2 to 1.5; the residual matching threshold ranges from 5 to 15 cm; the number of consecutive operating cycles for the degradation rate assessment value ranges from 30 to 90; the alarm threshold for the overall abnormal index of the elevator group is taken as the 99th percentile of the distribution of the overall abnormal index of the elevator group obtained from the vibration characteristic data of no less than 30 elevators in normal working condition for 7 consecutive days, and is adjusted to the value corresponding to a missed detection rate of less than 5% through backtracking verification; the high-risk threshold is 0.7, the low-risk threshold is 0.4, and the decision threshold adjustment range is 0.05 to 0.10.

[0027] This invention utilizes sparse representation theory to construct an overcomplete dictionary and uses a matching pursuit algorithm to sparsely reconstruct the impact component from a mixed vibration signal. It combines velocity integration to convert the time-domain signal into a spatial-domain signal, and then uses a guide rail spatial spectrum estimation periodic impact positioning algorithm to achieve high-resolution spatial positioning of multiple impact excitation sources. Furthermore, it continuously quantifies and tracks the degradation trend of defects through degradation rate evaluation values. This invention solves the technical problem of being unable to accurately locate the spatial position of elevator guide rail defects and evaluate their degradation trend in real time under conditions of multiple concurrent impact sources and insufficient signal-to-noise ratio.

[0028] Sparse representation theory utilizes the sparse coefficients of impact signals in an overcomplete dictionary domain, naturally separating impact and noise components in the coefficient domain, fundamentally overcoming the feature aliasing problem caused by time-domain coupling. Spatial domain transformation eliminates the frequency modulation effect caused by variable speed operation, restoring the spatial periodicity of each impact excitation source to a fixed spatial frequency, providing a physically meaningful input for subsequent spatial spectrum estimation. Based on the orthogonality principle of signal and noise subspaces, the multi-signal classification spatial spectrum function can still estimate the spatial coordinates of each excitation source at high resolution even under conditions of multiple sources coexisting and low signal-to-noise ratio, breaking through the resolution bottleneck of traditional frequency domain methods in multi-source scenarios. The degradation rate assessment value, through normalization of the slope of the linear regression of the spectral peak amplitude sequence within a continuous operating cycle, continuously quantifies the degradation trend of early weak defects, compensating for the inability of a single detection to reflect the dynamic evolution of defects. In summary, this invention solves the technical problem mentioned in the background art of being unable to accurately locate the spatial position of elevator guide rail defects and assess their degradation trend in real time under conditions of multiple concurrent impact sources and insufficient signal-to-noise ratio. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method of the present invention.

[0030] Figure 2 This is a graph of the spatial spectrum function for multiple signal classification in an embodiment of the present invention.

[0031] Figure 3 This is a graph showing the sequence of peak amplitudes of abnormal impact sources and the linear regression fitting curve in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0033] like Figure 1 The diagram shown is a flowchart of a method for monitoring abnormal vibration in an elevator car provided by the present invention. This method includes the following steps:

[0034] S01. Construct an overcomplete dictionary of the impact signal using sparse representation theory, sparsely reconstruct the impact component from the mixed vibration signal using the matching pursuit algorithm, and locate the optimal filtering frequency band by combining the spectral kurtosis map.

[0035] S02. Convert the impact components obtained in step S01 into spatial domain signals based on the car speed integral results, estimate the spatial position coordinates of each impact excitation source using the guide rail spatial spectrum estimation periodic impact positioning algorithm, and output the degradation rate evaluation value.

[0036] S03. For all online elevators in the elevator group, construct a sample covariance matrix of the multidimensional vibration feature matrix within the same time window, and use the random matrix theory to derive a collaborative monitoring algorithm for the vibration state of the elevator group to calculate the overall abnormal index of the elevator group and locate the abnormal elevator number.

[0037] S04. Input the impact component characteristics obtained in step S01, the spatial location coordinates obtained in step S02, and the degradation rate evaluation value into the physical constraint embedding time-series reasoning model, and output the fault category and fault severity score.

[0038] S05. Using a lock-free circular buffer queue and a task directed acyclic graph asynchronous parallel scheduling framework, the feature extraction task and model inference task in steps S01 to S04 are dynamically allocated to multi-core processors.

[0039] S06. Based on the fault severity score output in step S04, the spatial location coordinates and degradation rate assessment value output in step S02, calculate the comprehensive risk index through the fault severity score adaptive adjustment function, and generate a maintenance decision report.

[0040] Among them, sparse representation theory refers to the signal decomposition theory that represents a signal as a linear combination of a few atoms in an overcomplete dictionary. The number of atoms in the overcomplete dictionary is much larger than the signal dimension, so that the impulse signal has sparse coefficients in the dictionary domain, thereby achieving effective separation of impulse components and noise components in a high noise background.

[0041] The overcomplete dictionary refers to a dictionary matrix where the number of atoms exceeds the signal dimension. Each atom in the dictionary corresponds to an impact waveform template. The atom shape is designed based on the physical response characteristics of the elevator guide rail joint impact. The atom center frequency coverage range is 100–5000 Hz, and the atom duration coverage range is 1–20 ms. The specific parameters of the atom center frequency coverage range and the atom duration coverage range are obtained by collecting measured impact signals of different elevator models under different speed conditions, performing time-frequency analysis on the impact waveforms, and statistically analyzing the distribution range of waveform parameters.

[0042] Among them, the matching pursuit algorithm is a greedy iterative decomposition algorithm. In each iteration, the atom with the largest inner product with the current residual signal is selected from the overcomplete dictionary, its coefficient is estimated and subtracted from the residual, and the iteration is repeated until the residual energy is lower than the preset energy threshold. The preset energy threshold is determined by taking 1.2 to 1.5 times the average noise energy after conducting multiple statistical experiments on pure noise signals under no-load conditions.

[0043] Among them, the spectral kurtosis map refers to a two-dimensional spectrum with the filter center frequency and bandwidth as the horizontal and vertical axes, and the kurtosis value of the filtered signal as the color depth. The frequency band with the highest kurtosis value is the optimal filter frequency band.

[0044] Among them, spatial domain signal refers to the signal based on the distance traveled by the car. acceleration signal of independent variable The conversion formula is expressed as follows: ;in For the speed of the car, For integration time variable, For the current moment, the conversion eliminates the frequency modulation effect caused by variable speed operation.

[0045] The principle of the guide rail spatial spectrum estimation periodic impact positioning algorithm is as follows: Constructing The spatial domain signal observation matrix is ​​divided into several spatial windows. The sample covariance matrix is ​​calculated and eigenvalue decomposition is performed on the observation matrix. Based on the orthogonality principle between the signal subspace and the noise subspace, a MUSIC spatial spectrum function is constructed. Spatial frequencies are scanned and spectral peak positions are searched to accurately estimate the spatial coordinates of each impact excitation source. The formula for the MUSIC spatial spectrum function is as follows: ;in For spatial frequency, For spatial orientation vectors, The noise subspace matrix is ​​represented by the superscript. This indicates the conjugate transpose. The spatial spectral peak positions are matched with the designed spacing of the guide rail joints using residuals. Spectral peaks with residuals exceeding the residual matching threshold are identified as abnormal impact sources. The residual matching threshold ranges from 5 to 15 cm. The specific value is determined by conducting on-site measurements on elevators with different guide rail installation accuracy levels, statistically analyzing the maximum deviation of the guide rail joint installation position, and multiplying it by a safety factor of 1.5 to 2.0. The degradation rate assessment value is the normalized result of the slope obtained after linear regression of the spatial spectral peak amplitude sequence of each abnormal impact source over 30 to 90 consecutive operating cycles. The normalization method is to divide the original slope by the maximum value of the slope of historical normal samples; a larger value indicates a faster degradation rate of the corresponding guide rail defect. The technical benefits of the guide rail spatial spectrum estimation periodic impact location algorithm are as follows: after converting the time domain signal into a spatial domain signal, the spatial periodicity of the impact excitation can be restored to a fixed spatial frequency. The MUSIC algorithm uses the subspace orthogonality to separate the signal subspace from the noise subspace, so that even under the condition that multiple impact sources exist at the same time and the signal-to-noise ratio is low, the spatial coordinates of each impact excitation source can be located at high resolution. The degradation rate evaluation value can be used to continuously track early weak defects, providing maintenance personnel with accurate spatial positioning basis.

[0046] The principle of the collaborative monitoring algorithm for elevator group vibration state derived from random matrix theory is as follows: Assume the elevator group has a total of Each elevator collects data. 3D vibration characteristics, the number of sampling points within the time window is The construction dimension is The multidimensional vibration characteristic matrix is ​​used to calculate the sample covariance matrix. Based on the Marchenko-Pastur theorem, under the assumption that all elevators are operating under normal conditions, the theoretical boundary formula for the eigenvalues ​​of the sample covariance matrix is ​​as follows: ;in This is the noise variance estimate, determined by the mean of the sample variance of the baseline data under normal operating conditions. The baseline data is calculated by continuously collecting vibration data for each elevator under rated operating conditions for no less than 7 days. Overall abnormality index of the elevator group. The formula is expressed as follows: ;in It represents the largest eigenvalue of the measured sample covariance matrix. Timely determination of group abnormalities Location analysis is triggered when the overall elevator group anomaly index alarm threshold is exceeded. The overall elevator group anomaly index alarm threshold is determined as follows: Vibration characteristic data of no less than 30 elevators in normal operating condition are collected for 7 consecutive days. The overall elevator group anomaly index is calculated for each time window. The distribution of the obtained overall elevator group anomaly index is statistically analyzed, and the 99th percentile is taken as the initial alarm threshold. Then, the threshold is adjusted to the final value corresponding to a missed detection rate of less than 5% through backtracking verification of known fault cases. The amplitude of each component in the feature vector corresponding to the maximum eigenvalue exceeding the boundary reflects the contribution weight of each elevator to the group anomaly. The elevator number with the highest contribution weight is the abnormal elevator number. The technical benefits of the elevator group vibration state collaborative monitoring algorithm derived from random matrix theory are as follows: In the early stage of a single elevator, the vibration characteristics change weakly, and the signal-to-noise ratio is insufficient to support individual diagnosis. However, the collaborative monitoring of elevator groups, by aggregating the multidimensional vibration characteristics of multiple elevators, makes the statistical properties of the sample covariance matrix highly sensitive to group anomalies. Even if the feature offset of a single elevator is lower than the detection threshold of individual detection, the group effect will cause the maximum eigenvalue to deviate significantly from the theoretical boundary, thereby achieving collaborative enhanced detection under the condition of insufficient signal-to-noise ratio of a single elevator. Furthermore, the feature vector component analysis further realizes the accurate location of the abnormal elevator number.

[0047] The specific structure of the physical constraint-embedded temporal reasoning model is as follows: The input layer receives a multi-sensor vibration temporal feature sequence segmented by a sliding window. The sliding window length is 0.5–2.0 s, and the window step size is 0.1–0.5 s. The sliding window length is determined by taking 2–3 times the average duration of the impact event at different elevator operating speeds (0.5–4.0 m / s). The window step size is 10%–25% of the sliding window length. The front end uses a bidirectional long short-term memory layer with a hidden layer dimension of 128–256 and 2–3 layers to extract short-term local temporal dependencies and capture the contextual relationships of the guide rail impact transient events. The middle layer introduces a sparse self-attention Transformer module with 4–8 attention heads and a feedforward network dimension of 256–512. The window size of the local window attention mechanism covers 16–64 time steps to handle long temporal dependencies. The physical constraint embedding mechanism encodes prior physical parameters such as guide rail installation spacing and traction rope pitch into a position offset matrix and injects attention calculations. The formula for constructing the position offset matrix is ​​as follows: ;in The guide rail installation spacing, ranging from 3000 to 5000 mm, is determined from the installation drawings. Given the current car speed, For the corresponding time interval, The sampling time interval, and For time step index, For the position offset matrix, the first Line 1 The column elements, the position bias matrix, ensure that the attention distribution conforms to the spatial periodicity of the elevator mechanical structure. The network incorporates an internal iterative refinement module that performs three self-iterative corrections on the preliminary diagnosis results. Each iteration concatenates the fault probability vector from the previous round with the current features and re-infers the information, gradually eliminating the uncertainty of the preliminary diagnosis. The output layer simultaneously outputs the fault category and fault severity score, with the severity score being a continuous value between 0 and 1. Regarding the resource allocation of neuron activation functions, the activation function type of each gate unit in the bidirectional long short-term memory layer is dynamically selected based on the statistical kurtosis value of the current input feature. When the statistical kurtosis value is high, an activation function with stronger nonlinear expressive power is switched to adapt to the characteristics of the impact signal. Regarding the allocation of inter-layer gradient resources, the feedforward network dimension of each sparse self-attention Transformer module is dynamically scaled based on the inter-layer gradient norm ratio. When the gradient norm ratio exceeds the preset upper limit, the feedforward network dimension is compressed to release computational resources. The preset upper limit of the gradient norm ratio is determined by taking the 95th percentile after statistically analyzing the distribution of the gradient norm ratios of each layer during normal training on the training set. Regarding data loop resource allocation, the loop step size of the sliding window is dynamically adjusted based on the current input queue backlog depth; as the backlog depth increases, the loop step size is increased to reduce computational load. For memory allocation, the batch size of the batch normalization layer is dynamically configured based on the currently available memory capacity. For CUDA stream allocation, the forward and backward propagation of the bidirectional long short-term memory layer are allocated to different CUDA streams for parallel execution to reduce waiting latency. The CUDA thread block size is adaptively selected based on the current matrix multiplication scale to maximize computational unit utilization. The technical benefits of the physical constraint embedding temporal reasoning model are as follows: the bidirectional long short-term memory layer utilizes the contextual information of the impact event for simultaneous modeling, avoiding the loss of transient features caused by unidirectional modeling; the sparse self-attention Transformer module reduces computational load while retaining the ability to model long-term temporal dependencies through a local window attention mechanism; the physical constraint embedding mechanism encodes prior knowledge such as guide rail installation spacing into a position bias matrix, ensuring that the attention distribution is consistent with the spatial periodicity of the elevator mechanical structure, reducing the model's attention to irrelevant features, and improving the accuracy of early fault identification; the internal iterative refinement module gradually eliminates the uncertainty of the preliminary diagnosis through multiple rounds of self-iterative correction, improving the reliability of the output results.

[0048] The steps for establishing the training dataset for the physical constraint embedded temporal inference model specifically include: collecting triaxial acceleration vibration data from no fewer than 20 elevators under four operating conditions: normal operation, guide rail joint impact, guide shoe wear, and traction rope vibration. The data collection time for each elevator under each operating condition should be no less than 72 hours, and the data sampling frequency should be 1000–5000 Hz. The collected data should be manually labeled, including fault category labels and fault severity scores. The fault severity scores should be comprehensively evaluated by experienced maintenance engineers based on vibration values ​​and equipment inspection results. The labeled data should be divided into training, validation, and test sets in a 7:2:1 ratio. The training set data should be augmented using methods including random time shifting, random amplitude scaling, and random noise superposition. The random amplitude scaling coefficient should range from 0.8 to 1.2, expanding the training set size to 3–5 times the original size.

[0049] The specific steps for training the physical constraint-embedded temporal inference model include: using a weighted combination of cross-entropy loss and mean squared error loss as the total loss function, where cross-entropy loss is used for fault category classification and mean squared error loss is used for fault severity rating regression; the weight ratio of the two is determined through grid search on the validation set; the optimizer uses AdamW, with an initial learning rate range of [missing information]. ~ The training process employs a cosine annealing learning rate decay strategy; the number of training rounds ranges from 50 to 200; the early stopping strategy is monitored by the validation set comprehensive loss, and training is terminated early if there is no improvement after 10 consecutive rounds; after training is completed, the performance of the physical constraint embedded temporal reasoning model is evaluated on the test set.

[0050] The adaptive adjustment function for fault severity scoring is designed as follows: based on the fault severity score output in step S04... The degradation rate evaluation value output in step S02 And the overall abnormality index of the elevator group output in step S03 Calculate the comprehensive risk index The formula is expressed as follows: ;in , , Let be the weighting coefficient, satisfying , , , All are dimensionless normalized values. The values ​​range from 0 to 1. The three weighting coefficients are determined by normalizing the Pearson correlation coefficients between each indicator in historical maintenance records and the actual probability of failure, and are gradually optimized through iterative experiments. When physical constraints are embedded in the temporal reasoning model, the decision threshold during the reasoning stage is lowered by 0.05–0.10 to improve the fault recall rate; when At that time, the decision threshold of the physical constraint embedded temporal reasoning model during the reasoning stage remains unchanged; when To reduce the false positive rate, the decision threshold for embedding physical constraints into the inference stage of the temporal reasoning model is increased by 0.05–0.10. The specific value of the decision threshold adjustment is determined through a trade-off experiment on the validation set between precision and recall.

[0051] Among them, the bidirectional long short-term memory layer refers to a recurrent network layer that simultaneously calculates the forward and backward hidden states at each time step and concatenates them as the output representation, so as to utilize both past and future information of the time series.

[0052] Among them, the sparse self-attention Transformer module refers to the Transformer encoder module that introduces local window constraints on the basis of the standard multi-head self-attention mechanism. Each time step only calculates the attention weight within its local window, which reduces the computational cost from being proportional to the square of the sequence length to being linearly proportional to the sequence length.

[0053] Among them, the lock-free circular buffer queue refers to a fixed-capacity circular queue that is thread-safe using atomic operations instead of mutexes. Multiple collection threads concurrently write vibration data by atomically incrementing the write pointer, and the consumption threads read data by atomically incrementing the read pointer, thus avoiding thread blocking and context switching overhead caused by mutexes.

[0054] The task directed acyclic graph refers to an acyclic directed graph constructed with computational tasks as nodes and data dependencies as directed edges. The scheduling framework determines the task execution order based on the topological sorting of the directed graph and simultaneously distributes parallel nodes without data dependencies to multi-core processors for execution, eliminating the delay caused by sequential waiting in the pipeline bottleneck stage.

[0055] Among them, the comprehensive risk index It is scored based on the severity of the fault. Deterioration rate assessment value Overall abnormality index of elevator group The dimensionless score obtained by weighted calculation, with a value range of 0 to 1, is used in step S06 to drive the adaptive adjustment function of the fault severity score to dynamically correct the decision threshold of the physical constraint embedded temporal reasoning model, and serves as the basis for the maintenance priority ranking in the maintenance decision report.

[0056] Optionally, the present invention also provides a method for implementing an elevator car vibration anomaly monitoring system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions are used to execute the above-described method when the computer is run.

[0057] The specific implementation of step S01 is as follows: First, based on the physical response characteristics of the impact signal of the elevator guide rail joint, a complete dictionary of impact waveform atoms is designed. The atom center frequency covers 100–5000 Hz, and the atom duration covers 1–20 ms. The above parameter ranges are determined by performing time-frequency analysis on the measured impact signals collected from different elevator models under different speed conditions and statistically analyzing the distribution range of waveform parameters, ensuring that the dictionary atoms can cover the time-frequency characteristics of various impact waveforms in actual engineering. Second, a matching pursuit algorithm is used to perform sparse decomposition on the mixed vibration signal collected by multiple sensors. In each iteration, the inner product of the current residual and all atoms in the dictionary is calculated. The atom with the largest absolute value of the inner product is selected, its sparsity coefficient is estimated, and then subtracted from the residual. This process is repeated until the residual energy is lower than a preset energy threshold. The preset energy threshold is taken as 1.2–1.5 times the average energy of the pure noise signal under no-load conditions. Subsequently, the optimal filtering frequency band of the separated impulse component was located using the spectral kurtosis map. The spectral kurtosis map uses the center frequency and bandwidth of the filter as the horizontal and vertical axes, and the kurtosis value of the filtered signal as the color depth. The frequency band with the highest kurtosis value was selected as the optimal filtering frequency band, and the impulse component was bandpass filtered to further improve the signal-to-noise ratio of the impulse component.

[0058] The specific implementation of step S02 is as follows: First, the speed signal collected by the car speed sensor is integrated over time to convert the impact component from the time domain into a spatial domain signal with the travel distance as the independent variable, eliminating the frequency modulation effect caused by variable speed operation, so that the impact events of each impact excitation source exhibit equidistant spatial periodicity in the spatial domain. Second, a system containing... The spatial domain signal observation matrix of each spatial window is used to calculate the sample covariance matrix and perform eigenvalue decomposition. Based on the orthogonality principle between the signal subspace and the noise subspace, a multi-signal classification spatial spectrum function is constructed. The system scans spatial frequencies and searches for spectral peak positions to accurately estimate the spatial coordinates of each impact excitation source. Residual matching is performed between the spatial spectral peak positions and the designed spacing of the guide rail joints. Spectral peaks with residuals exceeding a residual matching threshold (range 5–15 cm) are identified as abnormal impact sources. The degradation rate assessment value is obtained by performing linear regression on the spatial spectral peak amplitude sequences of each abnormal impact source over 30–90 consecutive operating cycles. The resulting slope is normalized by dividing by the maximum slope value of historical normal samples before being output.

[0059] The specific implementation of step S03 is as follows: For all online elevators in the elevator group, extract the multi-dimensional vibration features of each elevator within the same time window, and construct a dimension of... The multidimensional vibration characteristic matrix, where For the feature dimension of a single elevator, This refers to the number of elevators. This represents the number of sampling points within the time window. The sample covariance matrix of the multidimensional vibration characteristic matrix is ​​calculated, and the theoretical boundary value of the eigenvalues ​​under normal operating conditions is calculated based on the Marchinko-Pastul law. The noise variance The anomaly index is determined by the mean variance of baseline vibration data collected over at least 7 consecutive days under rated operating conditions for each elevator. The overall anomaly index of the elevator group is then calculated. , Timely determination of group abnormalities When the alarm threshold is exceeded, the eigenvector corresponding to the largest eigenvalue exceeding the boundary is analyzed, and the amplitude of each component is extracted as the contribution weight of each elevator to the group anomaly. The elevator number with the highest contribution weight is the abnormal elevator number. The alarm threshold is taken as the 99th percentile of the overall anomaly index distribution of the elevator group obtained from data of no less than 30 elevators in normal operating condition for 7 consecutive days, and is adjusted to a final value with a false negative rate of less than 5% through backtracking verification.

[0060] The specific implementation of step S04 is as follows: The input to the physical constraint-embedded temporal inference model is a multi-sensor vibration temporal feature sequence segmented by a sliding window, with a window length of 0.5–2.0 s and a window step size of 0.1–0.5 s. The model front end is a bidirectional long short-term memory layer with a hidden layer dimension of 128–256 and 2–3 layers, used to extract short-term local temporal dependencies and capture the contextual information of the impact transient event. The middle part is a sparse self-attention converter module with 4–8 attention heads, a feedforward network dimension of 256–512, and a local window attention covering 16–64 time steps, used to process long-term temporal dependencies. The physical constraint embedding mechanism is based on the guide rail installation spacing. With the current car speed Calculate the impact time interval , and then A position bias matrix is ​​constructed and attention calculations are incorporated to ensure that the attention distribution conforms to the spatial periodicity of the elevator's mechanical structure. An internal iterative refinement module is included in the network to perform three self-iterative corrections on the preliminary diagnostic results. Each iteration concatenates the fault probability vector from the previous round with the current features and re-infers the results, gradually eliminating diagnostic uncertainty. The output layer simultaneously outputs the fault category and a fault severity score ranging from 0 to 1.

[0061] The specific implementation of step S05 is as follows: A lock-free circular buffer queue utilizes atomic operations to achieve thread safety. Multiple acquisition threads concurrently write vibration data using atomically incrementing write pointers, while consumption threads read data using atomically incrementing read pointers, avoiding thread blocking and context switching overhead caused by mutex locks. The task directed acyclic graph uses the computational tasks from steps S01 to S04 as nodes and data dependencies as directed edges. The scheduling framework determines the task execution order based on topological sorting, simultaneously distributing parallel nodes without data dependencies to multi-core processors, eliminating latency caused by sequential waiting at pipeline bottleneck stages. The bidirectional long short-term memory layer's forward and backward propagation are allocated to different computational acceleration unified device architectures for parallel execution. The matrix multiplication scale adaptively selects the thread block size to maximize computational unit utilization. The sliding window loop step size is dynamically adjusted based on the input queue backlog depth, and the batch size of the batch normalization layer is dynamically configured based on available video memory capacity.

[0062] The specific implementation of step S06 is as follows: based on the fault severity score output in step S04. The degradation rate evaluation value output in step S02 And the overall abnormality index of the elevator group output in step S03 All three are dimensionless values ​​after normalization, obtained through... Calculate the comprehensive risk index Weighting coefficient , , satisfy The value was determined by normalizing each indicator with the Pearson correlation coefficient of the actual failure probability in historical maintenance records. When physical constraints are embedded in the time-series inference model, the decision threshold is lowered by 0.05–0.10 to improve fault recall; when When the decision threshold remains unchanged; when The decision threshold is adjusted upwards by 0.05 to 0.10 to reduce the false alarm rate. Finally, the maintenance priority of each elevator is ranked according to the comprehensive risk index, and a maintenance decision report is generated that includes fault type, spatial location coordinates, deterioration rate assessment value and maintenance priority.

[0063] It should be noted that the key technologies of this invention include: impact component separation achieved through sparse representation and matching pursuit algorithms, which essentially utilizes the sparsity of the impact signal in the overcomplete dictionary domain to naturally decouple the impact from noise in the coefficient domain, overcoming the feature extraction failure problem caused by time-domain aliasing; multi-source spatial localization achieved through spatial domain transformation and multiple signal classification algorithms, which transforms the time-domain non-stationarity caused by variable speed modulation into the fixed periodicity of the spatial domain, and then utilizes the orthogonality of subspaces to locate each excitation source with high resolution, overcoming the deficiency of insufficient resolution of traditional frequency domain methods in multi-source low signal-to-noise ratio scenarios; and an elevator group collaborative monitoring algorithm derived from random matrix theory, which, by aggregating the covariance matrix statistical characteristics of multi-dimensional features of multiple elevators, significantly amplifies weak anomalies when the signal-to-noise ratio of a single elevator is insufficient under the group effect. The synergistic effect of the three is that impact component separation provides high-quality input for the latter two, spatial localization quantifies the degradation position of a single device, and group collaboration realizes early anomaly detection across devices. The three links are connected to form a complete perception link from signal-level separation to device-level localization to group-level early warning, and the overall detection sensitivity and localization accuracy are better than any single method.

[0064] It should be noted that when the elevator group is large and the early fault characteristics of each elevator are extremely weak, the vibration characteristic offset of a single elevator is often lower than the detection threshold of any single detection method, making the early fault completely invisible in the single-elevator diagnosis stage. The reason for this technical problem is that in the early fault stage, the guide rail defect has not yet developed to the point of causing significant changes in vibration value, and the characteristic signal-to-noise ratio of a single elevator is extremely low. Traditional single-elevator diagnosis methods rely on triggering a single elevator's characteristic exceeding the threshold, and cannot utilize the potential statistical correlation between multiple elevators to enhance detection sensitivity. A systematic blind zone exists in the range where the characteristic offset is lower than the detection threshold. The usual solution to this technical problem is to increase the sensitivity of a single sensor or increase the number of sensors to improve the single-elevator signal-to-noise ratio. However, this is limited by the upper limit of the sensor's physical sensitivity and engineering installation conditions, making it difficult to increase the single-elevator signal-to-noise ratio to a level sufficient to support early fault detection in actual deployment, and the cost of modifying existing systems is extremely high. This invention effectively solves this technical problem. Step S03 constructs a dimension... The multidimensional vibration feature matrix jointly models the multidimensional vibration features of all online elevators in the elevator group, so that the maximum eigenvalue of the sample covariance matrix simultaneously responds to the feature changes of all elevators in the group. Even if the feature offset of a single elevator is lower than the individual detection threshold, the covariance structure changes of multiple elevators will be significantly reflected through the deviation of the maximum eigenvalue of the matrix from the theoretical boundary of the Marchinko-Pastul theory. This enables the collaborative enhancement detection of early weak anomalies at the group level. Furthermore, the feature vector component analysis further traces the group anomalies precisely to the specific abnormal elevator number, which not only solves the problem of insufficient sensitivity in early fault detection, but also avoids the inefficiency of traditional methods that check one by one in multiple devices.

[0065] Specifically, the principle of this invention is:

[0066] The present invention can solve the above-mentioned technical problems. The fundamental reason is that each link of the solution forms a complete causal chain at the physical mechanism level.

[0067] First, the elevator guide rail impact signal physically exhibits short-time transient and band-limited localized waveform characteristics, which highly match the impact atoms designed according to the physical response characteristics in the overcomplete dictionary. Sparse representation theory utilizes this physical prior to decompose the mixed vibration signal into a sparse linear combination of dictionary atoms, making the impact component exhibit concentrated non-zero coefficients in the coefficient domain, while the coefficients of the broadband noise component are dispersed across all atoms. The matching pursuit algorithm iteratively extracts the atom with the largest inner product with the residual through greedy iteration, prioritizing the capture of the impact atom with the most concentrated energy in each iteration, thereby achieving effective separation of the impact component and the noise component. This is the prerequisite for all subsequent analysis steps to succeed.

[0068] Secondly, variable speed operation causes the time interval of impact events in the time domain to vary with speed fluctuations, thus masking the spatial periodicity of the impacts in the time domain. By integrating the car speed and mapping the time axis to the travel distance axis, the impact events in the spatial domain are restored to periodic impacts with equal intervals, whose spatial frequency directly corresponds to the reciprocal of the guide rail joint installation spacing. This transformation mathematically eliminates the frequency modulation caused by speed changes, allowing the spatial periodicity to be accurately represented in the spatial domain. This is the physical basis for spatial spectrum estimation to accurately locate the coordinates of each excitation source.

[0069] Furthermore, the multiple signal classification algorithm is based on the strict linear algebraic property that the signal subspace and noise subspace are orthogonal. When multiple periodic impulse sources exist in a spatial domain signal, their corresponding spatial frequency components span the signal subspace, while the orthogonal noise subspace does not contain these frequency components. The spatial spectral function of the multiple signal classification algorithm approaches infinity at the spatial frequency corresponding to the signal subspace, producing sharp spectral peaks, thus locating each excitation source with a resolution higher than that of the traditional Fourier transform. Even under low signal-to-noise ratio conditions, the subspace orthogonality can be robustly estimated through the eigenvalue decomposition of the sample covariance matrix, which is the essential reason why this method outperforms traditional methods in multi-source, low signal-to-noise ratio scenarios.

[0070] Finally, the degradation rate assessment value is normalized by the linear regression slope of the spectral peak amplitude within a continuous operating cycle, transforming a single amplitude snapshot into a trend quantity reflecting the dynamics of defect degradation. This allows early, weak defects to be identified through slope changes before their amplitude exceeds the absolute threshold, thus enabling continuous quantitative tracking of the degradation process. Overall, this invention possesses rigorous physical or mathematical foundations in each of the four stages—signal separation, coordinate system transformation, subspace decomposition, and trend modeling—which are interconnected and together constitute a complete technical path for solving the core technical problem.

[0071] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0072] The specific implementation method of step S01 is as follows.

[0073] Constructing an overcomplete dictionary using sparse representation theory ,in Represents the real number field. For signal dimension, The number of atoms and Each atom in the dictionary corresponds to an impact waveform template, with the atom center frequency covering 100–5000 Hz and the atom duration covering 1–20 ms. These parameters were obtained by collecting measured impact signals from different elevator models under different speed conditions, performing time-frequency analysis on the impact waveforms, and statistically analyzing the distribution range of the waveform parameters. (Mixed vibration signals) It can be represented as:

[0074] ;

[0075] In the formula, It is a sparse coefficient vector. This represents the noise component vector. The iterative process of the matching pursuit algorithm is as follows: Initialize the residual. index set Number of iterations In each iteration, the atom index with the largest inner product with the current residual is selected:

[0076] ;

[0077] In the formula, For dictionary No. Atoms in the column For the first The residual vector of the next iteration For the first The next iteration selects the atomic indices; the index set is then updated. Estimate the coefficients and update the residuals:

[0078] ;

[0079] ;

[0080] In the formula, For index set Sub-dictionaries composed of corresponding columns The corresponding sparse coefficient subvector; the iteration termination condition is... ,in To preset the energy threshold, it was determined by taking 1.2 to 1.5 times the average noise energy after multiple statistical experiments on the pure noise signal under no-load conditions; the impact component was reconstructed after the iteration. ,in For the final index set, This is the final sparse coefficient subvector. The reconstructed impact component vector. The spectral kurtosis plot is based on the filter center frequency. With bandwidth The x and y axes represent the kurtosis values ​​of the filtered signal. The frequency band with the highest kurtosis value, representing color depth, is the optimal filtering frequency band, used for further frequency band localization of the impulse component.

[0081] The specific implementation method of step S02 is as follows.

[0082] Impact component Based on the car speed Conversion of integration results into spatial domain signals spatial distance The calculation formula is expressed as follows:

[0083] ;

[0084] In the formula, for Car speed at any time, unit , For integration time variable, in units , For the current moment, the unit , The distance traveled by the car, in units The conversion eliminates the frequency modulation effect caused by variable speed operation. (Construction) Observation matrix of each spatial window ,in Number of sampling points within each spatial window; calculate the sample covariance matrix:

[0085] ;

[0086] In the formula, The spatial domain sample covariance matrix, with superscript... Indicates transpose; for Perform eigenvalue decomposition:

[0087] ;

[0088] In the formula, The eigenvector matrix, The matrix is ​​a diagonal matrix composed of eigenvalues; based on the principle of orthogonality between the signal subspace and the noise subspace, a multi-signal classification spatial spectrum function is constructed:

[0089] ;

[0090] In the formula, Spatial frequency, unit , For spatial orientation vectors, The noise subspace matrix is ​​represented by the superscript. Indicates conjugate transpose; scan The system searches for spectral peak positions to accurately estimate the spatial coordinates of each impact excitation source. The spatial spectral peak positions are then matched with the designed spacing of the guide rail joints using residuals. If the residual exceeds the residual matching threshold (5–15), the result is considered satisfactory. The spectral peaks were identified as an abnormal impact source. Deterioration rate assessment value. The slope of the normalized result obtained by linear regression of the amplitude sequence of the spatial spectrum peaks of each abnormal impact source within 30 to 90 consecutive operating cycles is calculated. Let the first... The spatial spectrum peak amplitude of the abnormal impact source corresponding to each operating cycle is: The running cycle number is ( , The total number of selected continuous operating cycles (ranging from 30 to 90) is expressed by the linear regression formula as follows:

[0091] ;

[0092] In the formula, For the first The linear regression fit value of the peak amplitude of the spectrum in each operating cycle The original regression slope, Intercept; degradation rate assessment value The normalization formula is expressed as follows:

[0093] ;

[0094] In the formula, This represents the maximum slope of the historical normal sample. and Same dimensions This is a dimensionless degradation rate assessment value; the larger the value, the faster the corresponding guide rail defect deteriorates.

[0095] The specific implementation method of step S03 is as follows.

[0096] Elevator group Each elevator collects data. 3D vibration characteristics, the number of sampling points within the time window is Construct a multidimensional vibration feature matrix Calculate the sample covariance matrix:

[0097] ;

[0098] In the formula, Let be the covariance matrix of the vibration characteristic samples of the elevator group. Based on the Marchenko-Pascal theorem, the theoretical boundary formula for the eigenvalues ​​of the sample covariance matrix is ​​as follows:

[0099] ;

[0100] In the formula, This is the noise variance estimate, determined by the mean of the sample variance of the baseline data under normal operating conditions. The baseline data is calculated by continuously collecting vibration data for each elevator under rated operating conditions for no less than 7 days. and Same dimensions; overall abnormality index of elevator group The formula is expressed as follows:

[0101] ;

[0102] In the formula, The measured sample covariance matrix The largest eigenvalue, and Same dimensions Dimensionless anomaly index Timely determination of group abnormalities When the alarm threshold is exceeded, a location analysis is triggered. The alarm threshold is obtained by collecting vibration characteristic data from no fewer than 30 elevators under normal operating conditions for 7 consecutive days and then statistically analyzing the data. The alarm threshold is distributed and the 99th percentile is used as the initial alarm threshold. Then, it is adjusted to the final value corresponding to a false negative rate of less than 5% through backtracking verification of known failure cases. The amplitude of each component in the corresponding feature vector reflects the contribution weight of each elevator to the group anomaly, and the elevator number with the highest contribution weight is the abnormal elevator number.

[0103] The specific implementation method of step S04 is as follows.

[0104] The input layer of the physical constraint-embedded temporal inference model receives multi-sensor vibration temporal feature sequences segmented by a sliding window, with a sliding window length ranging from 0.5 to 2.0. The window step size is 0.1 to 0.5. The front-end uses a bidirectional long short-term memory layer to extract short-term local temporal dependencies at each time step. Output Forward hidden state With backward hidden state splicing:

[0105] ;

[0106] In the formula, This represents a vector concatenation operation. , The unidirectional hidden layer dimension ranges from 128 to 256. A sparse self-attention transformer module is introduced in the middle, with 4 to 8 attention heads. The feedforward network dimension is 256 to 512. The local window attention mechanism has a window size covering 16 to 64 time steps. The formula for calculating the local window attention weight is as follows:

[0107] ;

[0108] In the formula, For querying the matrix, The key matrix, For value matrices, The dimension of the key vector. For the reason The position offset matrix is ​​constructed, and attention calculation is performed only within a local window; the physical constraint embedding mechanism determines the guide rail mounting spacing. (Value range: 3000-5000) (Originally obtained from the installation drawings) Encoded as a position offset matrix Its construction formula is expressed as follows:

[0109] ;

[0110] ;

[0111] In the formula, For the corresponding time interval of the guide rail joint, in units of , The current car speed, in units , unit , The sampling time interval is expressed in units of... , and For time step index, Position offset matrix No. Line 1 Column elements, The number of time steps is dimensionless. The time step is dimensionless. The dimensionless position bias matrix ensures that the attention distribution conforms to the spatial periodicity of the elevator's mechanical structure. The network incorporates an internal iterative refinement module that performs three self-iterative corrections on the initial diagnostic results. Each iteration concatenates the fault probability vector from the previous round with the current features before re-inferring the information. The output layer simultaneously outputs the fault category and fault severity score. , It is a continuous value between 0 and 1.

[0112] The specific implementation method of step S05 is as follows.

[0113] An asynchronous parallel scheduling framework using a lock-free circular buffer queue and a directed acyclic graph (DAG) is employed to dynamically allocate the feature extraction and model inference tasks in steps S01 to S04 to multi-core processors. The lock-free circular buffer queue utilizes atomic operations to implement a thread-safe, fixed-capacity circular queue. Multiple acquisition threads concurrently write vibration data by atomically incrementing the write pointer, while consumption threads read data by atomically incrementing the read pointer. The DAG uses computational tasks as nodes and data dependencies as directed edges. The scheduling framework determines the task execution order based on topological sorting, simultaneously distributing parallel nodes without data dependencies to multi-core processors for execution.

[0114] The specific implementation method of step S06 is as follows.

[0115] Scoring based on the severity of the fault Deterioration rate assessment value Overall abnormality index of elevator group The comprehensive risk index is calculated using an adaptive adjustment function based on the severity score of the fault. The formula is expressed as follows:

[0116] ;

[0117] In the formula, , , Let be the weighting coefficient, satisfying , , , All are normalized dimensionless values. The values ​​range from 0 to 1; the three weighting coefficients are determined by normalizing the Pearson correlation coefficients between each indicator in historical maintenance records and the actual probability of failure, and are gradually optimized through iterative experiments; when When the decision threshold during the model inference stage is lowered by 0.05–0.10, the failure recall rate can be improved; when When, keep the decision threshold unchanged; when In such cases, the decision threshold is raised by 0.05–0.10 to reduce the false alarm rate; the final decision is based on… The system generates maintenance decision reports and prioritizes maintenance for elevators with abnormal conditions.

[0118] To better understand and implement this invention, a specific application scenario of the invention is provided below as an example 2: To verify the effectiveness of the invention, technicians set up a test environment and selected a traction elevator in a building with a rated speed of 2.0 m / s. Under the condition that there were 3 known joint position deviation defects in the guide rail, vibration anomaly monitoring was performed sequentially according to steps S01 to S06 of the invention. An accelerometer was installed at the bottom of the car, with a sampling frequency of 2000 Hz, and the car speed was collected in real time by an encoder. In step S01, the atomic center frequency of the complete dictionary covered 100–5000 Hz, the atomic time width covered 1–20 ms, the number of atoms was 8 times the signal dimension, and the preset energy threshold of the matching pursuit algorithm was 1.3 times the average energy of the no-load noise. After 12 iterations, the residual energy of the matching pursuit algorithm dropped below the threshold, successfully separating 3 sets of impact components from the mixed vibration signal. The spectral kurtosis diagram showed that the optimal filter band center frequency was approximately 1200 Hz, and the bandwidth was approximately 400 Hz.

[0119] In step S02, the car speed integral result converts the time-domain signal into a spatial-domain signal, with a travel distance resolution of approximately 0.5 mm, effectively eliminating the frequency modulation effect caused by speed change. An observation matrix of 32 spatial windows is constructed. After eigenvalue decomposition of the sample covariance matrix, the spatial spectrum function of the multiple signal classification shows sharp spectral peaks at three spatial frequencies, with corresponding spatial coordinates shown in Table 1. Residual matching between the peak positions and the designed spacing of the guide rail joints is performed. The residuals of all three peaks exceed the residual matching threshold of 8 cm, identifying them as three abnormal impact sources, consistent with known defect locations. After linear regression of the peak amplitudes over 60 consecutive operating cycles, the degradation rate assessment values ​​for the three abnormal impact sources are 0.12, 0.31, and 0.08, respectively, with the second defect exhibiting the fastest degradation rate.

[0120] Table 1 Spatial location coordinates and degradation rate assessment values ​​of abnormal impact sources in guide rails

[0121]

[0122] In step S03, the building has 8 online elevators. 12-dimensional vibration features are extracted from each elevator, with 512 sampling points within the time window. The multidimensional vibration feature matrix has dimensions of 96×512. Baseline noise variance. The average value was calculated from 7 days of rated operating data for each elevator, and is 0.0034. The Marcinko-Pastur theoretical boundary value... The calculated result is 0.081, which is the largest eigenvalue of the measured sample covariance matrix. The overall abnormality index of the elevator group is 0.127. The value was 0.568, exceeding the alarm threshold of 0.21, triggering a location analysis. Feature vector component analysis showed that elevator number 3 had the highest contribution weight and was identified as an abnormal elevator number.

[0123] In step S04, the sliding window length of the physical constraint-embedded temporal reasoning AI model is 1.0s, the window step size is 0.2s, the bidirectional long short-term memory layer hidden layer dimension is 192, the number of layers is 2, the sparse self-attention converter module has 6 attention heads, the feedforward network dimension is 384, and the local window covers 32 time steps. (Guide rail installation spacing) The installation drawings indicate a distance of 4000mm. The current car speed is 2.0m / s. The position offset matrix time interval is... The time was 2.0s. After three internal iterations of refinement, the model output fault category as guide rail joint impact, with a fault severity score of 0.64.

[0124] In step S05, the lock-free circular buffer queue capacity is set to 4096 frames, 8 acquisition threads write concurrently, the consumption thread latency is stabilized below 1.2ms, the task directed acyclic graph contains a total of 14 nodes, among which the sparse decomposition task in step S01 and the matrix eigenvalue decomposition task in step S03 have no data dependency and are simultaneously distributed to 2 physical cores for parallel execution, the overall inference latency is shortened by about 38% compared with the sequential execution method.

[0125] In step S06, the normalized fault severity score The degradation rate assessment value is 0.64. The overall abnormality index of the elevator group is 0.31. After normalization, the result is 0.57. The three weighting coefficients were determined by normalizing the Pearson correlation coefficient. , , Comprehensive risk index The calculated result is 0.60, which is... Within the specified range, the decision threshold remains unchanged. The maintenance decision report output is shown in Table 2. Elevator No. 3 has the highest comprehensive risk index and is listed as the highest priority maintenance target. The second abnormal impact source has the highest degradation rate assessment value and is recommended for priority handling.

[0126] Table 2 Maintenance Decision Report Output Table

[0127]

[0128] like Figure 2 As shown, in this embodiment, the spatial spectrum function curve of the multiple signal classification in step S02 shows sharp spectral peaks at the spatial frequency positions corresponding to the three abnormal impact sources. There is a significant amplitude difference between the spectral peaks and the background noise layer, which intuitively verifies the high-resolution positioning capability of the guide rail spatial spectrum estimation periodic impact positioning algorithm under multi-source low signal-to-noise ratio conditions. Figure 3 As shown, in this embodiment, the spectral peak amplitude sequence of the second abnormal impact source for 60 consecutive operating cycles shows a linear upward trend. The linear regression fitting result is consistent with the measured data. The degradation rate assessment value of 0.31 is significantly higher than that of the first and third locations, reflecting that the second defect has the fastest degradation speed and needs to be prioritized for maintenance.

[0129] Compared to traditional time-domain statistical feature threshold judgment methods, this invention achieves several fundamental advancements at the technical principle level. Traditional methods directly perform statistical analysis on vibration signals in the time domain, failing to separate the contributions of each impact component from aliased signals, leading to feature aliasing in multi-source scenarios. This invention, however, achieves coefficient-level separation in the dictionary domain through sparse representation theory, ensuring that feature extraction of each impact component is no longer interfered with by other components. Traditional frequency-domain methods fail to reflect spatial periodicity under variable-speed conditions due to frequency modulation effects. This invention, however, maps the time-domain signal to a spatial-domain signal through velocity integration, eliminating modulation effects at their source and giving spatial positioning physical determinism. Traditional methods rely on triggering by exceeding the threshold of a single device, resulting in a systemic blind spot in the early stages of weak faults. This invention, however, utilizes the group statistical characteristics of the elevator group's covariance structure through random matrix theory, amplifying and detecting weak anomalies under synergistic effects. These three advancements are all rigorously based on physical mechanisms, rather than relying on empirical improvements brought about by parameter tuning.

[0130] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0131] Table 3. Variable Explanation Table (Part 1)

[0132]

[0133] Table 4. Variable Explanation Table (Part Two)

[0134]

[0135] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of monitoring elevator car vibration anomalies, characterized by, Includes the following steps: An overcomplete dictionary of impact signals is constructed using sparse representation theory. The impact components are sparsely reconstructed from the mixed vibration signals using a matching pursuit algorithm. The optimal filtering frequency band is located by combining the spectral kurtosis diagram. The impact components are converted into spatial domain signals based on the car speed integral result. The spatial position coordinates of each impact excitation source are estimated using the guide rail spatial spectrum estimation periodic impact positioning algorithm, and the degradation rate evaluation value is output. For all online elevators in the elevator group, a sample covariance matrix of the multidimensional vibration feature matrix is ​​constructed within the same time window. The elevator group vibration state collaborative monitoring algorithm is derived using random matrix theory to calculate the overall abnormal index of the elevator group and locate the abnormal elevator number. The impact component characteristics, spatial location coordinates, and degradation rate assessment values ​​are input into the physical constraints and embedded into the time-series inference model, which outputs the fault category and fault severity score. A lock-free circular buffer queue and a task directed acyclic graph asynchronous parallel scheduling framework are adopted to dynamically allocate feature extraction tasks and model inference tasks to multi-core processors; Based on the fault severity score, spatial location coordinates, and degradation rate assessment value, a comprehensive risk index is calculated using an adaptive adjustment function for the fault severity score, and a maintenance decision report is generated.

2. The method of monitoring elevator car vibration anomalies of claim 1, wherein, The sparse representation theory refers to the signal decomposition theory that represents a signal as a linear combination of a few atoms in an overcomplete dictionary. The number of atoms in the overcomplete dictionary is much greater than the signal dimension, so that the impact signal has sparse coefficients in the dictionary domain.

3. The method of monitoring elevator car vibration anomalies of claim 2, wherein, Each atom in the overcomplete dictionary corresponds to an impact waveform template. The atomic shape is designed based on the physical response characteristics of the impact of the elevator guide rail joint. The atomic center frequency coverage range and atomic time width coverage range are obtained by performing time-frequency analysis on the measured impact signal and statistically analyzing the waveform parameter distribution range.

4. The method for monitoring abnormal vibration of an elevator car according to claim 3, characterized in that, The matching pursuit algorithm refers to selecting the atom with the largest inner product with the current residual signal from the overcomplete dictionary in each iteration, estimating its coefficient and subtracting it from the residual, and iterating cyclically until the residual energy is lower than a preset energy threshold, which is a multiple of the average energy of the noise energy.

5. The method for monitoring abnormal vibration of an elevator car according to claim 4, characterized in that, The spatial domain signal refers to the acceleration signal with the car's travel distance as the independent variable. The travel distance is obtained by integrating the car's speed over time, and the frequency modulation effect caused by the variable speed operation is eliminated after the conversion.

6. The method for monitoring abnormal vibration of an elevator car according to claim 5, characterized in that, The guide rail spatial spectrum estimation periodic impact positioning algorithm specifically involves constructing a spatial domain signal observation matrix with multiple spatial windows, calculating the sample covariance matrix of the observation matrix and performing eigenvalue decomposition, constructing a multi-signal classification spatial spectrum function based on the orthogonality principle of the signal subspace and noise subspace, scanning spatial frequencies and searching for spectral peak positions.

7. The method for monitoring abnormal vibration of an elevator car according to claim 6, characterized in that, The degradation rate assessment value is the normalized result of the slope obtained by linear regression of the spatial spectrum peak amplitude sequence of each abnormal impact source in multiple consecutive operating cycles. The normalization method is to divide the original slope by the maximum value of the slope of the historical normal sample.

8. The method for monitoring abnormal vibration of an elevator car according to claim 7, characterized in that, The proposed algorithm for collaborative monitoring of elevator group vibration status, derived from the theory of random matrices, specifically determines the theoretical boundary of the eigenvalues ​​of the sample covariance matrix based on the Marchenko-Pastur law. The ratio of the deviation between the measured maximum eigenvalue and the theoretical boundary is used as the overall abnormality index of the elevator group. The abnormal elevator number is located by the component with the highest amplitude in the eigenvector corresponding to the maximum eigenvalue that exceeds the boundary.

9. The method for monitoring abnormal vibration of an elevator car according to claim 8, characterized in that, The physical constraint embedded temporal reasoning model includes a bidirectional long short-term memory layer for extracting short-term local temporal dependencies, a sparse self-attention converter module for processing long-term temporal dependencies, and a physical constraint embedding mechanism that encodes prior physical parameters such as guide rail installation spacing into position bias matrices and injects them into attention calculations. It also includes an internal iterative refinement module that performs multiple self-iterative corrections to the preliminary diagnostic results.

10. The method for monitoring abnormal vibration of an elevator car according to claim 9, characterized in that, The position offset matrix is ​​constructed by calculating the time interval corresponding to the impact event based on the guide rail installation spacing and the current car speed, then calculating the cosine value using the ratio of the time step index difference to the time interval as the independent variable, and using the obtained value as the element of the position offset matrix.

Images