An elevator steel wire rope tension intelligent monitoring method fusing sines geometry topological features and Mamba-KAN network

Abstract

The application provides an elevator steel wire rope tension intelligent monitoring method fusing symplectic geometry topology features and a Mamba-KAN network, first, denoising and standardization pretreatment are performed on collected multi-source heterogeneous data; second, two-stage progressive multi-dimensional operation excitation decoupling is performed on the pretreated multi-source heterogeneous data to identify a real steel wire rope tension imbalance sequence; then, symplectic geometry modal decomposition and phase space reconstruction are performed on the real steel wire rope tension imbalance sequence to extract string vibration principal component, and the string vibration principal component is mapped into a two-dimensional topology feature image through a Gram angle field algorithm; again, the two-dimensional topology feature image is input into the Mamba-KAN network, a Mamba encoder extracts long-range time sequence dependence with linear complexity, a KAN regression layer explicitly fits a nonlinear mapping between string vibration frequency and physical tension through a learnable spline function, and a double-path parallel output outputs a single rope tension prediction value and four types of health states; finally, three-stage progressive safety intervention is performed according to a tension deviation rate and a state classification result, and the deviation is strictly limited within 5%.

Description

Technical Field

[0001] This invention relates to the field of elevator operation safety monitoring technology, specifically to an intelligent monitoring method for elevator wire rope tension that integrates symplectic geometric topological features and Mamba-KAN networks. Background Technology

[0002] The elevator traction steel wire rope system is the core load-bearing component of the elevator. Its safety status directly determines the reliability of the entire elevator operation and the safety of passengers' lives and property. According to the strict requirements of the "Safety Specifications for Elevator Manufacturing and Installation" and the "Rules for Supervision and Periodic Inspection of Elevators", the tension deviation between multiple suspension ropes must always be controlled within a specific safety boundary (such as 5%) of the average value. If the tension deviation exceeds the limit, it will not only accelerate the fatigue breakage of individual steel wire ropes and abnormal wear of traction sheave grooves, leading to a waste of high resources and resource costs for replacing steel wire ropes in batches, but also cause safety hazards such as loosening of steel wire ropes, severe physical flutter, or even jumping out of the groove during operation.

[0003] Existing elevator wire rope tension monitoring technologies mainly include two categories: traditional sensor direct measurement and string vibration theory soft measurement. Traditional sensor direct measurement involves connecting a pressure sensor in series at the rope end or using a mechanical tension gauge. This method has limited installation space, high maintenance costs, and difficulty in achieving dynamic monitoring throughout the entire stroke. String vibration theory soft measurement utilizes the mapping relationship between the first-order natural frequency of the wire rope and the physical tension to inversely deduce the tension from the vibration signal. Although this method can achieve non-contact dynamic monitoring, the multi-degree-of-freedom and highly nonlinear dynamic characteristics of the wire rope system cause its early weak excitation characteristics to be submerged by shaft wind resistance, low-frequency resonance of the traction machine, and strong electromagnetic background noise. Existing signal decomposition algorithms such as EMD, FMD, and VMD lack a topology preservation mechanism for the original dynamic structure. When stripping away strong noise, they easily destroy the original manifold structure of string vibration, resulting in poor agreement between theoretical model calculation results and experimental measurement data.

[0004] In addition, during actual elevator operation, there are typical sources of interference such as sudden dynamic load changes caused by passengers walking in the car and transient mechanical impacts when the car passes through the guide rail joints. These sudden pulses caused by normal operation excitation are highly similar in timing and vibration characteristics to the actual tension imbalance, loosening, and slippage precursors of the wire rope. Traditional analysis models are prone to misdiagnosis due to the lack of a multi-dimensional spatiotemporal interference stripping mechanism.

[0005] Meanwhile, conventional deep learning models such as CNN or Transformer exhibit quadratic-level computational complexity when processing ultra-long time-series high-frequency vibration sequences, making it difficult to achieve millisecond-level real-time deployment on elevator controllers with limited computing power. Furthermore, the internal logic of these black-box models based on multilayer perceptrons is invisible, and they cannot provide clear mathematical analytical expressions, making it difficult for theoretical model calculation data to meet the basic requirements for safety verification of elevator components. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent monitoring method and system for elevator wire rope tension that integrates symplectic geometric topological features and Mamba-KAN networks. To achieve the above objective, this invention employs the following technical solutions: A method for intelligent monitoring of elevator wire rope tension that integrates symplectic geometric topological features and Mamba-KAN networks includes the following steps: S1. Multi-source signal acquisition and preprocessing: Radial and axial vibration signals of multiple steel wire ropes, car weighing load signals and car position signals are acquired by high-frequency MEMS vibration sensors, car bottom load weighing sensors and high-precision absolute encoders respectively; and the multi-source heterogeneous signal data are denoised and dimensionless normalized. S2. Multidimensional operation excitation decoupling: The first level is load transient time domain decoupling, which identifies candidate signals of wire rope tension imbalance by constructing a sliding energy integral window based on the load change rate of the car weighing load data; the second level is spatial position joint verification, which combines the above-mentioned candidate signals of wire rope tension imbalance with the absolute displacement coordinates of the car to perform spatial distribution statistics and then identify the real wire rope tension imbalance sequence. S3, Symmetric Geometric Topology Reconstruction: Perform symmetric geometric mode decomposition and phase space reconstruction on the real wire rope tension imbalance sequence, extract the principal component of the string vibration, and map the one-dimensional principal component sequence into a two-dimensional topological feature image through the Gram angle field algorithm. S4, Mamba-KAN tension prediction: The two-dimensional topological feature image is input into the Mamba encoder to extract long-range temporal dependencies, and then the KAN regression layer fits the nonlinear mapping between the string vibration frequency and the physical tension through a learnable spline function, outputting the single rope tension prediction value and health status classification. S5. Closed-loop safety intervention: Calculate the tension deviation rate based on the single rope tension prediction value, and implement a three-level progressive intervention of continuous monitoring, early warning push, and cutting off the safety loop in combination with the health status classification results.

[0007] Preferably, the preprocessing process in step S1 includes using the Laida criterion to remove electromagnetic interference extreme noise points and using Z-Score normalization to eliminate dimensional differences.

[0008] Preferably, the first-stage load transient time-domain decoupling in step S2 includes calculating the first-order time derivative of the load signal from the weighing sensor to obtain the load change rate, constructing a sliding energy integral window to quantify the coupling strength between load mutation and vibration response, and setting an impulse threshold and a derivative peak threshold as discrimination boundaries; when the sliding energy is lower than the impulse threshold and the load change rate exceeds the derivative peak threshold, it is determined to be a pseudo-fault pulse caused by passengers walking or jumping in the car and is eliminated; when the sliding energy is stable and the high-frequency vibration envelope of the wire rope is continuously abnormal, it is identified as a candidate signal of wire rope tension imbalance.

[0009] Preferably, the second-level spatial position joint verification in step S2 includes combining the candidate signals of wire rope tension imbalance output from the first level with the absolute displacement coordinates of the car to perform spatial distribution statistics, setting the tolerance window range at the guide rail joint, and establishing a guide rail joint position fingerprint database; if the candidate signal is fixed within the same guide rail joint tolerance range in multiple consecutive up and down strokes, it is determined to be an occasional environmental excitation caused by mechanical impact of the guide rail joint and is excluded; if the candidate signal continues to reappear in the non-joint area throughout the entire stroke and the joint reappearance probability exceeds the set probability threshold, it is diagnosed as a real wire rope tension imbalance sequence.

[0010] Preferably, the symplectic geometric mode decomposition and phase space reconstruction in step S3 includes mapping the real wire rope tension imbalance sequence to a Hankel trajectory matrix according to the embedding dimension, constructing a covariance matrix and solving the symplectic orthogonal transformation through the Hamiltonian matrix, performing eigenvalue decomposition on the manifold matrix, selecting the top few principal components whose cumulative contribution rate reaches a threshold or higher as the optimal string vibration principal component sequence based on the energy contribution rate of each modal component, and maintaining the nonlinear manifold topology while filtering out background noise.

[0011] Preferably, the Gram angle field algorithm mapping in step S3 includes normalizing the optimal string vibration principal component sequence and converting it to a polar coordinate system, constructing a Gram angle field matrix through trigonometric function inner product, mapping the one-dimensional principal component time series to a two-dimensional topological feature image, and converting the implicit tension fluctuation into a spatial visual texture tensor.

[0012] Preferably, the Mamba encoder in step S4 includes defining a continuous-time state-space equation and discretizing it through a zero-order hold to achieve a long-range temporal recursive approximation with linear complexity, and using a selective mechanism to dynamically adjust the state transition parameters according to the input content to extract the long-range temporal excitation dependence of multiple wire ropes.

[0013] Preferably, the KAN regression layer in step S4 includes replacing the fixed weight matrix of the traditional neural network with a multidimensional superimposed learnable spline function. By superimposing the basic activation function and the B spline function, the nonlinear mapping relationship between the string vibration frequency and the physical tension is explicitly approximated, so that the theoretical model calculation data and the experimental measurement data are in high agreement.

[0014] Compared with the prior art, the present invention has the following advantages: 1. Symmetric geometric mode decomposition is adopted to replace the traditional frequency domain decomposition. The one-dimensional non-stationary vibration sequence is mapped to the Hankel trajectory matrix. The covariance matrix is ​​constructed and solved by the Hamiltonian matrix to obtain the symmetric orthogonal transformation. The eigenvalue decomposition of the manifold matrix is ​​performed to extract the optimal string vibration principal component with a cumulative contribution rate of more than 85%. This method filters out strong background wind resistance and mechanical structure noise without distorting the topology of the original multi-degree-of-freedom nonlinear string vibration manifold. It largely preserves the real physical excitation characteristics of the wire rope tension micro-change and the precursor of the slip, so that the theoretical model calculation data matches the experimental measurement data.

[0015] 2. A two-stage progressive decoupling mechanism is constructed. The first stage is load transient time-domain decoupling: the load change rate of the load sensor load signal is calculated, and a sliding energy integral window is constructed to quantify the coupling strength between load mutation and vibration response, and identify it as a candidate signal of tension imbalance. The second stage is spatial position joint verification: the candidate signal is combined with the absolute displacement coordinate of the car to perform spatial distribution statistics and identify the real wire rope tension imbalance sequence. Through the above two-stage decoupling, the load transient interference and guide rail joint impact are effectively separated, thereby improving the system's anti-interference ability to normal operation excitation and the sensitivity to the identification of real faults.

[0016] 3. In terms of computational efficiency, the Mamba selective state-space model is adopted, defining a continuous-time state-space linear ordinary differential equation and discretizing it through a zero-order hold. This alleviates the problem of quadratic explosion in computational complexity when traditional Transformer networks process ultra-long time-series high-frequency vibration sequences. In terms of physical interpretability, the KAN Kolmogorov-Arnold network is adopted, using multi-dimensional superimposed learnable spline functions to replace the fixed weight matrix. By superimposing the basic activation function and the B-spline function, the nonlinear mapping law between the string vibration frequency and the physical tension is explicitly approximated. This solves the black box problem of the invisible internal logic of traditional deep learning models, giving the tension prediction a clear mathematical and physical meaning. This makes the theoretical model calculation data highly consistent with the experimental measurement data, meeting the basic requirements for safety verification of elevator components. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0018] Figure 2This is a schematic diagram of the hardware layout and active safety monitoring link for multi-degree-of-freedom generalized sensing of steel wire ropes in an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram of the multidimensional operation excitation spatiotemporal decoupling and interference stripping logic in an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of the topological reconstruction principle of nonlinear manifold based on symplectic geometric decomposition and Gram angle field polar coordinate mapping in an embodiment of the present invention.

[0021] Figure 5 This is a schematic diagram of the Mamba-KAN state-space network structure and the tension physical prediction topology based on explicit spline function approximation in an embodiment of the present invention. Detailed Implementation

[0022] The embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0023] To verify the generalization ability of the method of this invention, tests were conducted on a scaled-down experimental platform that satisfies the principles of geometric similarity, tension similarity, and string vibration frequency similarity. The diameter of the experimental wire rope was set to 0.8 mm to 1.4 mm, and mapped to the actual elevator wire rope specification range using a scaling factor. Forty different diameter wire ropes were distributed proportionally within the 0.8 mm to 1.4 mm diameter range, with four of each diameter. These four ropes of the same diameter were assembled into one elevator unit. The experiment included load conditions such as no load, half load, full load, and near the rated load, covering upward, downward, acceleration, constant speed, deceleration, guide rail joint area, non-joint area, and transient disturbance conditions caused by passenger movement or jumping. Typical states such as normal, tension deviation within limits, tension exceeding limits, severe loosening, and signs of impending elevator tripping were also constructed. Each elevator unit under independent operating conditions undergoes 10 full-stroke tests. During each run, multi-source heterogeneous data such as radial and axial vibration of the four steel wire ropes in the elevator unit, car load, and position coding are collected simultaneously. Through the above tests with multiple dimensions, multiple loads, multiple operating stages, and multiple health states, the applicability of this method in multiple scenarios is verified.

[0024] like Figure 1 Module 1 and Figure 2As shown, the monitoring system of this invention includes a high-frequency MEMS vibration sensor, a car bottom load weighing sensor, a high-precision absolute encoder, and an intelligent monitoring terminal. The high-frequency MEMS vibration sensor is rigidly fixed to the rope end assembly of each suspended wire rope, ensuring no relative displacement between the high-frequency MEMS vibration sensor and the rope body to accurately transmit micro-chord vibration signals. The frequency response range of the high-frequency MEMS vibration sensor is set to 0.1~5000Hz, and the sampling rate is 5kHz to cover the high-frequency chord vibration band of the wire rope. Signal acquisition follows the principle of radial and axial dual-channel, multi-rope synchronous triggering, that is, each wire rope is equipped with two orthogonal high-frequency MEMS vibration sensors to capture the radial oscillation and axial extension vibration of the wire rope respectively. All high-frequency MEMS vibration sensors are synchronously sampled through the same clock source to achieve multi-channel... The timing of the steel wire rope vibration signal is aligned; the arrangement and configuration of the car bottom load weighing sensor are as follows: a car bottom load weighing sensor with a range of 0~5000kg and an accuracy of ±0.5% is placed at the load-bearing support point between the bottom of the car and the car frame to obtain the dynamic load data of the car in real time and accurately. The sampling rate of the car bottom load weighing sensor is set to 1kHz to meet the requirements for capturing load transients; the arrangement and configuration of the high-precision absolute encoder are as follows: a high-precision absolute encoder with a resolution of 0.01mm is placed at the end of the traction machine main shaft or at the speed limiter wheel axle to accurately obtain the absolute position coordinates of the car. The sampling rate is set to 1kHz to achieve timing alignment with the load signal; the data cache depth of the intelligent monitoring terminal is set to a 10-second rolling cache to support historical data backtracking analysis.

[0025] After acquiring the raw data, the intelligent monitoring terminal performs the following preprocessing operations: First, the Laida criterion is used to remove transient extreme noise points caused by electromagnetic interference in the hoistway. For each data point in the sampling sequence of various signals, such as the radial and axial vibration signals of the wire rope, the car load signal, and the car position encoding signal, if its value exceeds the mean... If the range is ±3 times the standard deviation, it is judged as abnormal noise and removed, and filled with linear interpolation of the nearest valid data; secondly, the Z-Score normalization method is used to eliminate the dimensional differences of various signals such as the radial and axial vibration signals of the wire rope, the car load signal, and the car position encoding signal. The sampling sequence X of various signals such as the radial and axial vibration signals of the wire rope, the car load signal, and the car position encoding signal after processing by the Laida criterion is mapped to a unified dimensionless interval. The normalization formula is: ; In the formula: This is the original sampling sequence. To normalize the data, and These represent the mean and standard deviation of the corresponding sequences, respectively. The mean of the normalized data is 0 and the standard deviation is 1, which facilitates the subsequent fusion processing of multi-source heterogeneous data. Finally, the sampling sequences of the three heterogeneous signals—the radial and axial vibration signals of the wire rope, the car load signal, and the car position encoding signal—are aligned by timestamps and divided into frames with a fixed length (1024 points in this embodiment). Each frame of data serves as the basic unit for subsequent decoupling analysis.

[0026] like Figure 1 Module 2 and Figure 3 As shown, multi-source heterogeneous data, after denoising using the Laida criterion and Z-Score normalization preprocessing, undergoes multi-dimensional operational excitation decoupling to eliminate false alarms caused by normal operation excitations such as passenger movement within the car and impacts at guide rail joints. The decoupling process consists of two progressive discrimination stages: the first stage is time-domain decoupling of transient car load changes, which involves real-time calculation of the load signal detected by the car bottom load weighing sensor. First time derivative To obtain the rate of change of car load, characterizing the degree of transient change in load within the car, a construction length of [length missing] is [length missing]. Sliding energy integral window : ; in, For transient characteristics, time window For the load change rate, the impulse threshold can be set based on the mean-standard deviation criterion or the quantile criterion, and based on the no-load, half-load, full-load, passenger movement, jumping disturbance, normal full-journey operation, and calibrated tension imbalance samples during the system initialization phase. With derivative peak threshold And make rolling corrections based on car load, running speed, and historical operating data; when making a judgment, when and If the defect is determined to be a short-term non-physical impact caused by a passenger's momentary movement or jumping within the car, the defect assessment will be suspended. and If the load fluctuation is deemed normal, the defect assessment will be terminated. Furthermore, when the high-frequency vibration envelope of the wire rope remains abnormal within a continuous time window, it is identified as a candidate signal of wire rope tension imbalance and enters the second-level spatial position joint verification; when and When an anomaly occurs and it is impossible to directly distinguish between load disturbance and wire rope anomaly, the evaluation is not terminated, but instead proceeds to the second-level spatial position joint verification.

[0027] The second level is spatial position joint verification, used to spatially trace and identify the candidate signals of wire rope tension imbalance output from the first level. During the system initialization phase, through multiple full-stroke runs, the absolute position coordinates of each guide rail joint are automatically learned and calibrated. Establish a fingerprint database of guide rail joint locations; combine this with the absolute displacement coordinates of the car detected by a high-precision absolute encoder. Taking into account factors such as the resolution of the high-precision absolute encoder, the elastic elongation of the wire rope, and the vibration during car operation, the tolerance window range at the guide rail joints is set. For ±50mm, the spatial distribution of candidate signals for wire rope tension imbalance that passed the first-level screening was statistically analyzed. The specific judgment logic was as follows: assuming the absolute displacement range of the elevator's entire travel is ±50mm... The guide rail joint position is , … The seam area is then represented as: ; The non-seamless area is represented as: }; Let the absolute displacement coordinate of the car corresponding to the candidate signal of wire rope tension imbalance output by the transient time-domain decoupling of the first-stage load be... ,like If the anomaly occurs within the joint area, the process enters the joint impact investigation branch. At this point, the spatial distribution over multiple consecutive (5 times in this embodiment) up-and-down strokes is statistically analyzed. If the anomaly coordinates are consistently within the same joint area... If the issue is internal, it is determined to be an incidental environmental excitation caused by mechanical impact at the guide rail joint and is eliminated; otherwise, it is transferred to the non-joint area for processing. This means the anomaly occurred in the non-seam area. Calculate the joint recurrence probability of anomaly coordinates: ; If the joint recurrence probability If the failure rate is >95%, then occasional environmental interference is completely ruled out, and the problem is determined to be a defect in the wire rope itself or a tension imbalance. A sequence of defects or tension imbalances in the wire rope is then output. .

[0028] like Figure 1 Module 3 and Figure 4 As shown, the above sequence, determined through two-stage progressive decoupling, is identified as a defect or tension imbalance sequence within the wire rope itself. Perform symplectic geometric mode decomposition and phase space noise reduction, specifically by resolving the inherent defects or tension imbalance sequences of the wire rope. Mapped to the Hankel trajectory matrix with an embedding dimension of d=20: ; The selection of the embedding dimension is based on the spurious neighbor method in phase space reconstruction theory, ensuring that the reconstructed phase space can fully develop the dynamic characteristics of the original system. Then, the covariance matrix is ​​constructed. And construct the Hamiltonian matrix. Solve for the symplectic orthogonal transformation matrix. Satisfying the sine condition ,in For a standard symplectic matrix, To and Identity matrices of the same dimension, through symplectic orthogonal transformation Eigenvalue decomposition of the manifold matrix yields the eigenvalue matrix. ,in For each modal component, calculate its energy contribution rate based on its eigenvalue. 100%, select those that meet the conditions ≥85% of the front The principal components are used as the optimal string vibration principal component sequence. .

[0029] The optimal principal component sequence Normalizing to the interval [-1, 1] yields And transform to polar coordinates: ; ; In the formula, For polar coordinates, The radius is in polar coordinates. Given the sequence length, the Gram angle field feature image matrix G is constructed by calculating the trigonometric inner product between time steps i and j: ; This matrix is ​​a real symmetric matrix, and its element values ​​reflect the correlation strength between different time steps. In this embodiment, the feature image matrix G is transformed into a two-dimensional polar coordinate topological feature image of size 64×64 pixels, and the N×N matrix G is uniformly downsampled to 64×64. ; In the formula, For floor operations, The pixel coordinates of the downsampled image strike a balance between preserving sufficient texture details and controlling computational complexity. This transforms the hidden temporal tension fluctuations into a spatial visual texture tensor that reflects the dynamic stress on the wire rope, serving as the input to the subsequent Mamba-KAN network.

[0030] like Figure 1 Modules four and five in Figure 5As shown, a Mamba-KAN deep network model is constructed and deployed within a smart terminal. This model adopts a two-stage architecture. In the feature encoding stage, the Mamba selective state-space model is used to extract long-range temporal dependencies. In the regression and classification stage, the KAN Kolmogorov-Arnold network is used for white-box physical tension mapping. Specifically, a 64×64 pixel two-dimensional polar coordinate topological feature image and its corresponding load feature vector are concatenated and input into the Mamba encoder. For extremely long continuous high-frequency sampling features, a continuous-time hardware state-space linear ordinary differential equation is defined: ; In the formula, The hidden state vector is used to encode a compressed representation of historical time-series information. For input features, This is the state transition matrix, which controls the decay and continuation of memory. To input the control matrix and control the effect of new inputs on memory, a step size scalar is introduced because actual calculations require discretization. Using a zero-order hold for continuous matrices and Perform discretization mapping: ; ; The discretized recursive formula is: ; ; In the formula, Let k be the discrete hidden state at step k. For the output of step k, C and D are the output matrices. The selective mechanism of the Mamba encoder enables it to dynamically adjust the state transition parameters according to the input content, automatically focus on important vibration features and suppress irrelevant noise, and achieve low computational overhead for long-range string vibration feature extraction on a controller with limited computing power.

[0031] output from Mamba The input is fed into a KAN regression classifier for tension prediction. The KAN regression classifier uses a multi-dimensional superimposed learnable spline function to replace the fixed weight matrix of a traditional neural network. This classifier's nonlinear string vibration mapping function... Represented as: ; In the formula, the activation boundary function From the basic activation function With B-spline function Superimposed composition: ; in, , , The learnable parameters of this KAN regression classifier are adaptively updated during backpropagation. Through the combination of the spline functions mentioned above, the KAN regression classifier explicitly approximates the physical law of the nonlinear evolution mapping between string vibration frequency and physical tension. This breaks the limitation of the uninterpretability of the internal mechanism caused by the fixed internal weight matrix and lack of clear mathematical and physical meaning in traditional deep learning models, making the data calculated by the theoretical model highly consistent with the experimental measurement data. In the model training stage, the AdamW optimizer is used, with an initial learning rate of 0.001, a batch size of 32, and 200 training rounds. The learning rate decay strategy is cosine annealing, and the regularization uses weight decay of 0.01. After 200 rounds of backpropagation iterations, the learnable spline functions in this KAN regression classifier adaptively adjust their node coefficients and weight parameters, gradually optimizing the shape of the nonlinear mapping curve, and finally successfully fitting the nonlinear mapping relationship between string vibration frequency and physical tension that is highly consistent with the theoretical physical law of string vibration.

[0032] After training, the Mamba-KAN deep network model is deployed on an intelligent monitoring terminal, enabling real-time inference during actual operation. The intelligent monitoring terminal receives the quantitative prediction value of the dynamic tension of a single rope in real time. And calculate the average tension of each suspension rope in real time: ; Then, the tension deviation rate was calculated for each individual component: ; In the formula, Let be the predicted tension of the i-th wire rope. The average tension of all wire ropes, Let be the tension deviation rate of the i-th wire rope. During state determination, the deep feature tensor output by the Mamba encoder is input into the KAN regression classifier. The Softmax classifier in the KAN regression classifier outputs probability vectors for four states: normal, tension out of tolerance, severe loosening, and signs of slippage. Specifically, during model training, measured single-rope tension, tension deviation rate, manually calibrated health status, and corresponding vibration characteristics are used as supervision labels. Joint training is employed using tension regression loss and state classification cross-entropy loss. During model inference, the system first calculates the predicted value of the single-rope tension. and tension deviation rate Then, the Softmax output probability is used for judgment. When the maximum class probability exceeds the set confidence threshold and is consistent with the tension deviation rate criterion, the corresponding state is output. When the model confidence is insufficient or the classification result is inconsistent with the physical threshold, the state to be reviewed is output and the conservative early warning process is entered.

[0033] The system is based on the tension deviation rate The system implements a three-tiered progressive intervention based on the status classification results: The first tier is normal monitoring. When the tension deviation of each rope is within 5% of the average value and the status classification is "normal," the system maintains online monitoring, does not output intervention commands, and continuously records health data to the local database. The second tier is compliance warning. When the predicted tension of any single wire rope causes the real-time tension deviation to exceed 5% of the average value but does not reach the extreme looseness threshold, and the status classification is "tension out of tolerance," the system pushes a compliance warning message with precise coordinates and deviation values, prompting manual load equalization adjustment to prevent the potential danger of batch rope replacement due to long-term uneven stress. The third tier is millisecond-level emergency stop. When the status classification is "severe looseness" or an extremely high-risk "tripping precursor" characteristic pulse is detected, the system cuts off the elevator safety circuit or directly triggers the traction machine brake command within 50ms, and simultaneously sends an emergency fault message, including the fault type, occurrence time, car position, and tension data snapshot. In cases of sensor malfunction, missing data, low confidence in model output, or inconsistencies between the tension deviation rate and classification results, the system does not directly classify it as a normal state. Instead, it treats it as an abnormal state requiring verification, triggering a fault alert and initiating either a warning or conservative shutdown procedure based on the risk level. After the safety circuit is disconnected, the system enters a locked state. It can only be unlocked and resumed operation after professional maintenance personnel have reset it on-site and confirmed that the tension has returned to equilibrium.

[0034] Specific working principle: The first step is multi-source signal acquisition and preprocessing: radial and axial vibration signals of multiple wire ropes are simultaneously acquired by a high-frequency MEMS vibration sensor array deployed at the rope end assembly; car load signals are acquired by a car bottom load sensor placed at the load-bearing support point at the bottom of the car; and car position signals are acquired by a high-precision absolute encoder placed at the traction machine spindle end or the speed limiter wheel axle. The acquired multi-source heterogeneous signal data are processed using the Laida criterion to remove electromagnetic interference extreme noise points, and Z-Score normalization is used to eliminate dimensional differences. The vibration, load, and position signals are aligned by timestamps and framed at fixed lengths as the basic unit for subsequent analysis.

[0035] The second step is multi-dimensional operational excitation decoupling: The first level is load transient time-domain decoupling. During the initialization phase, the system determines the impulse threshold and derivative peak threshold based on different loads, operating speeds, passenger walking or jumping disturbances, normal full-journey operation, and calibrated tension imbalance samples. It calculates the first derivative of the load signal from the weighing sensor in real time to obtain the load change rate and constructs a sliding energy integral window to quantify the coupling strength between load mutation and vibration response. When the sliding energy is lower than the impulse threshold and the load change rate exceeds the peak threshold, it is identified as a pseudo-fault pulse caused by passenger walking or jumping in the car and recorded. When the sliding energy is lower than the impulse threshold and the load change rate does not exceed the peak threshold, it is identified as normal load fluctuation. When the sliding energy exceeds the impulse threshold and the high-frequency vibration envelope of the wire rope is continuously abnormal, it is identified as a candidate signal of wire rope tension imbalance. When the load change and vibration envelope are abnormal at the same time and it is impossible to directly distinguish between load disturbance and wire rope abnormality, it enters the second level of spatial position joint verification. The second level is spatial position joint verification. During the system initialization phase, the absolute position coordinates of each guide rail joint are automatically learned and calibrated through multiple full-stroke runs, establishing a guide rail joint position fingerprint database. The above candidate signals are combined with the car absolute displacement coordinates output in real time by the high-precision absolute encoder for spatial distribution statistics. If the car absolute displacement coordinates of the candidate signals are fixed within the same guide rail joint tolerance range at the time of the abnormal occurrence in multiple consecutive up and down strokes, it is determined to be an occasional environmental excitation caused by mechanical impact of the guide rail joint and is excluded. If the abnormal vibration coordinates of the candidate signals continue to reproduce in the full stroke in the non-joint area and the joint reproduction probability exceeds the set threshold, it is diagnosed as a real wire rope tension imbalance sequence.

[0036] The third step is symplectic geometric topology reconstruction: the phase space reconstruction and symplectic geometric mode decomposition of the real wire rope tension imbalance sequence output in the second step are performed. The one-dimensional vibration sequence is mapped to the Hankel trajectory matrix according to the embedding dimension. The covariance matrix is ​​constructed and solved by the Hamiltonian matrix to obtain the symplectic orthogonal transformation. The eigenvalue decomposition of the manifold matrix is ​​performed. According to the energy contribution rate of each modal component, the top few principal components with a cumulative contribution rate of more than the threshold are selected as the optimal string vibration principal component sequence. While filtering out the shaft wind resistance and the low-frequency structural noise of the traction machine, the nonlinear manifold topology is maintained. Then, the optimal principal component sequence is normalized and transformed to the polar coordinate system. The Gram angle field matrix is ​​constructed by the inner product of trigonometric functions. The one-dimensional principal time series is mapped to a two-dimensional topological feature image, and the implicit tension fluctuation is transformed into a spatial visual texture tensor.

[0037] The fourth step, Mamba-KAN tension prediction: The two-dimensional topological feature image generated in the third step is input into the Mamba encoder. A continuous-time state-space equation is defined and discretized using a zero-order hold to achieve a long-range temporal recursive approximation with linear complexity. A selective mechanism is used to dynamically adjust the state transition parameters based on the input content to extract the long-range temporal excitation dependence of multiple wire ropes. The deep feature tensor output by Mamba is then input into the KAN regression layer. A multi-dimensional superimposed learnable spline function is used to replace the fixed weight matrix of the traditional neural network. The nonlinear mapping relationship between the string vibration frequency and the physical tension is explicitly approximated by superimposing the basic activation function and the B-spline function. The dynamic tension quantification prediction value of a single wire rope and the discrete assessment of four health states are output in parallel. The four health states are determined based on the measured tension label, the tension deviation rate label, the manually calibrated fault category, and the deep vibration features output by the Mamba encoder. After the Softmax classifier outputs the probability of each state, the final state is determined by combining the physical threshold of the tension deviation rate and the confidence threshold. If the model has low confidence or the classification result conflicts with the physical threshold, a state to be reviewed is output.

[0038] Step 5, Closed-Loop Safety Intervention: The system receives the dynamic tension prediction value of each rope and the discrete health status assessment category from the dual-channel output of Step 4 in real time. It calculates the average tension of each suspension rope and the tension deviation rate for each rope. Based on the health status classification results, it executes a three-level progressive intervention: When the tension deviation of each rope is within 5% of the average and the status classification is normal, it maintains online monitoring and continuously records health data; when the tension deviation of any single wire rope exceeds 5% of the average but does not reach the extreme looseness threshold and the status classification is tension exceeding tolerance, it pushes a compliance warning message; when the status classification is severe looseness or a sign of impending trip, it cuts off the elevator safety circuit or triggers the traction machine brake command within 50ms. For undetermined states such as sensor anomalies, missing data, low confidence in model output, or inconsistencies between classification results and tension deviation rates, the system treats these as abnormal states awaiting verification, triggering fault prompts and entering a warning or conservative shutdown process based on the risk level. After the system enters a locked state, it must be reset on-site by professional maintenance personnel and confirmed that the tension has returned to equilibrium before the interlock can be released and operation resumed.

[0039] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the concept of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent monitoring of elevator wire rope tension that integrates symplectic geometric topological features and Mamba-KAN networks, characterized in that... Includes the following steps: S1. Multi-source signal acquisition and preprocessing: Radial and axial vibration signals of multiple steel wire ropes, car weighing load signals and car position signals are acquired by high-frequency MEMS vibration sensors, car bottom load weighing sensors and high-precision absolute encoders respectively; and the multi-source heterogeneous signal data are denoised and dimensionless normalized. S2. Multidimensional operation excitation decoupling: The first level is load transient time domain decoupling, which identifies candidate signals of wire rope tension imbalance by constructing a sliding energy integral window based on the load change rate of the car weighing load data; the second level is spatial position joint verification, which combines the above-mentioned candidate signals of wire rope tension imbalance with the absolute displacement coordinates of the car to perform spatial distribution statistics and then identify the real wire rope tension imbalance sequence. S3, Symmetric Geometric Topology Reconstruction: Perform symmetric geometric mode decomposition and phase space reconstruction on the real wire rope tension imbalance sequence, extract the principal component of the string vibration, and map the one-dimensional principal component sequence into a two-dimensional topological feature image through the Gram angle field algorithm. S4, Mamba-KAN tension prediction: The two-dimensional topological feature image is input into the Mamba encoder to extract long-range temporal dependencies, and then the KAN regression layer fits the nonlinear mapping between the string vibration frequency and the physical tension through a learnable spline function, outputting the single rope tension prediction value and health status classification. S5. Closed-loop safety intervention: Calculate the tension deviation rate based on the single rope tension prediction value, and implement a three-level progressive intervention of continuous monitoring, early warning push, and cutting off the safety loop in combination with the health status classification results.

2. The intelligent monitoring method for elevator wire rope tension based on symplectic geometric topological features and Mamba-KAN network as described in claim 1, characterized in that: The preprocessing process in step S1 includes using the Laida criterion to remove electromagnetic interference extreme noise points and using Z-Score normalization to eliminate dimensional differences.

3. The intelligent monitoring method for elevator wire rope tension based on symplectic geometric topological features and Mamba-KAN network as described in claim 1, characterized in that: The first-level load transient time-domain decoupling in step S2 includes calculating the first-order time derivative of the load signal from the weighing sensor to obtain the load change rate, constructing a sliding energy integral window to quantify the coupling strength between load mutation and vibration response, and setting an impulse threshold and a derivative peak threshold as discrimination boundaries. When the sliding energy is lower than the impulse threshold and the load change rate exceeds the derivative peak threshold, it is determined to be a pseudo-fault pulse caused by passengers walking or jumping in the car and is eliminated. When the sliding energy is stable and the high-frequency vibration envelope of the wire rope is continuously abnormal, it is identified as a candidate signal of wire rope tension imbalance.

4. The intelligent monitoring method for elevator wire rope tension based on symplectic geometric topological features and Mamba-KAN network as described in claim 1, characterized in that: The second-level spatial position joint verification in step S2 includes combining the candidate signals of wire rope tension imbalance output from the first level with the absolute displacement coordinates of the car to perform spatial distribution statistics, setting the tolerance window range at the guide rail joint, and establishing a guide rail joint position fingerprint database. If the candidate signal is fixed within the same guide rail joint tolerance range in multiple consecutive up and down strokes, it is determined to be an occasional environmental excitation caused by mechanical impact of the guide rail joint and is excluded. If the candidate signal continues to reappear in the non-joint area throughout the entire stroke and the joint reappearance probability exceeds the set probability threshold, it is confirmed as a real wire rope tension imbalance sequence.

5. The intelligent monitoring method for elevator wire rope tension based on symplectic geometric topological features and Mamba-KAN network as described in claim 1, characterized in that: The symplectic geometric mode decomposition and phase space reconstruction in step S3 include mapping the real wire rope tension imbalance sequence to a Hankel trajectory matrix according to the embedding dimension, constructing a covariance matrix and solving the symplectic orthogonal transformation through the Hamiltonian matrix, performing eigenvalue decomposition on the manifold matrix, and selecting the top few principal components whose cumulative contribution rate reaches a threshold or above as the optimal string vibration principal component sequence based on the energy contribution rate of each modal component, thus maintaining the nonlinear manifold topology while filtering out background noise.

6. The intelligent monitoring method for elevator wire rope tension based on symplectic geometric topological features and Mamba-KAN network as described in claim 1, characterized in that: The Gram angle field algorithm mapping in step S3 includes normalizing the optimal string vibration principal component sequence and converting it to the polar coordinate system, constructing the Gram angle field matrix through the inner product of trigonometric functions, mapping the one-dimensional principal component time series into a two-dimensional topological feature image, and converting the implicit tension fluctuations into a spatial visual texture tensor.

7. The intelligent monitoring method for elevator wire rope tension based on symplectic geometric topological features and Mamba-KAN network as described in claim 1, characterized in that: The Mamba encoder in step S4 includes defining a continuous-time state-space equation and discretizing it through a zero-order hold to achieve a long-range temporal recursive approximation with linear complexity. It also uses a selective mechanism to dynamically adjust the state transition parameters according to the input content and extract the long-range temporal excitation dependence of multiple wire ropes.

8. The intelligent monitoring method for elevator wire rope tension based on symplectic geometric topological features and Mamba-KAN network as described in claim 1, characterized in that: The KAN regression layer in step S4 uses a multidimensional superimposed learnable spline function to replace the fixed weight matrix of the traditional neural network. By superimposing the basic activation function and the B-spline function, it explicitly approximates the nonlinear mapping relationship between the string vibration frequency and the physical tension, so that the theoretical model calculation data and the experimental measurement data are in high agreement.

Images