A method for monitoring the quality of elevator operation in real time

Abstract

The application provides a method for monitoring elevator passenger quality in real time, and belongs to the technical field of elevator monitoring. The method is characterized in that a multi-channel sensor array is arranged at both ends of an elevator car, a guide shoe and a traction rope to collect multiple types of signals. After edge end preprocessing and cloud column storage, a guide rail geometric defect inverse problem vibration deconvolution tracing algorithm is used to convert time domain vibration signals into spatial domain signals, and inverse filtering is used to restore the guide rail defect distribution. A traction rope tension wave time domain reconstruction algorithm based on partial differential equation regularization is used to reconstruct the full rope tension space-time distribution. Multi-channel signals are input into a multi-scale time series convolution and state space fusion artificial intelligence model to output a passenger quality evaluation score. Finally, a comprehensive report is formed by fusing the outputs of each module, and a hierarchical alarm is triggered. The technical problem that elevator passenger quality monitoring cannot trace and locate vibration responses to guide rail spatial defects is solved.

Description

Technical Field

[0001] This invention belongs to the field of elevator monitoring technology, and more specifically, relates to a method for real-time monitoring of elevator ride quality. Background Technology

[0002] Elevator passenger quality monitoring is a crucial means of ensuring safe elevator operation. Current technologies for detecting geometric defects in the guide rails typically employ accelerometers to collect car vibration signals, combining this with Fast Fourier Transform (FFT) or Short-Time Fourier Transform (SFT) to extract frequency domain features, and then using threshold comparisons or machine learning classifiers to determine the presence of anomalies. These methods have been widely applied in elevator condition monitoring and can reflect the presence of vibration anomalies across the global frequency dimension.

[0003] However, the aforementioned spectral analysis methods have fundamental flaws. Frequency domain analysis outputs a statistical description of global frequency components; its underlying mechanism is mapping time-domain signals to the frequency axis, thus losing the correspondence between vibration events and the spatial position of the car. Geometric defects such as guide rail joint impacts and waviness distribution are spatially discrete, and their excitation effects exhibit strong time-varying characteristics depending on the car's position. A purely frequency domain approach cannot establish a causal mapping from the vibration response to the spatial coordinates of the excitation source.

[0004] In current elevator guide rail inspections, traditional spectral analysis methods lack the ability to solve inverse spatial domain problems. This means that even if vibration anomalies are detected, maintenance personnel cannot obtain precise spatial locations and quantitative estimates of the defects' amplitudes from the monitoring system. They must rely on manual, segment-by-segment inspections, which is inefficient and makes it difficult to quantify the severity of the defects. In other words, existing technologies suffer from the technical problem of failing to trace the source of vibration responses to spatial defects in the guide rails during elevator ride quality monitoring. Summary of the Invention

[0005] In view of this, the present invention provides a method for real-time monitoring of elevator ride quality, which can solve the technical problem in the prior art that elevator ride quality monitoring cannot trace the source of vibration response to defects in the guide rail space.

[0006] This invention is implemented as follows: This invention provides a method for real-time monitoring of elevator ride quality, comprising the following steps:

[0007] Multi-channel sensor arrays are arranged in the elevator car, guide shoes and both ends of the traction rope to collect vertical vibration acceleration, electrical parameters and rope end acceleration signals;

[0008] The acquired multi-channel signals are preprocessed at the edge, the raw data is stored in a circular buffer, and lossy compression is completed by wavelet coefficient threshold truncation method. The compressed data is then uploaded to the cloud columnar storage module.

[0009] The vertical vibration acceleration signal of the car is subjected to the vibration deconvolution source tracing algorithm of the inverse problem of guide rail geometric defects. The time domain signal is converted into a spatial domain signal according to the car running speed. The spatial distribution of guide rail defects is restored by inverse filtering of the spatial domain transfer function matrix, and the quantitative estimation results of defect location, amplitude and wavelength are output.

[0010] A time-domain reconstruction algorithm for traction rope tension waves based on partial differential equation regularization is applied to the acceleration signals at both ends of the traction rope and the tension signal at the compensation wheel. The spatiotemporal distribution of the tension waves of the entire rope is reconstructed in reverse, and a probability heatmap of the damage location is output.

[0011] The multi-channel signals are input into the passenger transport quality status perception model, which outputs the passenger transport quality assessment score. The learning rate parameter of the passenger transport quality status perception model is adjusted in real time by the dynamic learning rate adjustment function.

[0012] The quantitative estimation results of guide rail defects, the probability heat map of damage location, and the passenger transport quality assessment score are integrated to form a comprehensive elevator passenger transport quality report, which triggers a graded alarm.

[0013] The multi-channel sensor array consists of a vertical acceleration sensor, a lateral acceleration sensor, a current sensor, and a tension sensor, and its installation position is determined by a static calibration experiment.

[0014] Specifically, the lossy compression step involves performing a discrete wavelet transform on the signal, setting the wavelet coefficients with absolute values ​​below the compression threshold to zero, and retaining only the coefficients carrying the main energy for reconstruction, with a compression ratio not lower than the compression ratio threshold.

[0015] The cloud-based columnar storage module uses the Apache Parquet format, constructs a composite index based on time partitioning and device number, stores hot data on solid-state drives, and migrates cold data to object storage.

[0016] Specifically, the vibration deconvolution source tracing algorithm for the inverse problem of guide rail geometric defects involves performing deconvolution operations on the output spectrum using the pseudo-inverse matrix of the transfer function in the frequency-wavenumber domain. The regularization method is Tikkonov, and the regularization parameters are automatically selected by generalized cross-validation.

[0017] The guide shoe elastic parameters in the spatial domain transfer function matrix are pre-calibrated by a static compression experiment. The static compression experiment applies a gradient load to the guide shoe and records the compression displacement, and then fits the guide shoe elastic parameters.

[0018] Specifically, the time-domain reconstruction algorithm for the traction rope tension wave employs a time-stepping Crank-Nicolson implicit finite difference scheme to inversely reconstruct the spatiotemporal distribution of the entire rope tension wave within the framework of the wave equation. The regularization term is the spatial second derivative of the rope tension distribution. Norm.

[0019] The passenger transport quality status perception model is a multi-scale temporal convolution and state space fusion model, which is composed of an improved temporal convolutional network and a selective state space mechanism stacked alternately, and the end fusion limit learning mechanism forms a fast online fine-tuning layer.

[0020] Among them, the odd-numbered layers of the multi-scale temporal convolution and state-space fusion model introduce a physical channel grouping mechanism in the channel dimension, which assigns vibration, electrical and thermodynamic sensor signals to different convolutional channel groups, and the groups exchange information through a lightweight cross-channel attention module.

[0021] In the multi-scale temporal convolution and state-space fusion model, the even-numbered layers embed the rated speed and car mass as fixed structural constraint parameters in the continuous state transition matrix, and scale-aware gated residual connections are set between adjacent odd and even layers.

[0022] Among them, the multi-scale temporal convolution and state space fusion model adopts a three-step lightweight process of weight pruning, INT8 quantization and knowledge distillation when it is deployed, which compresses the inference model volume to below the compression volume threshold.

[0023] Specifically, the dynamic learning rate adjustment function is a comprehensive adjustment index calculated by weighting three indicators: the validation set loss reduction rate of the current batch, the mean gradient norm, and the variance of the passenger transport quality assessment score. The function then performs hold, decay, or reheat operations based on the interval in which the comprehensive adjustment index is located.

[0024] Specifically, when the overall adjustment index is lower than the reheat trigger threshold, the learning rate is reset to the reheat ratio range of the initial value, and the cosine annealing scheduling is restarted.

[0025] The training dataset for the multi-scale temporal convolution and state-space fusion model was collected from continuous multi-channel sensor data on no fewer than 50 elevators of different models in service. Data augmentation based on variational autoencoders was used to expand the dataset for a few types of fault samples.

[0026] The compression ratio threshold is 20:1, and the compression threshold is determined by iterative statistical analysis with the constraint that the root mean square error of the reconstructed signal does not exceed 5% of the root mean square amplitude of the original signal; the reheat ratio range is 10% to 30% of the initial learning rate; and the compression volume threshold is one-eighth of the original volume.

[0027] This invention converts the vertical vibration acceleration signal of the car into a spatial domain signal under a known operating speed, constructs the spatial domain transfer function matrix of the car-guide shoe-guide rail system, and uses a pseudo-inverse matrix inverse filtering method combined with Tikhonov regularization to perform deconvolution operation in the frequency-wavenumber domain. This restores the vibration response to a spatial distribution estimate of the guide rail defect, solving the technical problem that traditional spectrum analysis methods cannot trace the vibration response to spatial defects in the guide rail.

[0028] The core of this invention lies in introducing a vibration deconvolution source tracing algorithm for the inverse problem of guide rail geometric defects. This algorithm uses the known operating speed as a bridge to map the time-domain vibration signal to the spatial axis, giving the vibration characteristics originally dispersed in the time-frequency domain a clear spatial physical meaning. The spatial domain transfer function matrix characterizes the modulation effect of the guide shoe's elastic properties on excitation propagation, and the inverse filtering operation solves for the spatial distribution of the excitation source under this physical constraint, thereby realizing causal tracing from the vibration response to the spatial coordinates of the guide rail defect. In summary, this invention solves the technical problem mentioned in the background art that elevator passenger quality monitoring cannot trace the vibration response source to the spatial defect of the guide rail. Attached Figure Description

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

[0030] Figure 2 This is a reconstruction result of the spatial distribution of geometric defects in the guide rail.

[0031] Figure 3 A heatmap showing the spatiotemporal distribution of tension and probability of damage location in the traction rope.

[0032] Figure 4 A framework diagram for multi-physics domain fusion analysis of the comprehensive report on elevator ride quality. Detailed Implementation

[0033] 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.

[0034] like Figure 1 The diagram shown is a flowchart of a method for real-time monitoring of elevator ride quality provided by the present invention. This method includes the following steps:

[0035] S01. A multi-channel sensor array is arranged in the elevator car, guide shoe and both ends of the traction rope to collect vertical vibration acceleration, electrical parameters and rope end acceleration signals. The sampling rate is 1000-8000Hz and the resolution is 16bit.

[0036] S02. Perform edge preprocessing on the acquired multi-channel signals, store the original data in a circular buffer, and use the wavelet coefficient threshold truncation method to complete lossy compression with a compression ratio of not less than 20:1. After compression, upload the data to the cloud columnar storage module.

[0037] S03. Perform the vibration deconvolution source tracing algorithm for the inverse problem of guide rail geometric defects on the vertical vibration acceleration signal of the car. Convert the time domain signal into a spatial domain signal according to the car running speed. Restore the spatial distribution of guide rail defects through inverse filtering of the spatial domain transfer function matrix and output the quantitative estimation results of defect location, amplitude and wavelength.

[0038] S04. Perform a time-domain reconstruction algorithm for the tension wave of the traction rope based on the partial differential equation regularization on the acceleration signals at both ends of the traction rope and the tension signal at the compensation wheel. Reconstruct the spatiotemporal distribution of the tension wave of the whole rope in reverse under the framework of the wave equation and output a probability heat map of the damage location.

[0039] S05. Input the multi-channel signal into the passenger transport quality status perception model, output the passenger transport quality assessment score from the passenger transport quality status perception model, and adjust the learning rate parameter of the passenger transport quality status perception model in real time by the dynamic learning rate adjustment function.

[0040] S06. The quantitative estimation results of guide rail defects output by S03, the probability heat map of damage location output by S04, and the passenger transport quality assessment score output by S05 are integrated to form a comprehensive elevator passenger transport quality report and trigger a graded alarm.

[0041] The multi-channel sensor array consists of a vertical acceleration sensor, a lateral acceleration sensor, a current sensor, and a tension sensor. The installation position is determined by a static calibration experiment to ensure that the physical correspondence between the signals of each sensor is clear.

[0042] The circular buffer is a first-in-first-out circular memory structure. When new data is written to a full-capacity buffer, the oldest data is automatically overwritten. This is used to continuously store real-time streaming data under the condition of limited memory in edge computing units.

[0043] The wavelet coefficient threshold truncation method is as follows: after performing discrete wavelet transform on the signal, wavelet coefficients with absolute values ​​lower than a preset threshold are set to zero, and only coefficients carrying the main energy are retained for reconstruction. The preset threshold is determined by conducting multiple rounds of experiments on historical vibration data under different working conditions, with the constraint that the root mean square error of the reconstructed signal does not exceed 5% of the root mean square amplitude of the original signal, and after iterative statistical analysis.

[0044] The columnar storage module uses the Apache Parquet format, constructs a composite index based on time partitioning and device number, stores hot data in non-volatile solid-state drives, migrates cold data to object storage, and uses a pre-computed downsampling aggregation table to achieve multi-resolution retrieval.

[0045] The principle and specific implementation of the vibration deconvolution source tracing algorithm for the inverse problem of guide rail geometric defects are as follows: Taking the vertical vibration acceleration signal of the elevator car as input, the car-guide shoe-guide rail system is modeled as a linear time-varying system, and the excitation source is the geometric defects of the guide rail, including the misalignment of the joints and the waviness distribution. The first step of the algorithm is to convert the time-domain vibration signal into a spatial-domain signal using the known car running speed, that is, multiplying the time axis by the velocity to obtain the position axis; the second step is to construct the spatial-domain transfer function matrix of the system, with the guide shoe elastic parameters pre-calibrated by static compression experiments, and to perform inverse filtering on the spatial-domain signal in the frequency-wavenumber domain, using the pseudo-inverse matrix of the transfer function to perform deconvolution on the output spectrum to reconstruct the spatial distribution estimate of the guide rail defects; the third step is to perform continuous wavelet transform on the reconstructed defect distribution, extract the amplitude at each characteristic wavenumber, map it to the specific guide rail joint position, and output a quantitative estimate report of the defect position, amplitude, and wavelength. Regularization uses the Tikhonov method to suppress noise amplification in the ill-conditioned inverse problem, and the regularization parameters are automatically selected by generalized cross-validation. The static compression experiment involves applying a gradient load of 0.5–5 kN to the guide shoe and recording the compression displacement to obtain the guide shoe's elastic parameters. Each elevator is calibrated independently after installation. The technical advantages of the algorithm are as follows: Traditional spectrum analysis can only provide global frequency components and cannot locate faults to the spatial coordinates of the guide rail; this algorithm converts time-frequency domain information into spatial domain defect distribution, realizing causal tracing from vibration response to excitation source. This allows for a quantitative description of the location and amplitude of geometric defects such as guide rail joint impact and abnormal waviness, providing maintenance personnel with precise guide rail repair locations and significantly improving the spatial resolution and interpretability of fault location.

[0046] The principle and specific implementation of the traction rope tension wave time-domain reconstruction algorithm based on partial differential equation regularization are as follows: The traction rope is modeled as a one-dimensional string vibration system, whose tension distribution satisfies a wave equation with a damping term. Linear density, static tension, damping coefficient, and external excitation are the equation parameters. The algorithm uses the acceleration sensor measurements on the traction wheel side and the car side as boundary conditions, and the tension measurement at the compensation wheel as the internal constraint. It uses a time-stepping Crank-Nicolson implicit finite difference scheme to inversely reconstruct the spatiotemporal distribution of the entire rope tension wave within the framework of partial differential equations; the regularization term is the spatial second derivative of the rope tension distribution. The norm is used to suppress spatial noise amplification. By tracking the propagation speed and reflection behavior of tension waves in the rope, the location of local damage and strand slack in the rope are detected in real time, and a spatiotemporal distribution map of the whole rope tension and a probability heatmap of the damage location are output. The linear density, static tension, and damping coefficient are initially provided by the manufacturer's specifications and then corrected by fitting the no-load operation test data after installation. The correction method is to minimize the difference between the measured boundary acceleration and the simulated boundary acceleration. Norm. The Crank-Nicolson implicit finite difference scheme is a time-stepping numerical method that considers the spatial difference between the current time and the next time step simultaneously within each time step, ensuring that the numerical scheme remains unconditionally stable at any time step size and does not diverge as the time step size increases. The technical effect of the algorithm is as follows: As the core load-bearing component of the elevator, the tension distribution inside the traction rope cannot be obtained globally through single-point measurement; this algorithm uses boundary measurement values ​​to drive the forward and inverse solution of partial differential equations, reconstructing the tension distribution of the entire rope into a spatiotemporal two-dimensional map, making the local tension depression caused by rope strand slack and the wave velocity mutation caused by damage quantifiable detection features, providing a global physical basis for assessing the remaining life of the rope, and overcoming the fundamental defect that single-point tension sensors cannot detect mid-section damage.

[0047] The passenger transport quality status perception model is a multi-scale temporal convolutional and state-space fusion model. The specific structure of this model is as follows: the model uses an improved temporal convolutional network and a selective state-space mechanism, alternately stacked at multiple scale levels, with a total of eight stacked layers. Odd-numbered layers are dilated causal convolutional blocks, and even-numbered layers are improved selective state-space layers. The dilated causal convolutional blocks in odd-numbered layers introduce a physical channel grouping mechanism at the channel dimension, assigning vibration, electrical, and thermodynamic sensor signals to different convolutional channel groups. Within each group, convolutional kernel weights are shared, and information is exchanged between groups through a lightweight cross-channel attention module. The receptive field of each odd-numbered layer increases exponentially with the number of layers, covering multi-scale temporal patterns from short-term impacts to long-term trends. The receptive field range is determined experimentally; specifically, after statistical analysis of the typical duration of different fault types, the minimum receptive field required to cover 95% of the fault duration is selected as the upper limit of the receptive field of the top-level convolutional block. Even-numbered selective state-space layers embed elevator physical parameters into the continuous state transition matrix. Specifically, rated speed and car mass are used as fixed structural constraint parameters, reducing the number of free parameters and introducing prior physical knowledge. Rated speed and car mass are taken from the elevator nameplate parameters. Scale-aware gated residual connections are set between adjacent odd and even layers. The activated residual bypass is adaptively selected according to the current operating stage. The operating stage is determined by the slope of the speed curve, and the judgment threshold is determined by statistical analysis of the standard operating curve: the acceleration stage strengthens the bypass weights of short-term convolutional features, the constant speed stage strengthens the bypass weights of long-range state-space modeling, and the deceleration stage is between the two. At the end of the model, an extreme learning machine is fused to form a fast online fine-tuning layer. The weights of the hidden layer neurons of the extreme learning machine are randomly initialized and fixed during the training stage. Only the output layer connection weights participate in the least squares update. The online update of the last layer is completed in milliseconds using the current batch of data, achieving continuous adaptation. The model deployment employs a three-step lightweight process: weight pruning, INT8 quantization, and knowledge distillation, compressing the inference model volume to less than one-eighth of its original size. The weight pruning threshold is determined through iterative search under the constraint that the accuracy drop on the validation set should not exceed 1%. The scaling factor for INT8 quantization is obtained statistically from the dynamic range of the calibration dataset. The soft-label temperature coefficient for knowledge distillation is determined through a grid search within the range of 2 to 8, with the minimum validation set distillation loss as the criterion. On the field-application embedded platform, the cross-channel multiply-accumulate operations of the convolution kernels are allocated to the digital signal processing hard core of the field-programmable gate array (FPGA). Convolution calculations of different layers are scheduled in parallel in a pipelined manner. The intermediate activation tensors required by each layer are statically allocated to on-chip block random access memory partitions based on the inter-layer data dependencies. The least-squares update matrix operation of the extreme learning machine is allocated to a general-purpose processor core to avoid the FPGA reconstruction overhead.The steps for establishing the training dataset for the multi-scale temporal convolution and state-space fusion model specifically include: collecting continuous multi-channel sensor data for at least 3 months from no fewer than 50 in-service elevators of different models and floors, covering at least 5 types of operating conditions, including normal operation, guide rail joint impact, guide shoe wear, rope strand slack, and traction machine bearing wear; manually annotating each data segment by an engineer with elevator maintenance qualifications, including the operating condition type, fault level, and time of occurrence; dividing the data into training, validation, and test sets in a 7:1:2 ratio, and expanding the number of fault samples to be comparable to the number of normal samples using a data augmentation method based on variational autoencoders. The training steps for the multi-scale temporal convolution and state-space fusion model specifically include: using the weighted sum of cross-entropy loss and root mean square error loss as the total loss function, with an initial learning rate set to... ~ Within the specified range, a cosine annealing learning rate scheduling strategy is adopted; the batch size is set in the range of 32 to 128, and the specific value is determined according to the on-chip random access memory capacity of the embedded platform; the total number of training rounds is not less than 100 rounds, and an early stopping mechanism is adopted, terminating training when the validation set loss does not decrease for 20 consecutive rounds; in the knowledge distillation stage, the complete large model is used as the teacher network, and the lightweight model is used as the student network, and the mixing ratio of distillation loss and hard label loss is determined through validation set ablation experiments.

[0048] The technical benefits of the multi-scale temporal convolution and state-space fusion model are as follows: elevator vibration signals exhibit both non-stationarity and non-linearity, and single-scale convolution or cyclic structures cannot simultaneously account for short-term impacts and long-term trends. The model, through alternating stacking of odd and even layers, allows short-term convolutional receptive fields and long-term state-space modeling to function in parallel during the same forward propagation. The embedding of physical parameters provides prior constraints to the state transition matrix, avoiding overfitting under small sample conditions. The extreme learning machine's terminal layer enables the model to still adapt online even when there are significant individual differences in elevators, thereby achieving accurate discrimination of non-stationary fault characteristics under lightweight deployment constraints.

[0049] The dynamic learning rate adjustment function is calculated as follows: a comprehensive adjustment index is calculated by weighting three indicators: the validation set loss reduction rate of the current batch, the mean gradient norm, and the variance of the passenger transport quality assessment score. The specific weighting coefficients were determined through correlation analysis between the above three indicators and the model convergence speed during training. The ratio of the absolute values ​​of the correlation coefficients was used as the initial weights, which were then fine-tuned through five rounds of cross-validation. When the learning rate remains unchanged, the learning rate remains at its current value; when When the learning rate is multiplied by a fixed decay factor, the decay factor ranges from 0.80 to 0.95, and the specific value is determined by the validation set convergence experiment; when When the learning rate is multiplied by a large decay factor, the decay factor ranges from 0.50 to 0.79, and the specific value is determined by similar experiments; when When the learning rate is reset, the learning rate is reset to 10% to 30% of its initial value, and the cosine annealing schedule is restarted to escape local extrema.

[0050] The empirical mode decomposition is an adaptive signal decomposition method that does not require preset basis functions. By repeatedly screening the envelope of extreme points, the non-stationary signal is decomposed into several components that satisfy the intrinsic mode function conditions. Each component has a single amplitude modulation and frequency modulation characteristic, which is suitable for nonlinear non-stationary vibration analysis.

[0051] The Hilbert instantaneous frequency analysis involves constructing an analytical signal from a single-component signal, obtaining the instantaneous frequency by taking the time derivative of the phase function, and thus characterizing the evolution of the signal frequency over time.

[0052] The mutual information maximum relevance minimum redundancy feature selection algorithm calculates the mutual information between each feature and the fault category label (measuring relevance) and the mutual information between features (measuring redundancy), and iteratively selects a subset of features that maximizes relevance to the label while minimizing redundancy with the selected features.

[0053] The sparse autoencoder is a neural network that imposes sparsity constraints on the hidden layer activations of the autoencoder, forcing most hidden nodes to have near-zero activations under any input, thereby learning a low-dimensional sparse representation of the data for nonlinear dimensionality reduction.

[0054] The SHAP value originates from the Shapley value concept in game theory. By taking a weighted average of the marginal contributions of all feature subsets, each feature is assigned a contribution to the model prediction result, thereby achieving feature-level interpretability evaluation of the model prediction.

[0055] The physical channel grouping is a design method that assigns sensor signals from different physical mechanisms to independent convolutional channel groups. Within each group, convolutional kernel parameters are shared to take advantage of the signal similarity of similar sensors, and cross-physical domain information is transmitted between groups through an attention mechanism.

[0056] The extreme learning machine is a single-hidden-layer feedforward network. The hidden layer weights and biases are randomly initialized and then fixed. The output layer weights are solved only once by the least squares method, resulting in extremely fast training speed, which is suitable for online real-time update scenarios.

[0057] The Tikhonov regularization here refers to adding the solution to the objective function of the inverse problem. The norm penalty method balances the data fitting error and the smoothness of the solution by introducing a regularization parameter, thereby suppressing the noise amplification effect when inverting the ill-conditioned matrix.

[0058] The generalized cross-validation is an automatic regularization parameter selection method that does not require an independent validation set. It determines the optimal regularization parameter by minimizing the generalized cross-validation scoring function (a rotation-invariant approximation of leave-one cross-validation error).

[0059] The pseudo-inverse matrix is ​​a type of generalized matrix inverse. When the system transfer function matrix is ​​not invertible or ill-conditioned, the optimal approximate solution in the least squares sense is used to replace the exact inverse solution, and Tikhonov regularization is combined to ensure numerical stability.

[0060] The knowledge distillation method is a compression method that uses the softened output probability distribution of the large model (teacher network) as a supervision signal to train the small model (student network). The softening operation is achieved by scaling the Softmax input with a temperature coefficient, so that the student network learns the inter-class similarity structure captured by the teacher network.

[0061] The INT8 quantization is a model compression method that maps neural network weights and activation values ​​from 32-bit floating-point numbers to 8-bit integer representations. It maintains the approximate equivalence of values ​​before and after quantization through a linear scaling factor, which significantly reduces storage usage and multiplication-accumulation operation overhead on embedded platforms.

[0062] Optionally, the present invention also provides a method for implementing a system for real-time monitoring of elevator ride quality using a computer, wherein the computer is provided with a readable storage medium storing program instructions, and the program instructions can execute the above-described method when running on the computer.

[0063] The specific implementation of step S01 is as follows: Based on the static calibration test results, the installation positions of each sensor are determined at the center of the elevator car bottom, the four corner guide shoe mounting seats, and the ends of the traction rope sheave and car side ropes. A multi-channel sensor array is then arranged. The multi-channel sensor array consists of a vertical acceleration sensor, a lateral acceleration sensor, a current sensor, and a tension sensor. The installation direction of each sensor is strictly aligned with the direction of the target physical quantity to ensure a clear physical correspondence of the signals. The sampling rate is set in the range of 1000–8000 Hz, with a resolution of 16 bits. The sampling parameters are determined based on the elevator's rated speed and the guide rail joint spacing, ensuring that the Nyquist frequency is higher than the highest frequency component of interest. The static calibration test verifies the rationality of the installation position by applying a gradient load of 0.5–5 kN to the guide shoes and recording the responses of each sensor. This test is performed independently for each elevator after installation.

[0064] The specific implementation of step S02 is as follows: A circular buffer is established in the edge computing unit to continuously store real-time streaming data under memory-constrained conditions. When new data is written to a full-capacity buffer, the oldest data is automatically overwritten to ensure real-time performance. Lossy compression is achieved by performing wavelet coefficient threshold truncation on the original multi-channel signal in the buffer: First, the signal is subjected to discrete wavelet transform to obtain wavelet coefficients at each scale; then, using a preset threshold as the truncation standard, coefficients with absolute values ​​below the threshold are set to zero, retaining only the coefficients carrying the main energy; finally, the compressed signal is reconstructed from the retained coefficients. The compression threshold is determined by iterative statistical analysis of historical vibration data under multiple operating conditions, with a compression ratio of no less than 20:1, constrained by the root mean square error of the reconstructed signal not exceeding 5% of the root mean square amplitude of the original signal. The compressed data is uploaded to the cloud storage module in Apache Parquet columnar format, and a composite index is built according to time partitioning and device number. Hot data is stored on solid-state drives, and cold data is migrated to object storage. Simultaneously, multi-resolution retrieval is achieved in conjunction with a pre-calculated downsampling aggregation table.

[0065] The specific implementation of step S03 is as follows: Using the vertical vibration acceleration signal of the car as input, the vibration deconvolution source tracing algorithm for the inverse problem of guide rail geometric defects is executed. First, the time axis of the time-domain vibration signal is multiplied by the velocity using the known car running speed, converting it into a spatial domain signal with spatial position as the horizontal axis, thus establishing an explicit correspondence between vibration events and guide rail positions. Second, the spatial domain transfer function matrix of the car-guide shoe-guide rail system is constructed, with the guide shoe elastic parameters in the matrix taken from the static compression experiment calibration results. In the frequency-wavenumber domain, the pseudo-inverse matrix of the transfer function is used to perform deconvolution on the spatial domain signal spectrum, and the Tikhonov regularization method is used to suppress noise amplification of the ill-conditioned inverse problem. The regularization parameters are automatically selected by generalized cross-validation, thereby reconstructing the spatial distribution estimate of guide rail defects. Third, continuous wavelet transform is performed on the reconstructed defect spatial distribution, extracting the amplitude at each characteristic wavenumber and mapping it to the specific guide rail joint position, outputting a quantitative estimation report of defect position, amplitude, and wavelength.

[0066] The specific implementation of step S04 is as follows: The traction rope is modeled as a one-dimensional string vibration system, whose tension distribution satisfies the wave equation with damping. Using the acceleration sensor measurements on the traction sheave side and the car side as boundary conditions, and the tension measurement at the compensation sheave as the internal constraint, a time-stepped Crank-Nicolson implicit finite difference scheme is used to inversely reconstruct the spatiotemporal distribution of the entire rope tension wave within the framework of partial differential equations. The Crank-Nicolson scheme considers the spatial difference between the current and next time steps simultaneously within each time step, maintaining unconditional stability at any time step. The regularization term is the spatial second derivative of the rope tension distribution. Norm, used to suppress spatial noise amplification. The linear density, static tension, and damping coefficient in the wave equation are initially provided by the manufacturer's specifications, and then corrected by fitting experimental data from no-load operation after installation. The correction method is to minimize the difference between the measured boundary acceleration and the simulated boundary acceleration. Norm. The algorithm outputs a spatiotemporal distribution map of the entire rope tension and a probability heatmap of the damage location, which are used to locate rope strand slack and local damage.

[0067] The specific implementation of step S05 is as follows: Multi-channel sensor signals are input into the passenger transport quality state perception model, i.e., a multi-scale temporal convolution and state-space fusion model. The model consists of eight stacked layers: odd-numbered layers are dilated causal convolutional blocks, and even-numbered layers are selective state-space layers. Scale-aware gated residual connections are established between odd and even layers. Odd-numbered layers introduce a physical channel grouping mechanism, assigning vibration, electrical, and thermodynamic sensor signals to different convolutional channel groups. Information is exchanged between groups through a lightweight cross-channel attention module. Even-numbered layers embed rated speed and car mass as fixed structural constraint parameters into the continuous state transition matrix. At the end of the model, an extreme learning mechanism is fused to form a fast online fine-tuning layer, using the current batch data to complete the online update of the last layer in milliseconds. Before deployment, the model undergoes three lightweighting processes: weight pruning, INT8 quantization, and knowledge distillation, compressing the inference model volume to less than one-eighth of its original volume. After the model outputs the passenger transport quality assessment score, the dynamic learning rate adjustment function calculates a comprehensive adjustment index by weighting three indicators: the validation set loss reduction rate, the mean gradient norm, and the variance of the passenger transport quality assessment score. The learning rate is adjusted in real time according to the range of the comprehensive adjustment index: when the comprehensive adjustment index is not lower than 0.85, the learning rate remains unchanged; when the comprehensive adjustment index is between 0.60 and 0.85, it is multiplied by a decay factor; when the comprehensive adjustment index is lower than 0.30, a reheat mechanism is triggered, resetting the learning rate to 10% to 30% of the initial value and restarting cosine annealing scheduling.

[0068] The specific implementation of step S06 is as follows: The quantitative estimation results of the guide rail defect location, amplitude, and wavelength output in step S03, the probability heatmap of the entire rope damage location output in step S04, and the passenger transport quality assessment score output in step S05 are aligned and merged according to the elevator equipment number and timestamp to form a comprehensive elevator passenger transport quality report. The report triggers tiered alarms based on the quantitative indicators output by each submodule: a maintenance alarm is triggered when the guide rail defect amplitude or damage probability exceeds the alarm threshold; a passenger transport quality warning is triggered when the passenger transport quality assessment score is lower than the quality threshold; and an emergency shutdown alarm is triggered when multiple indicators exceed limits simultaneously. This tiered alarm mechanism ensures that fault information of different severity levels is communicated to maintenance personnel in a differentiated manner, achieving precise maintenance.

[0069] It should be noted that the key technologies of this invention include: the vibration deconvolution source tracing algorithm for the inverse problem of guide rail geometric defects, which restores the vibration response to the spatial distribution of guide rail defects through spatiotemporal transformation and pseudo-inverse matrix deconvolution, overcoming the limitation of traditional spectrum analysis that can only describe global frequencies and cannot locate spatial defects; the traction rope tension wave temporal domain reconstruction algorithm, which drives the inverse solution of partial differential equations through boundary measurement, reconstructs the tension distribution of the entire rope into a spatiotemporal two-dimensional map, overcoming the defect that single-point tension sensors cannot detect mid-section damage; and the multi-scale temporal convolution and state-space fusion model, which, through alternating stacking of odd and even layers and embedding of physical parameters, takes into account both short-term impact and long-term trend modeling and suppresses small-sample overfitting. The three key technologies work synergistically in the same data stream: spatial domain defect localization provides accurate spatial labeling priors for model training, the tension spatiotemporal map provides global physical constraints for passenger quality scoring, and the artificial intelligence model fuses multi-physical domain information into an interpretable comprehensive evaluation, enabling the overall monitoring system to form a mutually reinforcing technical synergy in three dimensions: positioning accuracy, global perception, and adaptive capability.

[0070] It should be noted that during long-term elevator operation, significant differences in mechanical parameters exist between different elevators of the same model. These differences include varying degrees of guide shoe wear, inconsistent traction rope tension, and varying dynamic load distribution within the car, leading to individual deviations in the statistical characteristics of vibration signals for each elevator. In such cases, when a passenger quality status perception model trained on a large-scale dataset is transferred to a new elevator, a systematic deviation exists between the decision boundary of its output layer and the actual signal distribution of the new elevator. This results in a decrease in the model's accuracy in identifying specific fault modes for that particular elevator. The reason for this technical problem is that the weights of the deep neural network are optimized to match the distribution of the training data during the training phase. When the signal distribution of the target elevator deviates from the training distribution, the fixed-weight network cannot automatically adjust the decision boundary. Retraining the entire network is neither feasible nor economical in real-time operation. A common solution to this problem is to use transfer learning to fine-tune the target domain data, i.e., updating the weights of all or part of the network layers with a small amount of labeled data from the new elevator. However, full fine-tuning requires significant backpropagation computation, making it difficult to complete online updates within milliseconds under the computational constraints of embedded edge platforms. Furthermore, limited target domain data can easily lead to overfitting of shallow feature extraction layers, impairing the model's generalization ability to common fault patterns. This invention effectively solves this technical problem by fusing a Extreme Learning Machine (ELM) at the end of a multi-scale temporal convolutional and state-space fusion model to form a fast online fine-tuning layer, concentrating the computational burden of individual adaptation on the output layer. The hidden layer weights of the ELM are randomly initialized and fixed after training; only the output layer connection weights are solved once using the least squares method, eliminating the need for backpropagation iterations and resulting in extremely low computational cost. Updates can be completed within milliseconds, meeting the real-time constraints of embedded platforms. Simultaneously, the deep feature extraction weights remain unchanged, ensuring that the generalization ability to common fault patterns is not affected by the individual adaptation process. This enables the model to continuously adapt to individual elevator differences in online deployment scenarios with limited computational resources.

[0071] Specifically, the principle of this invention is:

[0072] This invention solves the aforementioned core technical problems by establishing a complete physical inverse problem solution link from vibration response to guide rail excitation source. Specifically, the vertical vibration acceleration signal of the car is the convolution result of the geometric defects of the guide rail transmitted to the car through the guide shoe elastic system. In the time domain, the defect location information is modulated by the running speed and embedded in the time axis. When analyzed directly in the time or frequency domain, the mapping relationship between the spatial coordinates of the defect and the vibration characteristics is implicit. This invention first uses the known car running speed to multiply the time axis of the time domain signal by the velocity to convert it into a position axis, thus establishing an explicit correspondence between vibration events and spatial positions. The physical basis of this transformation is that a linear time-varying system can be approximated as a linear spatial domain system under uniform or slowly changing speed conditions.

[0073] After completing the time-space transformation, this invention constructs the spatial domain transfer function matrix of the car-guide shoe-guide rail system. This matrix describes the transfer characteristics of the guide shoe's elastic parameters to different wavenumber excitations. These parameters are pre-calibrated by static compression experiments, ensuring the accuracy of the physical model. Subsequently, the spatial domain signal is inversely filtered in the frequency-wavenumber domain, and the output spectrum is deconvolved using the pseudo-inverse matrix of the transfer function to reconstruct the spatial distribution of guide rail defects. Since the transfer function matrix exhibits ill-conditioned behavior at high frequencies or certain wavenumbers, direct inversion would lead to severe noise amplification. Therefore, this invention introduces the Tikhonov regularization method, adding the solution to the objective function. The norm penalty term strikes a balance between data fitting accuracy and solution smoothness, while the regularization parameter is automatically selected by generalized cross-validation, avoiding the subjectivity of manual parameter tuning.

[0074] Finally, continuous wavelet transform is performed on the reconstructed spatial distribution of defects to extract the amplitude at each characteristic wavenumber, which is then mapped to the specific guide rail joint location, outputting a quantitative estimate report of the defect location, amplitude, and wavelength. The above algorithmic chain logically constitutes a complete causal tracing closed loop: physical modeling formalizes the relationship between system response and excitation; spatiotemporal transform gives the signal spatial physical meaning; inverse filtering restores the excitation source; regularization ensures numerical stability; and wavelet transform extracts spatial features. Each step has a clear physical or mathematical basis, and the overall scheme is logically self-consistent, thus fundamentally overcoming the limitation of traditional spectrum analysis methods that can only provide a global frequency description but cannot locate spatial defects.

[0075] 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.

[0076] The specific implementation of step S01 is as follows: Multi-channel sensor arrays are respectively arranged on the elevator car floor, the guide shoe mounting base, and both ends of the traction rope. The arrays consist of vertical acceleration sensors, lateral acceleration sensors, current sensors, and tension sensors. The installation positions are determined by static calibration experiments. The sampling rate is set to 1000–8000 Hz, and the resolution is 16 bits.

[0077] The specific implementation of step S02 is as follows: Edge preprocessing uses a first-in-first-out circular buffer to continuously store streaming data. After performing discrete wavelet transform on the signal, data with absolute values ​​below a preset threshold is processed. The wavelet coefficients are set to zero, retaining only the coefficients carrying the main energy, with a preset threshold. The constraint that the root mean square error of the reconstructed signal does not exceed 5% of the root mean square amplitude of the original signal is expressed by the following formula:

[0078] ;

[0079] In the formula, For the first The original signal values ​​of each sampling point (unit: ), To reconstruct signal values ​​(unit: ), This represents the total number of signal sampling points. Both sides are dimensionless ratios. This was determined through multiple rounds of iterative statistical analysis of historical vibration data. The compression ratio is no less than 20:1. The compressed data is uploaded to the cloud in Apache Paquet format, and a composite index is built according to time partitioning and device number. Hot data is stored on solid-state drives, and cold data is migrated to object storage.

[0080] The specific implementation of step S03 is as follows: the vibration deconvolution source tracing algorithm for the inverse problem of guide rail geometric defects uses the vertical vibration acceleration signal of the car. As input, link the time axis with the velocity axis. Multiplication transforms the signal to the spatial domain, yielding the spatial domain acceleration signal. The position coordinates , Dubble variables for integration (unit: In the frequency-wavenumber domain, the system output spectrum... Spatial distribution of guide rail defects Transfer function matrix in the spatial domain The correlation is expressed in the following formula:

[0081] ;

[0082] In the formula, Space wavenumber (unit: ), Output acceleration spatial spectrum (unit: ), The system space domain transfer function matrix (unit: ), Spatial distribution spectrum of guide rail defects (unit: ), To measure the noise spectrum (unit: ). Based on the elastic parameters of the guide shoe (unit: ), guide shoe damping coefficient (unit: and car quality (unit: ) constitutes, its first The elements are:

[0083] ;

[0084] In the formula, The imaginary unit, Rated operating speed of the car (unit: ), The dimensionless transfer function matrix is ​​the first... Elements have the same dimensions for their numerators and denominators, while their ratios are dimensionless. and By applying 0.5 to 5 to the guide shoe The results were obtained through static compression experiments with gradient loading and recorded compressive displacement, with each elevator calibrated independently after installation. The ill-conditioned inverse problem was solved using the Tikhonov regularization method, with the objective function being:

[0085] ;

[0086] In the formula, Spatial distribution estimation of guide rail defects (unit: ), is the regularization parameter (dimensionless). Acceleration spectrum reference quantity (unit: ), Defect spectrum reference quantity (unit: Both terms are dimensionless after being normalized to reference values, and the objective function is dimensionless. The method is automatically selected by the generalized cross-validation method, i.e., minimizing the generalized cross-validation scoring function. :

[0087] ;

[0088] In the formula, The number of observation equations, For the trace operation of a matrix, It is the identity matrix. for The transpose of has dimensionless numerator and denominator. Dimensionless. Ultimately Solving using the pseudo-inverse matrix:

[0089] ;

[0090] In the above formula, The dimensions are ,and dimension After multiplying, we get dimension This is consistent with the defect spectrum. For the reconstructed... Perform continuous wavelet transform to extract the amplitude at each characteristic wavenumber, map it to the specific guide rail joint position, and output the quantitative estimation results of the defect location, amplitude and wavelength.

[0091] The specific implementation of step S04 is as follows: the traction rope is modeled as a damped one-dimensional string vibration system, and the tension wave spatiotemporal distribution is... It satisfies the wave equation, as expressed in the following formula:

[0092] ;

[0093] In the formula, Rope axis coordinates (unit: ), For time (unit: ), Total rope length (unit: ), For time reference (unit: ), Wave speed (unit: ), Linear density of traction rope (unit: ), Linear density reference quantity (unit: ), Static tension (unit: ), Normalized reference value for tension (unit: ), Damping coefficient (unit: ), External excitation per unit length (unit: All values ​​are dimensionless after being normalized to reference values. , , Initial values ​​are provided in the manufacturer's specifications, and then the measured boundary acceleration is minimized. With simulated boundary acceleration The difference The norm is adjusted, and the formula is expressed as follows:

[0094] ;

[0095] In the formula, The measured boundary acceleration vector (unit: ), The simulation boundary acceleration vector (unit: ), Boundary acceleration reference quantity (unit: The objective function is dimensionless. Using the acceleration measurements on the traction sheave side and car side as boundary conditions and the tension measurement at the compensation sheave as internal constraints, a time-stepped solution is performed using a Crank-Nicholson implicit finite difference scheme. The solution is obtained at the [missing information - likely a specific time step]. The spatial node, the first The discrete equations for each time step are:

[0096] ;

[0097] In the formula, For the first The spatial node Tension values ​​at each time step (unit: ), Spatial step size (unit: ), Time step (unit: Each term is dimensionless after being normalized to the reference value. The complete objective function with regularization constraints is:

[0098] ;

[0099] In the formula, The boundary tension values ​​extracted by numerical solution (unit: ), Measured boundary tension value (unit: ), Boundary tension reference value (unit: ), The space regularization parameter is dimensionless, and both parameters are dimensionless. The objective function is dimensionless. By tracking the propagation velocity and reflection behavior of tension waves, the location of local damage is detected, and the spatiotemporal distribution map of the whole rope tension and the probability heatmap of the damage location are output.

[0100] The specific implementation of step S05 is as follows: The passenger transport quality status perception model is a multi-scale temporal convolution and state-space fusion model with a total of 8 stacked layers. Odd-numbered layers are dilated causal convolutional blocks, and even-numbered layers are improved selective state-space layers. The odd-numbered layers assign vibration, electrical, and thermodynamic sensor signals to different convolutional channel groups, with shared convolutional kernel weights within each group. Let the input tensor be... , For the number of channels, For time steps, the first The odd-numbered layer The group's output is:

[0101] ;

[0102] In the formula, For the first Layer Group at time index The output at the location (units consistent with the input signal). For the first Layer The convolution kernel weights of the group (dimensionless). The kernel length is 1. For convolution kernel index, For the first The expansion factor of the layer and , This is the bias term (units consistent with the input signal). For the first Group input channel signals. Groups exchange information and attention weights via a lightweight cross-channel attention module. By the Group and No. The dot product of the group features is obtained after normalization. The weights are dimensionless. The selective state-space layer of even-numbered layers has a continuous state transition matrix. In-line embedded rated speed With car quality As fixed structural constraint parameters, the discretized state equation is:

[0103] ;

[0104] ;

[0105] In the formula, The hidden state vector. For the input vector, For the output vector, , , , Let be the discretized state matrix, where The structure is composed of and Parameterization Speed ​​reference quantity (unit: ), For quality reference quantity (unit: The scale-aware gated residual connection output between adjacent odd and even layers is:

[0106] ;

[0107] In the formula, For the first Layer input features, For the first Layer output features, For the first Layer output features, and Gating weights and satisfying , From the slope of the velocity curve (unit: ) and judgment threshold (unit: After comparison, linear interpolation is performed. The method was determined through statistical analysis of the standard operating curve. The model's terminal fusion limit learning mechanism forms a fast online fine-tuning layer, with the hidden layer output matrix set as follows: , This represents the batch sample size. The hidden layer neurons are 256 (an empirical value), and the output layer weights are... Solve using the least squares method:

[0108] ;

[0109] In the formula, For the label matrix, for transpose, This represents the output layer weight matrix. After three lightweighting steps—weight pruning, INT8 quantization, and knowledge distillation—the model's size is compressed to less than one-eighth of its original size. The formula for the softened output probability distribution of knowledge distillation is as follows:

[0110] ;

[0111] In the formula, For the first The softened output probability (dimensionless) of the class. For Teachers' Network The logical value of a class (dimensionless). For Teachers' Network The logical value of a class (dimensionless). For category indexing, The temperature coefficient is dimensionless, ranging from 2 to 8. The weighted total loss, calculated by combining distillation loss and hard label loss, is:

[0112] ;

[0113] In the formula, For dimensionless total loss, For hard-label cross-entropy loss, For soft label distillation loss, To avoid loss of reference value, all three have the same dimensions. The distillation-mixing ratio (dimensionless) was determined through validation set ablation experiments, with an empirical value of 0.3–0.7. The gradient scaling factor (dimensionless). The dynamic learning rate adjustment function is a comprehensive adjustment exponent. The calculation is as follows:

[0114] ;

[0115] In the formula, The current batch of validation set loss reduction rate (dimensionless). The mean of the gradient norms (dimensionless, calculated by dividing the actual gradient norm by the reference gradient norm). get), The variance of the passenger transport quality assessment score (dimensionless, calculated by dividing the actual variance by the reference variance). get), , , The weighting coefficients (dimensionless) were determined by taking the ratio of the absolute values ​​of the correlation coefficients and then fine-tuning them through five rounds of cross-validation. This is a dimensionless comprehensive adjustment index. When... The learning rate remains constant; when Multiply by a decay factor of 0.80 to 0.95; when Multiply by a decay factor of 0.50 to 0.79; when When the learning rate is triggered, the reheat mechanism is activated, resetting it to 10% to 30% of the initial value and restarting the cosine annealing schedule.

[0116] The specific implementation of step S06 is as follows: the quantitative estimation result of guide rail defects output by S03, the probability heat map of damage location output by S04, and the passenger transport quality assessment score output by S05 are fused from multiple sources to form a comprehensive elevator passenger transport quality report, and a graded alarm is triggered based on the fusion result.

[0117] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effect of the invention, technicians set up a test environment and deployed a multi-channel sensor array on a traction passenger elevator with a rated speed of 2.5 m / s and a rated load of 1600 kg in operation in a 32-story building. The signals from the elevator running continuously for 30 days were collected and processed to verify the entire process of the method of the invention.

[0118] In step S01, technicians place one vertical acceleration sensor at the center of the car bottom, one vertical acceleration sensor and one lateral acceleration sensor at each of the four corner guide shoe mounting bases, one acceleration sensor at each of the rope ends on the traction sheave side and the car side, one tension sensor at the compensation sheave, and one current sensor in the traction machine main circuit, for a total of 12 channels. The sampling rate is uniformly set to 4000Hz, and the resolution is 16bit. A static compression experiment is performed by applying five levels of gradient loads (0.5kN, 1kN, 2kN, 3kN, and 5kN) to the four corner guide shoes, fitting the elastic parameters of each guide shoe, and storing the results in the equipment file for subsequent algorithm initialization.

[0119] In step S02, a circular buffer with a sliding window capacity of 10 seconds is established within the edge computing unit. Three-level Daubechies-4 wavelet decomposition is performed on the original signal of each channel. The truncation threshold for each channel is determined with the constraint that the reconstruction root mean square error does not exceed 5% of the original signal's root mean square amplitude. The compression ratio reaches at least 24:1, meeting the requirement of no less than 20:1. The compressed data is uploaded to the cloud in Apache Parquet format, indexed by day and device number.

[0120] In step S03, technicians extract the vertical vibration acceleration signals of the car from day 7 to day 9 and process them using a vibration deconvolution source tracing algorithm based on the inverse problem of guide rail geometric defects. After converting the time-domain signal into a spatial-domain signal using the measured car running speed curve, a spatial-domain transfer function matrix based on the calibrated guide shoe elastic parameters is constructed, and Tikhonov regularized inverse filtering is performed in the frequency-wavenumber domain. The algorithm output shows that there is a joint misalignment defect approximately 8.4m from the pit, with an estimated amplitude of 1.2mm and a wavelength of 0.38m. Figure 2 As shown. Maintenance personnel then manually remeasured the location, and the measured misalignment was 1.1 mm, verifying the accuracy of the algorithm's positioning and amplitude estimation.

[0121] In step S04, the acceleration measurements at the boundaries of the traction sheave and car sides, and the tension measurements at the compensation sheave, are used as inputs to execute a time-domain reconstruction algorithm for the traction rope tension wave based on partial differential equation regularization. The algorithm employs a Crank-Nicolson implicit finite difference scheme with a spatial step size of 0.05 m and a time step size of 0.25 ms. The reconstruction results of the spatiotemporal distribution of the entire rope tension show a local tension depression approximately 3.2 m from the traction sheave on day 21. The damage probability heatmap shows a damage probability of 0.87 at this location. Figure 3 As shown. Technicians inspected the rope segment and found a slight slack in one strand, consistent with the algorithm's positioning results.

[0122] In step S05, the 12-channel signal from 30 days is sliced ​​into 4-second sliding window slices and input into the multi-scale temporal convolution and state-space fusion model. Before deployment, the model underwent three lightweight processing steps: weight pruning, INT8 quantization, and knowledge distillation. The inference model volume is one-seventh that of the original model, and the single inference latency on the embedded platform is 12ms. The Extreme Learning Machine's online fine-tuning layer performs a least-squares update after each batch of running data. The passenger quality assessment score output by the model tends to stabilize after the elevator has been in operation for three days, indicating that the model has successfully adapted to the individual characteristics of the elevator. The dynamic learning rate adjustment function triggered a reheat mechanism twice during training, both occurring during periods when the variance of the passenger quality assessment score fluctuation significantly increased. After the reheat, the model validation set loss recovered its downward trend within five rounds. The distribution of passenger quality assessment scores under various operating conditions during the 30-day operation period is shown in Table 1.

[0123] Table 1. Statistical Table of Passenger Transport Quality Assessment Scores for Each Working Condition

[0124]

[0125] In step S06, the outputs of the above modules are aligned and merged according to timestamps to form a comprehensive elevator ride quality report for 30 days. The report triggered three maintenance alarms, corresponding to the following: guide rail joint defect exceeding limits on day 7; rope strand slack damage probability exceeding limits on day 21; and ride quality assessment score continuously falling below the quality threshold on day 28. All three alarms were consistent with the results of manual retesting, with no missed or false alarms. The multi-physics domain fusion analysis framework of the comprehensive report is as follows: Figure 4 As shown.

[0126] The advancements of this invention compared to traditional methods are reflected in the following aspects: Traditional elevator vibration monitoring methods can only determine the existence of anomalies in the global frequency dimension. Their fundamental limitation lies in the loss of the correspondence between signals and spatial locations in frequency domain analysis, making it impossible for maintenance personnel to obtain the precise spatial coordinates of defects from the monitoring system. This invention, through time-space transformation and inverse problem solving, restores the vibration response to a defect distribution map with spatial physical meaning, making the location and amplitude of defects quantifiable parameters, fundamentally changing the way maintenance personnel obtain fault information. For traction rope detection, traditional single-point tension sensors can only sense the tension value at the installation point, failing to reflect the damage state of the rope's middle section. This invention reconstructs the spatiotemporal distribution of the entire rope tension using partial differential equations as the physical framework, transforming tension depressions and wave velocity mutations caused by local damage into quantifiable detection features, filling the technological gap of single-point sensor perception blind spots. The passenger quality status perception model, through physical parameter embedding and online fine-tuning using an extreme learning machine, maintains its ability to generalize to common fault modes while achieving rapid adaptation to individual elevator differences, solving the inherent defect of fixed-weight models in decision boundary shifts during individual migration scenarios.

[0127] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0128] Table 2. Variable Explanation Table (Part 1)

[0129]

[0130] Table 3. Variable Explanation Table (Part Two)

[0131]

[0132] 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 for real-time monitoring of elevator ride quality, characterized in that, Includes the following steps: Multi-channel sensor arrays are arranged in the elevator car, guide shoes and both ends of the traction rope to collect vertical vibration acceleration, electrical parameters and rope end acceleration signals; The acquired multi-channel signals are preprocessed at the edge, the raw data is stored in a circular buffer, and lossy compression is completed by wavelet coefficient threshold truncation method. The compressed data is then uploaded to the cloud columnar storage module. The vertical vibration acceleration signal of the car is subjected to the vibration deconvolution source tracing algorithm of the inverse problem of guide rail geometric defects. The time domain signal is converted into a spatial domain signal according to the car running speed. The spatial distribution of guide rail defects is restored by inverse filtering of the spatial domain transfer function matrix, and the quantitative estimation results of defect location, amplitude and wavelength are output. A time-domain reconstruction algorithm for traction rope tension waves based on partial differential equation regularization is applied to the acceleration signals at both ends of the traction rope and the tension signal at the compensation wheel. The spatiotemporal distribution of the tension waves of the entire rope is reconstructed in reverse, and a probability heatmap of the damage location is output. The multi-channel signals are input into the passenger transport quality status perception model, which outputs the passenger transport quality assessment score. The learning rate parameter of the passenger transport quality status perception model is adjusted in real time by the dynamic learning rate adjustment function. The quantitative estimation results of guide rail defects, the probability heat map of damage location, and the passenger transport quality assessment score are integrated to form a comprehensive elevator passenger transport quality report, which triggers a graded alarm.

2. The method for real-time monitoring of elevator ride quality according to claim 1, characterized in that, The multi-channel sensor array consists of a vertical acceleration sensor, a lateral acceleration sensor, a current sensor, and a tension sensor. The installation position is determined by a static calibration experiment.

3. The method for real-time monitoring of elevator ride quality according to claim 2, characterized in that, The lossy compression step specifically involves performing a discrete wavelet transform on the signal, setting the wavelet coefficients with absolute values ​​below the compression threshold to zero, and retaining only the coefficients carrying the main energy for reconstruction, with a compression ratio not lower than the compression ratio threshold.

4. The method for real-time monitoring of elevator ride quality according to claim 3, characterized in that, The cloud-based columnar storage module uses the Apache Parquet format, constructs a composite index based on time partitioning and device number, stores hot data on solid-state drives, and migrates cold data to object storage.

5. The method for real-time monitoring of elevator ride quality according to claim 4, characterized in that, The proposed vibration deconvolution source tracing algorithm for the inverse problem of guide rail geometric defects specifically involves performing deconvolution operations on the output spectrum using the pseudo-inverse matrix of the transfer function in the frequency-wavenumber domain. Regularization employs the Tikhonov method, and the regularization parameters are automatically selected by generalized cross-validation.

6. The method for real-time monitoring of elevator ride quality according to claim 5, characterized in that, The guide shoe elastic parameters in the spatial domain transfer function matrix are pre-calibrated by a static compression experiment, in which a gradient load is applied to the guide shoe and the compression displacement is recorded, and the guide shoe elastic parameters are obtained by fitting.

7. The method for real-time monitoring of elevator ride quality according to claim 6, characterized in that, The aforementioned time-domain reconstruction algorithm for traction rope tension waves specifically employs a time-stepping Crank-Nicolson implicit finite difference scheme to inversely reconstruct the spatiotemporal distribution of the entire rope tension wave within the framework of the wave equation. The regularization term is the spatial second derivative of the rope tension distribution. Norm.

8. The method for real-time monitoring of elevator ride quality according to claim 7, characterized in that, The passenger transport quality status perception model is a multi-scale temporal convolution and state space fusion model, which is composed of an improved temporal convolutional network and a selective state space mechanism stacked alternately, and the end fusion limit learning mechanism forms a fast online fine-tuning layer.

9. The method for real-time monitoring of elevator ride quality according to claim 8, characterized in that, The odd-numbered layers of the multi-scale temporal convolution and state-space fusion model introduce a physical channel grouping mechanism in the channel dimension, which assigns vibration, electrical, and thermodynamic sensor signals to different convolutional channel groups, and the groups exchange information through a lightweight cross-channel attention module.

10. The method for real-time monitoring of elevator ride quality according to claim 9, characterized in that, The even-numbered layers of the multi-scale temporal convolution and state-space fusion model embed rated speed and car mass as fixed structural constraint parameters in the continuous state transition matrix, and scale-aware gated residual connections are set between adjacent odd and even layers.

Images