Elevator fault prediction method and system

Abstract

The invention relates to the technical field of elevator fault prediction, in particular to an elevator fault prediction method and system. The method comprises the following steps: collecting an operation vibration signal of an elevator brake through a vibration sensor, performing frequency domain conversion, performing vibration frequency disorder structure analysis, analyzing abnormal vibration frequency intensity, analyzing elevator braking force increment loss, and obtaining time sequence increment data; thirdly, incremental gradient nonlinear induction is carried out on the time sequence incremental data, and an elevator fault prediction model is constructed based on a K-nearest neighbor algorithm; the elevator fault prediction technology is optimized, so that the elevator fault prediction technology is more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of elevator fault prediction, and in particular to an elevator fault prediction method and system. Background Art

[0002] Elevators generate various types of vibration signals during operation. These signals reflect the operating status of various elevator components, including the braking system, drive system, and door control system. Real-time monitoring and analysis of these vibration signals can effectively identify abnormal elevator conditions. High-precision vibration sensors collect vibration signals in real time during elevator operation. Combined with advanced signal processing methods such as frequency and time domain analysis, more effective information can be extracted from complex vibration signals, enabling more accurate prediction of potential elevator failures. Furthermore, using machine learning algorithms to analyze and model vibration data enables big data-based elevator failure prediction, improving both accuracy and real-time performance. However, traditional elevator failure prediction methods suffer from low accuracy in analyzing the loss of braking force caused by abnormal braking vibration, resulting in large errors in the accuracy of elevator failure prediction. Summary of the Invention

[0003] Based on this, it is necessary to provide an elevator fault prediction method and system to solve at least one of the above technical problems.

[0004] To achieve the above object, a method for predicting elevator failures is provided, the method comprising the following steps: Step S1: Deploy a vibration sensor on the elevator brake to collect the running vibration signal, and then perform frequency domain conversion to obtain running vibration cleaning frequency domain data; perform vibration frequency disorder structure analysis on the running vibration cleaning frequency domain data to obtain vibration frequency disorder structure data; Step S2: performing abnormal vibration frequency intensity analysis based on the vibration frequency disordered structure data to obtain abnormal vibration frequency intensity data; performing elevator braking force incremental loss analysis based on the abnormal vibration frequency intensity data to obtain braking force loss time series incremental data; Step S3: performing incremental gradient nonlinear induction on the braking force loss time series incremental data to obtain loss incremental gradient nonlinear induction data; constructing an elevator fault prediction model on the loss incremental gradient nonlinear induction data based on the K-nearest neighbor algorithm to obtain an elevator fault prediction model; and sending the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

[0005] Preferably, step S1 includes the following steps: Step S11: deploying a vibration sensor on the elevator brake to collect an operating vibration signal to obtain an elevator brake operating vibration signal; Step S12: performing signal cleaning processing on the elevator brake operation vibration signal to obtain a brake operation vibration cleaning signal; Step S13: performing frequency domain conversion on the brake operation vibration cleaning signal to obtain operation vibration cleaning frequency domain data; Step S14: performing vibration frequency disorder structure analysis on the frequency domain data of the vibration cleaning operation to obtain vibration frequency disorder structure data.

[0006] Preferably, step S2 includes the following steps: Step S21: performing abnormal vibration frequency intensity analysis on the frequency domain data of the vibration cleaning operation according to the vibration frequency disorder structure data to obtain abnormal vibration frequency intensity data; Step S22: performing rotor shaft deformation offset gradient calculation based on the abnormal vibration frequency intensity data to obtain rotor shaft deformation offset gradient data; Step S23: performing an elevator braking force incremental loss analysis based on the rotor shaft deformation offset gradient data to generate braking force incremental loss data; Step S24: performing time series incremental regression analysis on the braking force incremental loss data to obtain braking force loss time series incremental data.

[0007] Preferably, step S22 includes the following steps: Step S221: identifying the rotor shaft vibration energy conduction vector based on the abnormal vibration frequency intensity data to obtain the rotor shaft vibration energy conduction vector; Step S222: performing radial vibration energy geometric fluctuation analysis on the rotor shaft vibration energy conduction vector to obtain radial vibration energy geometric fluctuation data; Step S223: performing rotor shaft cross-section ellipse energy deviation calculation based on radial vibration energy geometric fluctuation data to obtain cross-section ellipse energy deviation data; Step S224: performing rotor shaft cyclic torsional stress coupling based on the cross-sectional ellipse energy deviation data and the radial vibration energy geometric fluctuation data to obtain cyclic torsional stress coupling data; Step S225: obtaining basic material fatigue limit parameters of the rotor shaft; performing rotor shaft deformation offset gradient calculation on the basic material fatigue limit parameters according to the cyclic torsional stress coupling data to obtain rotor shaft deformation offset gradient data.

[0008] Preferably, step S23 includes the following steps: Step S231: performing radial / axial offset skewness analysis based on the rotor shaft deformation offset gradient data to obtain radial / axial offset skewness data; Step S232: performing torsional deformation gradient analysis on the radial / axial offset skewness data to obtain torsional deformation gradient data; Step S233: calculating the rotor shaft deformation curvature radius partition variance based on the torsion angle deformation gradient data to obtain the deformation curvature radius partition variance; Step S234: Deducing the braking torque loss index based on the variance of the deformation curvature radius partitions to obtain the braking torque loss index; Step S235: performing an elevator braking force incremental loss analysis based on the braking torque loss index to generate braking force incremental loss data.

[0009] Preferably, step S234 includes the following steps: The braking rotation centrifugal force imbalance fluctuation index is calculated according to the variance of the deformation curvature radius partition, and the braking rotation centrifugal force imbalance fluctuation index is obtained; The braking rotation axial force change rate is calculated proportionally based on the variance of the deformation curvature radius partition, and the proportional data of the axial force change rate is obtained; Perform logarithmic linearization on the proportional data of the axial force variation rate to obtain the logarithmic linear data of the force variation; The braking torque loss index is deduced based on the braking rotation centrifugal force imbalance fluctuation index and the logarithmic linear data of the drift force change to obtain the braking torque loss index.

[0010] Preferably, step S24 includes the following steps: Step S241: analyzing the loss increment time series relationship of the braking force increment loss data to obtain the loss increment time series relationship; Step S242: performing trend increment heteroscedasticity analysis on the loss increment time series relationship to generate loss trend increment heteroscedasticity features; Step S243: performing a change-point lognormal distribution analysis on the loss increment time series relationship based on the loss trend increment heteroscedasticity feature to obtain change-point lognormal distribution data; Step S244: performing time series incremental regression analysis on the change point log-normal distribution data to obtain time series incremental data of braking force loss.

[0011] Preferably, step S3 includes the following steps: Step S31: normalizing the braking force loss time series increment data to obtain normalized braking force loss time series increment data; Step S32: performing incremental gradient nonlinear induction on the normalized data of the braking force loss time series increment to obtain loss incremental gradient nonlinear induction data; Step S33: constructing an elevator fault prediction model based on the loss increment gradient nonlinear induction data based on the K nearest neighbor algorithm to obtain an elevator fault prediction model; Step S34: Send the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

[0012] Preferably, step S32 includes the following steps: Step S321: Drawing a loss time series increment curve for the normalized braking force loss time series increment data to generate a loss time series increment curve; Step S322: performing a logarithmic difference calculation of the amplitude change ratio of adjacent segments on the loss time series incremental curve to obtain the logarithmic difference of the amplitude change ratio; Step S323: performing time series reverse slip accumulation on the loss time series increment curve according to the logarithmic difference of the amplitude change ratio to obtain loss time series reverse slip accumulation data; Step S324: performing incremental gradient nonlinear induction according to the loss time series reverse slip accumulation data to obtain loss incremental gradient nonlinear induction data.

[0013] Preferably, the present invention further provides an elevator fault prediction system for executing the elevator fault prediction method described above, the elevator fault prediction system comprising: The vibration frequency disorder structure analysis module is used to deploy a vibration sensor on the elevator brake to collect the running vibration signal, and then perform frequency domain conversion to obtain the running vibration cleaning frequency domain data; the vibration frequency disorder structure analysis is performed on the running vibration cleaning frequency domain data to obtain the vibration frequency disorder structure data; The braking force incremental loss analysis module is used to analyze the abnormal vibration frequency intensity based on the vibration frequency disordered structure data to obtain abnormal vibration frequency intensity data; based on the abnormal vibration frequency intensity data, the elevator braking force incremental loss analysis is performed to obtain the braking force loss time series incremental data; The fault prediction model construction module is used to perform incremental gradient nonlinear induction on the braking force loss time series incremental data to obtain loss incremental gradient nonlinear induction data; construct an elevator fault prediction model based on the loss incremental gradient nonlinear induction data based on the K nearest neighbor algorithm to obtain an elevator fault prediction model; and send the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

[0014] The present invention has the beneficial effect of deploying vibration sensors on elevator brakes, enabling real-time acquisition of vibration signals during elevator operation. This method effectively monitors the elevator's dynamic state, particularly the operating condition of the brake system, and can capture subtle changes in elevator operation. Frequency-domain conversion of the acquired signals produces operational vibration cleansing frequency domain data, which helps eliminate noise and interference, improves the signal-to-noise ratio, and yields clearer and more accurate vibration signatures. Based on this, vibration frequency disorder structure analysis reveals the complex frequency distribution and disorder characteristics of the vibration signal, providing accurate baseline data for subsequent fault diagnosis and prediction. Analysis of the vibration frequency disorder structure data allows the extraction of abnormal vibration frequency intensities during elevator operation. This data analysis not only identifies potential anomalies in the elevator system but also distinguishes between normal and abnormal vibration patterns, enabling timely detection of fault signs within the elevator system. Based on this abnormal vibration frequency intensity data, further analysis of the incremental loss of braking force can quantify the gradual loss of braking force during operation and reveal the performance degradation of the elevator's braking system. This is crucial for predicting elevator failures, as braking force loss is directly related to elevator safety and reliability. Early detection of these changes can effectively prevent sudden failures during high-load operation. Applying incremental gradient nonlinear summarization to the time-series incremental data of braking force loss can deeply explore the regularity of braking force loss changes. Nonlinear analysis of this data can also identify potential factors affecting elevator performance. This analytical approach captures the nonlinear relationships and complex dynamic behavior of elevator systems, helping to establish more accurate fault prediction models. Processing this incremental gradient nonlinear summarization data using the K-nearest neighbor algorithm allows efficient prediction of elevator failures using historical data, identifying failure trends in advance and reducing elevator downtime and repair costs. Once the elevator fault prediction model is delivered to the terminal, real-time monitoring of the elevator status is achieved, ensuring safer and more stable elevator operation. This data-driven approach further improves the accuracy and response speed of elevator fault warnings. Therefore, the present invention is an optimization process for a traditional elevator fault prediction method, which solves the problem that the traditional elevator fault prediction method has low accuracy in analyzing the braking force loss caused by abnormal braking vibration, thereby causing large errors in the accuracy of elevator fault prediction. It improves the accuracy of the analysis of the braking force loss caused by abnormal braking vibration and reduces the accuracy error of elevator fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic flow chart of the steps of an elevator fault prediction method; Figure 2 for Figure 1 Detailed implementation steps of step S2 in FIG. Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG. DETAILED DESCRIPTION

[0016] See also Figures 1 to 3 , a method for predicting elevator failures, the method comprising the following steps: Step S1: Deploy a vibration sensor on the elevator brake to collect the running vibration signal, and then perform frequency domain conversion to obtain running vibration cleaning frequency domain data; perform vibration frequency disorder structure analysis on the running vibration cleaning frequency domain data to obtain vibration frequency disorder structure data; Step S2: performing abnormal vibration frequency intensity analysis based on the vibration frequency disordered structure data to obtain abnormal vibration frequency intensity data; performing elevator braking force incremental loss analysis based on the abnormal vibration frequency intensity data to obtain braking force loss time series incremental data; Step S3: performing incremental gradient nonlinear induction on the braking force loss time series incremental data to obtain loss incremental gradient nonlinear induction data; constructing an elevator fault prediction model on the loss incremental gradient nonlinear induction data based on the K-nearest neighbor algorithm to obtain an elevator fault prediction model; and sending the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

[0017] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of an elevator fault prediction method according to the present invention. In this example, the elevator fault prediction method includes the following steps: Step S1: Deploy a vibration sensor on the elevator brake to collect the running vibration signal, and then perform frequency domain conversion to obtain running vibration cleaning frequency domain data; perform vibration frequency disorder structure analysis on the running vibration cleaning frequency domain data to obtain vibration frequency disorder structure data; In the embodiment of the present invention, a vibration sensor is selected and fixedly installed on the outer surface of the bearing shell of the elevator brake body by means of a threaded connection. The installation position is 10 mm to the left of the middle section of the brake rotor shaft to ensure high sensitivity capture of the vibration signal during the braking process. When the elevator frequently starts and stops, the signal is collected at a sampling rate of 48 kHz and the sampling duration is 120 seconds. The collected original vibration signal is filtered by a Butterworth bandpass filter with a filter bandwidth set to 20 Hz to 8 kHz to remove the influence of power supply interference and background noise. The processed signal is then processed by a fast Fourier transform (FFT). Frequency domain conversion. After conversion, the frequency domain data is short-term cleaned using a multi-channel peak hold algorithm with a cleaning cycle of 500ms, and the main peak frequency and its harmonic components are retained. Finally, the running vibration cleaning frequency domain data is obtained. The frequency domain data is input into the vibration frequency disordered structure identification program based on the multi-scale entropy (MSE) analysis method. By calculating the change trend of sample entropy at different scales, the vibration frequency disordered structure areas in different frequency bands are identified, and the intrinsic mode function (IMF) in the local frequency band is extracted in combination with the empirical mode decomposition (EMD) to confirm the randomness enhancement area of ​​the frequency domain energy distribution, and finally the vibration frequency disordered structure data is output.

[0018] Step S2: performing abnormal vibration frequency intensity analysis based on the vibration frequency disordered structure data to obtain abnormal vibration frequency intensity data; performing elevator braking force incremental loss analysis based on the abnormal vibration frequency intensity data to obtain braking force loss time series incremental data; In an embodiment of the present invention, the obtained vibration frequency disordered structure data is analyzed for abnormal vibration frequency intensity of the original frequency domain data based on the continuous wavelet transform (CWT). The Morlet mother wavelet function is selected to enhance the detection sensitivity of high-frequency changes. The scale factor range is set to 1 to 128. A local frequency interval intensity distribution map is constructed using a multi-scale energy density spectrum. The local maximum amplitude change gradient of the energy abnormal mutation region is calculated to form abnormal vibration frequency intensity data. Based on this intensity data, a conventional difference method is used to construct the rate of change of the vibration intensity time series. Further, combined with the dynamic braking force curve measured by the elevator brake, a corresponding relationship between abnormal intensity points and brake torque attenuation changes is established. By constructing a first-order difference ratio sequence of braking force changes and frequency intensity changes, the braking force increment change value is extracted. Then, the sliding window time series cumulative difference method is used to analyze its change trend. The window width is 30 data points and the sliding step size is 5 data points. Finally, the braking force loss time series incremental data is obtained.

[0019] Step S3: performing incremental gradient nonlinear induction on the braking force loss time series incremental data to obtain loss incremental gradient nonlinear induction data; constructing an elevator fault prediction model on the loss incremental gradient nonlinear induction data based on the K-nearest neighbor algorithm to obtain an elevator fault prediction model; and sending the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

[0020] In the embodiment of the present invention, the time series incremental data of the braking force loss is normalized, and the Z-score normalization method is used to convert all sample values ​​into a data distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the influence of the numerical dimension on subsequent analysis. The normalized data is used to construct a time series incremental change curve, and the segmented regression analysis method is applied to extract the local slope change of the curve. Each segment is set to a fixed length of 50 sampling points, and the logarithmic difference of the amplitude change ratio between adjacent segments is calculated to capture the sections in the curve showing a nonlinear transition trend. The difference sequence is further subjected to cumulative reverse slip processing, and the reverse time window reconstruction method is used to reversely accumulate and compress the forward curve change trend to form a slip accumulation sequence. , and a nonlinear incremental piecewise curvature estimation function is established based on the amplitude gradient of the change point to obtain the loss incremental gradient nonlinear induction data. Finally, the K nearest neighbor classification algorithm is used to construct an elevator fault prediction model. The K value is selected as 7. The training sample data is taken from 50 groups of historical elevator brake abnormal failure samples. The input features are the nonlinear incremental trend value and reverse slip cumulative value in the induction data. The distance measurement method uses Euclidean distance, and a weighting function is introduced to give inverse weights to neighboring samples. The state classification label of the current sample is obtained by the majority voting method. After training, the model is saved in binary form in the local database and sent to the elevator control terminal through the TCP communication protocol to complete the model deployment, so as to realize real-time elevator fault prediction function.

[0021] Step S1 includes the following steps: Step S11: deploying a vibration sensor on the elevator brake to collect an operating vibration signal to obtain an elevator brake operating vibration signal; Step S12: performing signal cleaning processing on the elevator brake operation vibration signal to obtain a brake operation vibration cleaning signal; Step S13: performing frequency domain conversion on the brake operation vibration cleaning signal to obtain operation vibration cleaning frequency domain data; Step S14: performing vibration frequency disorder structure analysis on the frequency domain data of the vibration cleaning operation to obtain vibration frequency disorder structure data.

[0022] In this embodiment of the present invention, a vibration sensor is fixedly installed at the center of the top of the elevator brake housing using an M6 threaded fastening structure. The installation point is 12 mm from the axis of the brake rotor shaft and perpendicular to the ground to ensure the collection of composite vibration signals in the axial and radial directions while preventing interference with the sensor caused by the rotation of the motor main shaft. The sampling equipment uses the NI PXIe-4492 data acquisition card, with a sampling frequency of 48 kHz, a data accuracy of 24 bits, and a sampling duration of 90 seconds. The collection time is selected within the complete cycle of the elevator's startup, operation, and braking stages from static to running. Sampling is repeated three times for each cycle to enhance data stability and cover all operating conditions. Finally, the total number of raw operating vibration signals obtained in each cycle is 12,960,000 data points. After collection, the signal is exported to a CSV format file and saved as a raw vibration data file. The original vibration data file is read and the signal is cleaned in three stages using the MATLAB signal processing toolbox. First, a fifth-order Butterworth bandpass filter is applied for filtering. The filter passband is set to 30Hz to 6500Hz, and the upper and lower cutoff frequencies of the filter are set to 25Hz and 7000Hz respectively to filter out the power supply frequency interference and the nonlinear response of the high-frequency sensor. The second step is to correct the baseline drift. The local average value under the 200-point sliding window is calculated by the sliding average method, and then the baseline is zeroed by subtracting the sliding average from the original signal. The third step is to remove burrs. Through continuous third-order difference detection, data points with differential values ​​exceeding three times the standard deviation of the local mean are identified, and these data points are interpolated and replaced with the neighboring average values. After three cleaning steps, the brake operation vibration cleaning signal with noise interference removed is obtained. The cleaned time domain signal is input into the frequency domain transformation process and transformed using the fast Fourier transform algorithm. The number of FFT transformation points is set to 131072 points. To meet the point number requirement, the signal length is zero-padded. The result after transformation is a spectrum amplitude array with a frequency resolution of 0.366Hz. The spectrum amplitude is processed using logarithmic normalization to enhance the identifiability of weak frequency components, and multi-segment frequency division processing is used to divide the spectrum into six segments, namely 30-200Hz, 200-800Hz, 800-1500Hz, 1500-3000Hz, 3000-5000Hz and 5000-6500Hz. The local peak frequency and energy distribution characteristics of each segment are extracted as subset records, and the output overall frequency domain data structure is named running vibration cleaning frequency domain data.The multi-scale entropy analysis method is used to identify the vibration frequency disorder structure of the running vibration cleaning frequency domain data. The number of entropy calculation scales is set to 1 to 20, the sampling point reconstruction dimension is set to 3, and the tolerance coefficient is set to 0.15 times the sample standard deviation. By calculating the relative entropy difference between the sample entropy value and the benchmark ordered spectrum sample at each scale, the frequency bands with high-frequency mutation and energy random distribution characteristics are identified, and these frequency bands are located as frequency disorder regions. The instantaneous frequency fluctuation trajectories in these frequency bands are reconstructed by combining the Hilbert transform, and the local frequency transition density and the power density change rate within the frequency band are analyzed. When the transition density exceeds the threshold of 1.5 times / Hz and the power change rate is greater than 0.4dB / Hz, it is judged to be a high disorder region. Finally, the vibration frequency disorder structure data including the start and end frequencies, main peak frequencies, energy distribution and transient transition rates of each frequency disorder segment are output.

[0023] Step S2 includes the following steps: Step S21: performing abnormal vibration frequency intensity analysis on the frequency domain data of the vibration cleaning operation according to the vibration frequency disorder structure data to obtain abnormal vibration frequency intensity data; Step S22: performing rotor shaft deformation offset gradient calculation based on the abnormal vibration frequency intensity data to obtain rotor shaft deformation offset gradient data; Step S23: performing an elevator braking force incremental loss analysis based on the rotor shaft deformation offset gradient data to generate braking force incremental loss data; Step S24: performing time series incremental regression analysis on the braking force incremental loss data to obtain braking force loss time series incremental data.

[0024] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes: Step S21: performing abnormal vibration frequency intensity analysis on the frequency domain data of the vibration cleaning operation according to the vibration frequency disorder structure data to obtain abnormal vibration frequency intensity data; In an embodiment of the present invention, the operating vibration cleaning frequency domain data is analyzed section by section according to the high disordered frequency band interval identified in the obtained vibration frequency disordered structure data. The spectrum data of each frequency band is used as input, and the transient enhanced frequency points in the disordered segment are identified by peak tracking and frequency energy curve fitting methods. Peak tracking adopts the first-order difference of the signal and the sliding window extreme value matching method. The window size is set to 512 frequency points and the moving step is 64 frequency points. The fitting method adopts cubic spline curve regression. The local energy of the spectrum line envelope in each sliding window is estimated and the abnormal enhancement points are identified. The frequency points whose energy change rate is greater than 1.7 times the local mean change rate of the conventional operating spectrum line are marked as abnormal frequency points, and their corresponding frequency coordinates and logarithmic amplitudes are recorded at the same time. Finally, the abnormal frequency set in all disordered frequency bands is extracted to construct abnormal vibration frequency intensity data. The data structure includes the frequency point position, transient energy rise rate, change duration and spectrum envelope nonlinear fitting residual.

[0025] In another embodiment, when performing abnormal vibration frequency intensity analysis on the frequency domain data of the running vibration cleaning based on the vibration frequency disorder structure data, the frequency domain data is first divided into 128 equal-width frequency bands, covering the range of 0 Hz to 2000 Hz, and the width of each frequency band is 15.625 Hz. Then, a fast Fourier transform with a 1024-point Hanning window and an overlap rate of 50% is applied to improve the frequency resolution. Then, a statistical deviation threshold analysis is performed on the frequencies in each frequency band, and frequencies with amplitude values ​​exceeding 3.5 times the standard deviation of the average amplitude of the frequency band are marked as abnormal. Then, a frequency intensity ratio is calculated by dividing the amplitude of each abnormal frequency by the standard deviation of its frequency band. Average amplitude, these ratios are then weighted according to their proximity to the mechanical resonance frequency of the elevator braking system. The weights are determined by a logarithmic distance function, where frequencies within ±5 Hz of the known resonance point receive a weight of 0.9 to 1.0, while frequencies exceeding ±50 Hz receive a weight less than 0.3. The resulting abnormal vibration frequency intensity data is represented as a vector containing frequency values ​​(Hz), normalized amplitudes (dimensionless values ​​between 0 and 1), and calculated intensity ratios (typically multiplication factors between 1.0 and 10.0). This data structure simultaneously captures the location and severity of abnormal vibration in the frequency domain, laying the foundation for subsequent mechanical deformation analysis.

[0026] Step S22: performing rotor shaft deformation offset gradient calculation based on the abnormal vibration frequency intensity data to obtain rotor shaft deformation offset gradient data; In this embodiment of the present invention, abnormal vibration frequency intensity data is used as input. A mapping relationship between the dynamic deformation response of the rotor shaft is established based on the frequency response of the mechanical system and the dynamic relationship between the rigid body shaft system. Referring to the structural parameters of the elevator brake, the rotor shaft material is set to 45# steel, the diameter is 32mm, the cantilever length is 160mm, and the end connection plate diameter is 120mm. An axial deformation response function is constructed based on the empirical curve of structural modal analysis. The position of the abnormal frequency point and its energy enhancement amplitude are further combined to simulate the total shaft end deformation caused by frequency disturbance using a piecewise integral accumulation superposition method. The influence range of each abnormal frequency point is set to the ±5Hz interval. The effective spectral energy density within this interval is calculated and multiplied by the equivalent offset coefficient obtained based on modal density fitting. The contribution of each frequency point to the shaft deformation is finally obtained. After summarization, the gradient offset trajectory of the shaft center in the spatial coordinate direction is reconstructed according to the frequency segment distribution. The output is the rotor shaft deformation offset gradient data. The data structure includes the three-axis offset, the frequency contribution source, the frequency distribution mean, and the second-order difference gradient value of the offset increment.

[0027] In another embodiment, in the process of calculating the rotor shaft deformation offset gradient based on abnormal vibration frequency intensity data, the abnormal vibration frequency intensity data is first converted into a vibration energy vector through energy density mapping. Each abnormal frequency point is assigned an energy contribution value, which is calculated by multiplying the square of its amplitude by the frequency value to obtain an energy value in J / kg. These energy vectors are decomposed into radial and axial components using a coordinate transformation matrix based on the physical installation direction of the vibration sensor. The sensor is located at 0°, 90°, 180° and 270° on the circumference of the rotor shaft. The measurement accuracy is ±0.01mm. The radial energy distribution is calculated by several The He wave pattern recognition algorithm is analyzed. The algorithm identifies wave nodes that appear at specific rotation angles. The node detection sensitivity is set to identify energy changes of at least 15% between adjacent points. The cross-sectional energy distribution is mapped to an elliptical model, where the major and minor axes represent the principal stress directions. The ratio between these axes (a perfect circle is 1.0, and the deformed axis is typically between 1.05 and 1.30) quantifies the degree of asymmetric deformation. The cyclic torsional stress coupling calculation combines these elliptical deviations with the radial energy waveform pattern through a tensor product operation to generate a stress field map with a circumferential resolution of 5° and an axial resolution of 10 mm. The fatigue limit parameters of the shaft material (typically SAE 4140 steel with a fatigue limit of 615 MPa and a number of cycles of 10^7) are used to normalize the stress field. The central difference method is used for gradient analysis with a spatial step size of 5 mm to generate a deformation gradient vector in μm / mm to characterize the spatial variation of the shaft deformation.

[0028] Step S23: performing an elevator braking force incremental loss analysis based on the rotor shaft deformation offset gradient data to generate braking force incremental loss data; In an embodiment of the present invention, the three-axis offset information in the rotor shaft deformation offset gradient data is used in combination with the mechanical contact pressure distribution law between the brake friction pad and the brake disc to deduce the braking force change process. The axis center offset is converted into the radial compression unevenness coefficient of the contact surface. According to the typical structure of the elevator brake, the initial state of the contact surface is uniform annular contact. The initial average unit area pressure of the braking force is set to 0.35 MPa. Under the bias pressure caused by the offset, the contact force field discrete integration method is used to calculate the contact surface pressure distribution function under different offsets. The total contact force change is solved for the pressure field change result using the multi-segment area method. The braking force reduction increment function corresponding to each unit offset gradient value is constructed. By sequentially traversing the offset caused by all frequency point sources and the corresponding contact force change, the overall braking force incremental loss data under the current vibration state is obtained. The output result is a braking force change sequence synchronized with the timestamp and accompanied by the offset source and corresponding influence weight.

[0029] In another embodiment, in the process of analyzing the incremental loss of braking force of an elevator based on the rotor shaft deformation offset gradient data, the rotor shaft deformation offset gradient data is first subjected to radial / axial offset skewness analysis using a sliding window with a 45° rotation window and a 50% overlap rate, and statistical moments are calculated to generate a skewness coefficient ranging from -2.5 to +2.5, where a negative value indicates deformation dominated by compression and a positive value indicates deformation dominated by tension. The torsional angle deformation gradient analysis uses a discrete differential method with an angle step of 2° to calculate the deformation change rate along the rotation direction, generating a gradient value generally in μm / degree. The gradient value of significant deformation exceeds 0.5 μm / degree. The rotor shaft deformation curvature radius partition variance calculation divides the shaft into 8 longitudinal segments and 12 For the circular segments, a three-point arc fitting algorithm is used to calculate the local curvature radius of each segment. The fitting error tolerance is ±0.5%. The difference is 0.001 to 0.1 under normal operating conditions and exceeds 0.15 under severe deformation conditions. The braking torque loss index is derived by utilizing these partitioned variances through a power law relationship, where the exponential term represents the sensitivity of torque loss to deformation and is typically between 1.2 and 1.8 depending on the brake mechanism design. Higher values ​​indicate greater sensitivity to deformation. The incremental braking force loss data is generated by applying this index to the nominal braking force value, considering the cumulative effect of deformation in repeated braking cycles. 120 points are sampled over a full rotation cycle, producing percentage loss values ​​ranging from 0.5% in the early degradation stage to 15% under critical pre-failure conditions.

[0030] Step S24: performing time series incremental regression analysis on the braking force incremental loss data to obtain braking force loss time series incremental data.

[0031] In an embodiment of the present invention, the acquired braking force incremental loss data is processed using a time series incremental regression analysis method. The time window width is set to 3 seconds and the moving step is set to 1 second. The incremental change rate and change amplitude in each window are counted, and an incremental ratio sequence between consecutive windows is constructed. The sequence is processed by first-order difference processing and then linear regression analysis is performed to capture the continuous nonlinear incremental trend. The regression method adopts least squares fitting, and the objective function is to minimize the product of the incremental value and the time interval. At the same time, the second-order sliding standard deviation is used to evaluate the degree of dispersion of the regression residual in each time period, identify the mutation characteristics in the short term and the long-term trend evolution direction, and finally output the braking force loss time series incremental data. The data content includes the loss value per unit time, the residual sum of squares, the regression trend slope, and the length of the second-order derivative change sign interval.

[0032] In another embodiment, when performing time series incremental regression analysis on the incremental loss of braking force data, the temporal relationship of the incremental loss is first extracted by sequential difference operation with a time step of 8 hours and a monitoring period of 30 days, creating a time series containing 90 data points to capture the acceleration or deceleration of the braking force degradation. The incremental loss time series relationship is then subjected to trend incremental heteroskedasticity analysis using a modified GARCH (generalized autoregressive conditional heteroskedasticity) process with parameters α=0.15 and β=0.75 to generate heteroskedasticity features to quantify the time-varying volatility in the degradation pattern. The variance value is typically in the range of 0.0001 to 0.05, and the change point lognormal distribution is analyzed using Bayesian The change point detection algorithm is applied to the time series relationship of loss increments. The prior probability of the change point occurrence is 0.05, the minimum segment length is 5 data points, and the log-normal distribution parameters μ (log mean) range from -3.5 to -1.2 and σ (log standard deviation) range from 0.2 to 0.8 are generated for each identified segment. The time series incremental regression analysis uses a piecewise exponential regression model for each segment identified in the previous step. The regression coefficient is calculated using weighted least squares. Newer data points are given exponentially higher weights than older points based on a decay factor of 0.95. The generated time series incremental data simultaneously captures the magnitude and rate of braking force loss, providing a 90% prediction confidence interval for the next 7 days of operation.

[0033] Step S22 includes the following steps: Step S221: identifying the rotor shaft vibration energy conduction vector based on the abnormal vibration frequency intensity data to obtain the rotor shaft vibration energy conduction vector; Step S222: performing radial vibration energy geometric fluctuation analysis on the rotor shaft vibration energy conduction vector to obtain radial vibration energy geometric fluctuation data; Step S223: performing rotor shaft cross-section ellipse energy deviation calculation based on radial vibration energy geometric fluctuation data to obtain cross-section ellipse energy deviation data; Step S224: performing rotor shaft cyclic torsional stress coupling based on the cross-sectional ellipse energy deviation data and the radial vibration energy geometric fluctuation data to obtain cyclic torsional stress coupling data; Step S225: obtaining basic material fatigue limit parameters of the rotor shaft; performing rotor shaft deformation offset gradient calculation on the basic material fatigue limit parameters according to the cyclic torsional stress coupling data to obtain rotor shaft deformation offset gradient data.

[0034] In the embodiment of the present invention, the main peak frequency points marked in the abnormal vibration frequency intensity data are used as frequency source inputs, and the vibration propagation path analysis is carried out in combination with the structural parameters of the rotor shaft in the elevator brake system. The rotor shaft is 480 mm long and 32 mm in diameter, and is made of 45# steel with a Young's modulus of 210 GPa and a density of 7850 kg per cubic meter. By combining the energy amplitude and frequency position corresponding to each frequency point and using the one-dimensional wave propagation model of the axisymmetric structure in the wave theory, the displacement vector propagation relationship along the axial direction is derived, and the vibration propagation velocity is introduced in the calculation process. The expression contains the specific functional form of the relationship between frequency, medium density and elastic modulus. At the same time, based on the phase delay and amplitude attenuation model of frequency domain energy propagation in the direction of the structural axis, an energy conduction vector matrix is ​​constructed. Each element in this vector matrix represents the unit energy flow density and phase angle change along the axial z direction at a specific frequency. The matrix dimension is N×3, where N is the number of frequency points and 3 is the corresponding three-axis energy vector distribution direction. Finally, the rotor shaft vibration energy conduction vector data is obtained, which records the directionality and spatial distribution path of the axial vibration displacement energy intensity changes caused by different frequencies. The rotor shaft vibration energy conduction vector obtained above is spatially projected and decomposed. Each vector is projected onto the radial plane perpendicular to the rotor shaft cross section, i.e., the XY plane, according to the right-hand coordinate system. All vectors are uniformly converted into the standard reference coordinate system using the geometric vector rotation transformation method, and their modulus is normalized. Then, a two-dimensional radial vibration energy distribution diagram is constructed. In this diagram, each data point represents the radial energy density corresponding to a frequency. By fitting the density distribution of the projection vectors corresponding to all frequency points, the roundness deviation is analyzed using the polar coordinate curve approximation method, and the energy concentration direction and fluctuation amplitude within the 360-degree angle range are further identified. The energy difference distribution value of each angular segment is calculated, and the maximum and minimum radial energy density differences within a sliding window with an angle range of 15 degrees are calculated. The angular change trend is recorded to finally form the radial vibration energy geometric fluctuation data. This data structure contains the radial equal-angle energy density, energy difference standard deviation, maximum offset angle and its corresponding energy density range.

[0035] Based on the acquired radial vibration energy geometric fluctuation data, the energy distribution deformation deviation of each section of the rotor shaft under dynamic excitation state is calculated using the ellipse fitting method. The fitting process adopts the least squares ellipse fitting method, and the radial distribution energy density is used as the fitting input value. It is mapped to the radial length of the ellipse at each angle point in the polar coordinate system, thereby forming a set of boundary points of the ellipse. The ellipse major axis is identified for this set, and the lengths of the major axis and minor axis as well as the major axis rotation angle are extracted. The ellipse shape factor and circular deviation measurement indicators are further calculated, such as eccentricity, the ratio of the ellipse area to the unit circle area and other parameters. The above indicators are summarized as ellipse energy deviation data, and the frequency correspondence and cross-sectional position index of each ellipse fitting result are recorded at the same time. The output data is named cross-sectional ellipse energy deviation data, which can be used for subsequent rotor shaft mechanical response analysis and form the basic basis for coupling analysis with torsional stress distribution. Based on the joint analysis of the above-mentioned elliptical energy deviation data and radial vibration energy geometric fluctuation data, a three-dimensional elliptical cross-section model is constructed at each axial section, and the main axis direction of the ellipse is mapped to the eccentric vibration action line. At the same time, based on the phase relationship of each frequency component in the conduction vector and the excitation period, an energy offset model under the time series is constructed. The periodic deformation variable is further equivalent to the cyclic shear deformation occurring within one cycle. The cyclic torsional stress distribution is calculated based on the relationship between the shear moment and the material deformation. The equivalent stress superposition method is used to calculate the periodically changing shear stress vector on each elliptical section. The input parameters include the polar moment of inertia of the circular section, the shear modulus of the material, the eccentricity of the ellipse, etc., and finally a three-dimensional stress tensor field is constructed. The time dimension is then subjected to Fourier analysis to extract the main cyclic frequency stress components, and the output is cyclic torsional stress coupling data. This data structure contains the torsional load peak value, stress amplitude, and shear stress distribution coefficient in the main axis direction experienced by different sections per unit time. First, the fatigue limit parameters of the rotor shaft material are obtained. These parameters can be obtained from material manuals or static fatigue tests. For 45 steel, the cyclic torsional fatigue limit at room temperature is approximately 200 MPa. With a safety factor of 1.5, and referring to the fatigue life standard SN curve parameters, the shear stress limit of this material at 10^6 cycles is 180 MPa. Then, the stress amplitude and corresponding frequency data of each point in the cyclic torsional stress coupling data are used as input. A stress-life curve is constructed by combining the number of cyclic loads and the stress amplitude. The damage accumulation factor corresponding to each stress amplitude is calculated. Miner's linear cumulative damage theory is used to integrate the damage factor per unit area of ​​all cross-sections. The strain gradient field is further constructed by mapping the shear stress offset to the local fatigue life change. Finally, the rotor shaft deformation offset gradient data is output. This data includes dimensional information such as the displacement gradient in the three axes, the shear stress amplitude distribution gradient, the slope of the cyclic stress response curve, and the predicted remaining fatigue life time.

[0036] Step S23 includes the following steps: Step S231: performing radial / axial offset skewness analysis based on the rotor shaft deformation offset gradient data to obtain radial / axial offset skewness data; Step S232: performing torsional deformation gradient analysis on the radial / axial offset skewness data to obtain torsional deformation gradient data; Step S233: calculating the rotor shaft deformation curvature radius partition variance based on the torsion angle deformation gradient data to obtain the deformation curvature radius partition variance; Step S234: Deducing the braking torque loss index based on the variance of the deformation curvature radius partitions to obtain the braking torque loss index; Step S235: performing an elevator braking force incremental loss analysis based on the braking torque loss index to generate braking force incremental loss data.

[0037] In an embodiment of the present invention, the offset components of each node in three-dimensional space are extracted based on the rotor shaft deformation offset gradient data. The nodes are classified according to the composite displacement directions in the axial direction (i.e., the Z-axis direction) and the radial direction (i.e., the XY plane). The sampling section interval is set to 10 mm, and 48 cross-sectional analysis planes are set on a rotor shaft with a total shaft length of 480 mm. Multiple points at the same radial distance are extracted from each cross-section, and the offset values ​​of the Z-axis component and the XY-axis component of the displacement vector are statistically calculated. A kernel density estimation method is used to fit a distribution curve for each offset value. A Gaussian kernel is used as the kernel function, and the bandwidth parameter is set to 1.5 mm. Based on the skewness coefficient of the fitting curve, segments with a skewness coefficient greater than 0.5 or less than -0.5 are marked as significantly skewed regions. The maximum value, mean, skewness direction, and peak position of the skewness density function in each skewed segment are recorded. The output is radial offset skewness data and axial offset skewness data. This data structure contains the displacement value distribution characteristics, extreme value location, and standard deviation information for each cross-sectional point. The radial and axial offset skewness data obtained in step S231 are respectively mapped to the corresponding section nodes in the rotor shaft space model, and a three-point method structural unit is constructed for the section node. The spatial projection vector of the torsion angle is formed between each group of three points. The rotation angle of the plane formed by the three points within the unit length is calculated through the spatial geometric relationship. The clockwise or counterclockwise torsion direction formed between the three points in the right-hand coordinate system is used as the basis for determining the direction of the deformation angle. The spacing between each group of three points is fixed at 10 mm. Normalization is performed according to the coordinate difference in the Z-axis direction to obtain the actual torsion angle change rate within the unit length. All torsion angle change rates are sorted into a continuous sequence to construct a one-dimensional spatial gradient map. The gradient map is subjected to first-order difference processing to analyze its change trend, and each local maximum and minimum position is extracted as the torsion angle concentration area. The output torsion angle deformation gradient data includes the angle change rate within each unit length, the gradient symbol sequence, the torsion concentration area interval length and the torsion direction.

[0038] Based on the torsion angle deformation gradient data, the curvature radius of the torsion angle change rate in each unit length segment is estimated. The geometric tangent approximation method is used to construct a vector angle between every three consecutive points. The curvature radius corresponding to the bending segment is approximately inferred by the central angle formula to obtain the discrete point data of the rotor shaft curvature change with length. The entire shaft is partitioned according to the similarity of the curvature change gradient. The minimum length of each partition is set to 50mm. The variance of the curvature radius data in each segment is calculated to obtain the difference in curvature stability between different areas. Each segment is numbered and recorded and output together with its position index, curvature average value, and standard deviation to form the deformation curvature radius partition variance data. This data structure contains the segment number, segment start and end position, average curvature radius, curvature radius fluctuation amplitude and normalized variance value, which is used as a reference for the subsequent establishment of the braking torque loss function input parameters. By inputting the deformation curvature radius partition variance data into the powertrain torque transmission efficiency model, and based on the influence of shaft deformation in the transmission system on the braking torque reaction force, the influence function is set as the exponential mapping function of the curvature radius variance to the torque loss rate. The comparative relationship between the deformation section and the braking torque change actually measured in the experimental data is used as the fitting basis, a set of calibration curves is constructed to determine the braking torque loss rate corresponding to different variance thresholds, and the degree of total braking torque loss is estimated by the segment weight accumulation method. The process adopts a five-segment partition integration method to divide the torque loss calculation into segments, and the output result is a braking torque loss index with an index value between 0 and 1, which reflects the degree of loss of the entire shaft body to the braking force transmission path. The result also includes the torque loss weight of each segment and its proportion in the entire shaft body. After obtaining the braking torque loss index, refer to the standard braking torque benchmark value of the elevator braking system. This benchmark value sets the elevator mass to 800kg and the safety braking torque requirement to be no less than 230Nm based on the elevator car mass and maximum load standards. Combined with the braking torque loss index and the dynamic operating speed data in the current state, the actual braking torque reduction in different operating stages is calculated, and the incremental braking force loss is obtained by the difference between the index and the rated braking torque value. The results are sorted in time series to construct a braking force incremental loss data sequence. The output data includes fields such as time point index, corresponding instantaneous speed, theoretical braking torque, actual deduced braking torque after loss, braking force incremental loss value, and braking state identification code, forming the final braking force incremental loss data of this step, providing a quantitative basis for subsequent elevator fault trend deduction.

[0039] Step S234 includes the following steps: The braking rotation centrifugal force imbalance fluctuation index is calculated according to the variance of the deformation curvature radius partition, and the braking rotation centrifugal force imbalance fluctuation index is obtained; The braking rotation axial force change rate is calculated proportionally based on the variance of the deformation curvature radius partition, and the proportional data of the axial force change rate is obtained; Perform logarithmic linearization on the proportional data of the axial force variation rate to obtain the logarithmic linear data of the force variation; The braking torque loss index is deduced based on the braking rotation centrifugal force imbalance fluctuation index and the logarithmic linear data of the drift force change to obtain the braking torque loss index.

[0040] In the embodiment of the present invention, the specific operation of calculating the braking rotation centrifugal force imbalance fluctuation index based on the deformation curvature radius partition variance is as follows: first, the average curvature radius and variance value of each segment in the deformation curvature radius partition variance data are extracted as initial variables, each segment is marked as an independent structural unit, and the unit number is set from 1 to N, with N being 10. The initial rotational angular velocity is set to 30 rad based on the instantaneous speed data of each structural unit of the rotor shaft in the rotating state, and the mass parameter is estimated every second based on the equivalent mass distribution of the segment. The equivalent mass of each segment is set to 1.2 kg, and the theoretical centrifugal force value generated by each structural unit is derived according to the centrifugal force formula, and then the value is calculated. The value is multiplied by the deformation curvature radius fluctuation amplitude coefficient of the current segment, which is equal to the ratio of the standard deviation to the mean. The centrifugal force imbalance values ​​of all the above structural units are weighted averaged, and the centrifugal force imbalance fluctuation index of the entire rotor structure in the rotating state is calculated through normalization. The unit of this index is Newton multiplied by meter, and the centrifugal force imbalance fluctuation index value is expressed in a dimensionless way between 0 and 1, representing the degree of influence of different deformation variables on the overall centrifugal load distribution. The larger the partition, the higher the weight in the calculation. Finally, the braking rotation centrifugal force imbalance fluctuation index is obtained. The data structure contains the time series of the local imbalance ratio of the centrifugal force value corresponding to the unit number and the total index value.

[0041] The specific operation of proportional calculation of the braking rotation axial force change rate according to the deformation curvature radius partition variance is as follows: first, the curvature difference between each segment is compared based on the average curvature radius change trend data of each segment extracted in the previous stage, and the curvature radius change between two adjacent segments is calculated according to the axial arrangement order of the rotor shaft. The structural damping constant and axial elastic modulus of each segment are set to 120N·s / m and 210GPa respectively. Combined with the above parameters, the axial reaction force change trend caused by the deformation of the shaft segment stiffness is divided into adjacent segments. The deformation difference between the two sections is multiplied by the stiffness change factor to obtain the local drift force change value. Then, the geometric series approximation method is used to assume that the axial drift force change rate shows a geometric growth trend along the axial direction. The change value of each section is recursively deduced according to the natural exponential function and the common ratio value is determined by fitting the minimum residual criterion. The result includes the axial drift force change value of each structural section and its geometric calculation ratio relative to the previous section, which finally forms the geometric data of the axial drift force change rate. The data structure includes the local stiffness change amount of the shaft section number, the local curvature variation value, the drift force change value and the geometric calculation common ratio sequence.

[0042] The operation of performing logarithmic linearization processing on the geometric data of axial drift force change rate is to use the geometric data formed in step 2 as input to perform natural logarithm conversion on the drift force change rate of all sections according to their numerical values, that is, to take the ln function for each geometrically calculated drift force change value to improve the linear fitting accuracy of subsequent analysis. Then, the least squares linear regression method is used to construct a linear regression function between the logarithmic value sequence of each section and its number in the axial direction, and the linear coefficient and bias term are extracted and expressed as the drift force change output by the logarithmic linear regression model. The logarithmic linear data contains the number of each structural section, its logarithmically transformed drift force value, regression slope intercept, fitting residual sum of squares, and linear fitting confidence interval. At the same time, it is recorded whether the overall linear trend is statistically significant. The significance judgment standard is a confidence level greater than 95%. The braking torque loss index is deduced based on the braking rotation centrifugal force imbalance fluctuation index and the logarithmic linear data of the drift force variation. The centrifugal force imbalance fluctuation index is used as the centrifugal interference factor input, and the linear slope of the logarithmic linear data of the drift force variation is used as the axial instability factor input to construct a one- and two-dimensional loss matrix. The output corresponding to each set of inputs in the matrix is ​​the braking torque loss rate. This matrix is ​​established through regression analysis of data collected at various stages of the actual elevator braking process, based on the response degree of the braking failure time to the input variables under different fault conditions. The boundary conditions are set as a maximum rotor speed of 1200 rpm and a minimum torque maintenance value of 180 Nm. A two-dimensional interpolation table is constructed based on historical data. A bilinear interpolation algorithm is used with the centrifugal imbalance value and the logarithmic slope as the coordinate axes to determine the torque loss degree under the current operating condition. Finally, the braking torque loss index under the current state is calculated. The index value is expressed in a normalized form from 0 to 1, representing the effective loss rate of braking output capacity per unit time. The final data structure contains the current centrifugal imbalance value, the logarithmic slope value, the loss matrix mapping position interpolation function residual value, and the deduced braking torque loss index value.

[0043] Step S24 includes the following steps: Step S241: analyzing the loss increment time series relationship of the braking force increment loss data to obtain the loss increment time series relationship; Step S242: performing trend increment heteroscedasticity analysis on the loss increment time series relationship to generate loss trend increment heteroscedasticity features; Step S243: performing a change-point lognormal distribution analysis on the loss increment time series relationship based on the loss trend increment heteroscedasticity feature to obtain change-point lognormal distribution data; Step S244: performing time series incremental regression analysis on the change point log-normal distribution data to obtain time series incremental data of braking force loss.

[0044] In an embodiment of the present invention, a timestamp and a corresponding loss value sequence are extracted from the incremental braking force loss data. The time unit is set to 0.5 seconds, and the time series sample length within the extraction period is 300 sampling points, i.e., a total duration of 150 seconds. After rearranging the sequence along the time axis to ensure data continuity, a difference sequence is constructed, i.e., the difference between the incremental braking force loss values ​​at two adjacent moments. Subsequently, a sliding window method is used to perform local fluctuation intensity statistics on the difference sequence, with a window size of 30 sampling points and a step size of 5 sampling points. The average absolute value of the difference within each window is calculated as a local fluctuation intensity indicator, and the time position corresponding to the window center point and the fluctuation value are extracted to form a fluctuation time series. An inflection point identification operation is performed based on the above fluctuation time series, and the inflection point determination condition is that the mean change rate of the fluctuation intensity of the two groups of windows is greater than 30%. Through the above method, the time positions where the fluctuation of the incremental braking force loss value changes significantly are identified and marked as key points in the loss increment time series relationship. Simultaneously, loss increment time series relationship data containing fields such as the original difference sequence, window fluctuation intensity, and inflection point time identifier are generated for subsequent analysis and processing. Based on the loss increment time series relationship data obtained in step S241, first extract the continuous subsequences between the inflection points as independent time interval segments, set the minimum segment length for analysis to 20 sampling points, and perform heteroscedasticity test on the incremental difference sequence in each segment. The specific method is to use the cumulative variance distribution evaluation method, that is, calculate the forward accumulated variance value curve of each sampling point in the segment and perform linear fitting, and then use the average deviation value between the actual variance curve and the fitting curve as the heteroscedasticity coefficient. If the coefficient is greater than the set threshold value of 0.1, the segment is marked as heteroscedastic. All segments with heteroscedastic characteristics are statistically extracted to extract their start time, end time, segment variance change direction and maximum variance change amplitude. On this basis, the loss trend incremental heteroscedasticity feature data is constructed. The data includes the time boundary, fluctuation intensity range, cumulative deviation amplitude and heteroscedasticity coefficient of each heteroscedastic segment, and is saved for further distribution change point detection operations.Based on the incremental heteroscedasticity feature data of the loss trend extracted in step S242, the incremental differential data within each heteroscedastic segment is first standardized, that is, the mean within the segment is subtracted and then divided by the standard deviation to eliminate the interference of the data scale on the fitting distribution. The standardized data are then arranged in sequence as a univariate time series. The sliding window change point identification strategy is adopted to set the window length to 40 sampling points and the step length to 5 sampling points. The data is fitted with a log-normal distribution in each window. The shape parameter σ and location parameter μ are calculated using the maximum likelihood estimation method, and the fitting residual and confidence level of each window are recorded. In multiple consecutive windows, the fitting parameter mutation points, that is, the points where the change rate of the shape parameter or location parameter is greater than 10%, are identified as candidate change point points. The significance of the squared residual difference between the windows on both sides is further determined to screen out the fitting mutation points with a small fluctuation range. The points finally retained are the change points. The log-normal distribution data includes the time index of each change point, the change amplitude of the fitting parameter before and after the mutation, the residual difference statistics, the confidence level, and the standardized value of the data corresponding to the center point of the fitting window, which are used for subsequent incremental trend modeling. Based on the logarithmic normal distribution data of the change points obtained in step S243, the incremental difference data before and after each change point are extracted to form an incremental subsequence before and after the change point and normalized according to the time series, and an incremental difference time curve is constructed. Then, linear regression fitting is performed on each subsequence to extract its slope-intercept determination coefficient and fitting residual, and the fitting significance level is set to 95%. Subsequences that fail the significance test are directly excluded. For subsequences that pass the fitting, the change rate is further evaluated based on their fitting slope values. The difference in the incremental change rate before and after each change point is calculated as the local change amplitude and used as the time series incremental value of the unit event. The time series incremental value is output and combined with the change point position to form a time series incremental sequence. The braking force loss time series incremental data finally constructed contains the fitting curve parameters before and after each change point position, the local change rate of the fitting residual value, the fitting effective interval and the time index coordinates, which are used as the basis for the input of the time series fault factor in the subsequent elevator brake fault trend prediction model.

[0045] Step S3 includes the following steps: Step S31: normalizing the braking force loss time series increment data to obtain normalized braking force loss time series increment data; Step S32: performing incremental gradient nonlinear induction on the normalized data of the braking force loss time series increment to obtain loss incremental gradient nonlinear induction data; Step S33: constructing an elevator fault prediction model based on the loss increment gradient nonlinear induction data based on the K nearest neighbor algorithm to obtain an elevator fault prediction model; Step S34: Send the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

[0046] As an example of the present invention, refer to Figure 3As shown, in this example, step S3 includes: Step S31: normalizing the braking force loss time series increment data to obtain normalized braking force loss time series increment data; In an embodiment of the present invention, continuous time series windows are first extracted from the existing incremental braking force loss time series data. The length of each window is set to 50 data points, with a step size of 10 data points. Adjacent windows are ensured to overlap to enhance the continuity of the time series and the accuracy of identifying changing trends. A range normalization operation is performed on the data in each window. That is, the maximum and minimum values ​​in the window are first calculated, and then all data in the window are converted to normalized values ​​between 0 and 1 by subtracting the minimum value and dividing by the range. The normalization operation is performed with floating-point precision to ensure that six significant digits are retained. The consistency of the normalized overlapped portions of the data across windows is maintained. The normalization offset problem of the cross-portion is resolved by averaging the normalized values ​​at the same time point. All normalized results are rearranged in their original time order to form the complete normalized data of the incremental braking force loss time series. At the same time, the window number of the window where the original normalized value of the timestamp corresponding to each data point and its maximum and minimum reference values ​​are recorded in the normalized data structure for subsequent nonlinear induction and prediction processing.

[0047] Step S32: performing incremental gradient nonlinear induction on the normalized data of the braking force loss time series increment to obtain loss incremental gradient nonlinear induction data; In the embodiment of the present invention, the normalized braking force loss time series incremental normalized data is subjected to nonlinear induction of incremental gradients. First, the normalized value difference between two adjacent data points is calculated according to the time series sequence to construct a primary incremental gradient sequence. This sequence reflects the local change rate of the normalized data. Then, a second-order gradient analysis interval with a sliding window length of 30 data points is set. A nonlinear morphological analysis operation is performed on the incremental gradient sequence within each sliding window. A cubic polynomial function is fitted to the gradient sequence within the window using a polynomial fitting method. The derivative coefficients and inflection point positions of each order of the function are extracted, and the local extreme points are identified and their positive and negative change trends are determined. The effectiveness of the nonlinear fitting is further judged by fitting residuals, and only the window fitting results with the residual sum of squares less than 0.002 are retained as valid samples. Then, the cubic polynomial coefficients extracted from each valid window and the number of inflection points of their derivative curves and the direction of extreme value change are combined to form a nonlinear feature vector. The central timestamp of each window is used as a reference to mark the positions of all feature vectors, and finally the loss incremental gradient nonlinear summary data is constructed. This data structure contains six-dimensional feature items of each feature vector, specifically the first-order coefficient, second-order coefficient, third-order coefficient, residual, number of extreme values ​​and number of slope inflection points, which are used to judge the trend of elevator fault status.

[0048] Step S33: constructing an elevator fault prediction model based on the loss increment gradient nonlinear induction data based on the K nearest neighbor algorithm to obtain an elevator fault prediction model; In the embodiment of the present invention, an elevator fault prediction model is constructed based on the loss incremental gradient nonlinear inductive data based on the K nearest neighbor algorithm. First, nonlinear inductive feature data corresponding to the time period containing the elevator brake fault record is extracted from the actual historical working condition data to form a labeled sample library. The labels are divided into four categories: normal, mild abnormality, moderate abnormality and severe fault, which are represented by 0, 1, 2 and 3 respectively. All sample data are divided into a training set and a test set with a ratio of 8 to 2. In the training stage, a Euclidean distance measurement matrix is ​​constructed to calculate the Euclidean distance between the nonlinear feature vector of each test sample and all sample vectors in the training set, and the K nearest distance is extracted. The K value is set to 7 for each sample, and the label values ​​of these 7 samples are classified by majority voting to determine the prediction category to which the test sample belongs. This process is repeated until all test samples are predicted, and evaluation indicators such as accuracy, recall rate, and F1 score are recorded to confirm the model prediction performance. The final training results are saved as the elevator fault prediction model under the condition that the accuracy rate is greater than 93%. At the same time, all support vectors identified during the training process are saved for model visualization and decision path analysis. The prediction model is encapsulated in the form of a JSON structure, which contains parameters such as K value, distance matrix index structure, label distribution characteristics, and training sample number for online fault detection.

[0049] Step S34: Send the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

[0050] In an embodiment of the present invention, a constructed elevator fault prediction model is sent to a terminal to execute an elevator fault prediction method. The prediction model is written into an address segment of an intelligent prediction processing module set in an elevator operation control mainboard via the industrial control bus Modbus TCP protocol using a serial communication mechanism. Control instructions are transmitted in a fixed format, including a model structure declaration byte, parameter length, check code, and model data body. The sending module first performs byte sequence block compression on the model. Each block is 256 bytes long, and the total model file size is approximately 96KB. The model is sent frame by frame according to the data packet sequence numbering from 1 to N. Each frame is accompanied by a data check bit, and the communication integrity is confirmed by a CRC16 check code. After receiving the data, the target terminal immediately decompresses and reconstructs the data to verify the integrity and writes the data into a cache unit for subsequent real-time elevator operation data comparison and identification operations. At the same time, a model version identification and terminal hardware compatibility verification mechanism are established. The model operation control logic is initialized through the terminal logic module. This process is completed within 15 seconds after the operating system is started and is executed regularly by a system background task. The prediction model interface is periodically called during the elevator braking cycle to perform online reasoning on the real-time collected nonlinear gradient feature data and generate real-time fault prediction results for upload and processing by the maintenance system.

[0051] Step S32 includes the following steps: Step S321: Drawing a loss time series increment curve for the normalized braking force loss time series increment data to generate a loss time series increment curve; Step S322: performing a logarithmic difference calculation of the amplitude change ratio of adjacent segments on the loss time series incremental curve to obtain the logarithmic difference of the amplitude change ratio; Step S323: performing time series reverse slip accumulation on the loss time series increment curve according to the logarithmic difference of the amplitude change ratio to obtain loss time series reverse slip accumulation data; Step S324: performing incremental gradient nonlinear induction according to the loss time series reverse slip accumulation data to obtain loss incremental gradient nonlinear induction data.

[0052] In an embodiment of the present invention, after obtaining the normalized data of the braking force loss time series increment, the horizontal axis is first set to the time index unit in seconds, the vertical axis is set to the normalized incremental value range in the range of 0 to 1, and the plot drawing function provided in the Matplotlib graphics library is used to construct continuous sample points on the time axis with a sample interval of 0.5 seconds. A total of 600 samples are collected, or a total time of 300 seconds. In the figure, each normalized incremental data point is marked with a blue solid circle in chronological order. The adjacent data points are smoothly connected by the cubic spline interpolation method to generate a continuous curve to enhance the visualization clarity of the local change trend. No fitting and data smoothing are performed in the drawing process. The gradient jump trend of the original data in the time series is retained, providing basic structural support for subsequent difference calculation and slip processing. The curve data generated by the drawing is stored in an array structure and is accompanied by the original normalized value, index position and time label of each sample point. It is also output as an image file for abnormal feature change trend comparison and manual verification. After the curve is drawn, the entire loss time series incremental curve is divided into continuous segments with an adjacent length of 30 points. The total number of segments is 570. In each segment, the maximum amplitude value of the current segment and the maximum amplitude value of the previous segment are extracted to form a segmented amplitude pair. The amplitude pair is then subjected to logarithmic difference calculation. The natural logarithm conversion function is set to take the natural logarithm of the two values ​​in each set of amplitude pairs and calculate their difference. The result is the logarithmic difference of the amplitude change ratio of the segment. In order to avoid infinitesimal values ​​in the logarithmic calculation, a very small amount is added to each amplitude value. It is set to 0.0001 to ensure that calculation can still be performed when the amplitude is zero. The logarithmic differences of all segments are arranged in sequence to form a logarithmic difference sequence with the starting time point index and current segment number of each segment. At the same time, the mean and variance of the logarithmic difference sequence are calculated for the subsequent determination of the initial direction of the reverse slip and the setting of the amplitude recognition interval. Based on the logarithmic difference of the amplitude change ratio obtained in step S322, a time series reverse slip accumulation process is performed on the loss time series increment curve. The slip processing direction is set to reversely trace back from the current segment to the historical segment to construct an asymmetric time reverse window. Each window contains the current segment and its previous five segments, totaling 180 points. Within each reverse slip window, the logarithmic difference is weighted and accumulated. The weighting factor is set as a linear function that increases with time, from 0.2 to 1, increasing by 0.16 each time. The accumulated value is equal to the logarithmic difference of each segment in the window multiplied by its corresponding weighting factor, and then the sum is normalized to the range of 0 to 1. This method strengthens the impact of recent changes on the overall trend judgment. The weighted accumulated values ​​output by all slip windows constitute a reverse slip accumulation data array. Each data contains the window start time, weighted accumulation result, corresponding segment number, slip amplitude weight vector, and the accumulated logarithmic difference residual index of the processing window. After the reverse slip data is formed, it is used to identify the asymmetric distribution of sudden and progressive characteristics in the local time series. Based on the reverse sliding cumulative data output in step S323, incremental gradient extraction and nonlinear induction processing are performed on the weighted cumulative results of each reverse window. First, a time series of the cumulative value of each reverse window is constructed. The difference between every two adjacent time points constitutes a first-order incremental gradient sequence. The minimum discernible gradient is set to 0.01. All gradient values ​​with absolute values ​​less than the threshold are eliminated to ensure that only gradient signals with significant change trends are retained. Then, a sliding polynomial fitting method is used to set the window length to 50 points and the overlap length to 25 points. A cubic polynomial fitting is performed on the first-order gradient data in each window. The fitting results are: The first-order derivative, second-order derivative and residual sum of squares are retained, and the inflection point position and its curvature change trend are recorded. The residual evaluation is performed on the fitting results in each window, and only the segments with fitting residuals less than 0.01 are retained as valid segments. The polynomial derivative coefficients, curvature extreme points and corresponding time indexes extracted from all valid fitting segments together constitute the final loss incremental gradient nonlinear summary data. This data structure contains complete parameter dimensions such as timestamp, nonlinear fitting parameters, first-order and second-order derivatives, extreme point numbers, residual square values, etc., which are used as the data basis input for the K nearest neighbor algorithm to construct a fault prediction model.

[0053] The present invention also provides an elevator fault prediction system for executing the elevator fault prediction method described above, the elevator fault prediction system comprising: The vibration frequency disorder structure analysis module is used to deploy a vibration sensor on the elevator brake to collect the running vibration signal, and then perform frequency domain conversion to obtain the running vibration cleaning frequency domain data; the vibration frequency disorder structure analysis is performed on the running vibration cleaning frequency domain data to obtain the vibration frequency disorder structure data; The braking force incremental loss analysis module is used to analyze the abnormal vibration frequency intensity based on the vibration frequency disordered structure data to obtain abnormal vibration frequency intensity data; based on the abnormal vibration frequency intensity data, the elevator braking force incremental loss analysis is performed to obtain the braking force loss time series incremental data; The fault prediction model construction module is used to perform incremental gradient nonlinear induction on the braking force loss time series incremental data to obtain loss incremental gradient nonlinear induction data; construct an elevator fault prediction model based on the loss incremental gradient nonlinear induction data based on the K nearest neighbor algorithm to obtain an elevator fault prediction model; and send the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

[0054] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting elevator failures, characterized in that: The following steps are involved: Step S1: Deploy a vibration sensor on the elevator brake to collect the running vibration signal, and then perform frequency domain conversion to obtain running vibration cleaning frequency domain data; perform vibration frequency disorder structure analysis on the running vibration cleaning frequency domain data to obtain vibration frequency disorder structure data; Step S2: performing abnormal vibration frequency intensity analysis based on the vibration frequency disordered structure data to obtain abnormal vibration frequency intensity data; performing elevator braking force incremental loss analysis based on the abnormal vibration frequency intensity data to obtain braking force loss time series incremental data; Step S3: performing incremental gradient nonlinear induction on the braking force loss time series incremental data to obtain loss incremental gradient nonlinear induction data; constructing an elevator fault prediction model on the loss incremental gradient nonlinear induction data based on the K-nearest neighbor algorithm to obtain an elevator fault prediction model; and sending the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

2. The elevator failure prediction method according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: deploying a vibration sensor on the elevator brake to collect an operating vibration signal to obtain an elevator brake operating vibration signal; Step S12: performing signal cleaning processing on the elevator brake operation vibration signal to obtain a brake operation vibration cleaning signal; Step S13: performing frequency domain conversion on the brake operation vibration cleaning signal to obtain operation vibration cleaning frequency domain data; Step S14: performing vibration frequency disorder structure analysis on the frequency domain data of the vibration cleaning operation to obtain vibration frequency disorder structure data.

3. The elevator failure prediction method according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: performing abnormal vibration frequency intensity analysis on the frequency domain data of the vibration cleaning operation according to the vibration frequency disorder structure data to obtain abnormal vibration frequency intensity data; Step S22: performing rotor shaft deformation offset gradient calculation based on the abnormal vibration frequency intensity data to obtain rotor shaft deformation offset gradient data; Step S23: performing an elevator braking force incremental loss analysis based on the rotor shaft deformation offset gradient data to generate braking force incremental loss data; Step S24: performing time series incremental regression analysis on the braking force incremental loss data to obtain braking force loss time series incremental data.

4. The elevator failure prediction method according to claim 3, characterized in that: Step S22 includes the following steps: Step S221: identifying the rotor shaft vibration energy conduction vector based on the abnormal vibration frequency intensity data to obtain the rotor shaft vibration energy conduction vector; Step S222: performing radial vibration energy geometric fluctuation analysis on the rotor shaft vibration energy conduction vector to obtain radial vibration energy geometric fluctuation data; Step S223: performing rotor shaft cross-section ellipse energy deviation calculation based on radial vibration energy geometric fluctuation data to obtain cross-section ellipse energy deviation data; Step S224: performing rotor shaft cyclic torsional stress coupling based on the cross-sectional ellipse energy deviation data and the radial vibration energy geometric fluctuation data to obtain cyclic torsional stress coupling data; Step S225: obtaining basic material fatigue limit parameters of the rotor shaft; performing rotor shaft deformation offset gradient calculation on the basic material fatigue limit parameters according to the cyclic torsional stress coupling data to obtain rotor shaft deformation offset gradient data.

5. The elevator failure prediction method according to claim 4, characterized in that: Step S23 includes the following steps: Step S231: performing radial / axial offset skewness analysis based on the rotor shaft deformation offset gradient data to obtain radial / axial offset skewness data; Step S232: performing torsional deformation gradient analysis on the radial / axial offset skewness data to obtain torsional deformation gradient data; Step S233: calculating the rotor shaft deformation curvature radius partition variance based on the torsion angle deformation gradient data to obtain the deformation curvature radius partition variance; Step S234: Deducing the braking torque loss index based on the variance of the deformation curvature radius partitions to obtain the braking torque loss index; Step S235: performing an elevator braking force incremental loss analysis based on the braking torque loss index to generate braking force incremental loss data.

6. The elevator failure prediction method according to claim 5, characterized in that: Step S234 includes the following steps: The braking rotation centrifugal force imbalance fluctuation index is calculated according to the variance of the deformation curvature radius partition, and the braking rotation centrifugal force imbalance fluctuation index is obtained; The braking rotation axial force change rate is calculated proportionally based on the variance of the deformation curvature radius partition, and the proportional data of the axial force change rate is obtained; Perform logarithmic linearization on the proportional data of the axial force variation rate to obtain the logarithmic linear data of the force variation; The braking torque loss index is deduced based on the braking rotation centrifugal force imbalance fluctuation index and the logarithmic linear data of the drift force change to obtain the braking torque loss index.

7. The elevator failure prediction method according to claim 6, characterized in that: Step S24 includes the following steps: Step S241: analyzing the loss increment time series relationship of the braking force increment loss data to obtain the loss increment time series relationship; Step S242: performing trend increment heteroscedasticity analysis on the loss increment time series relationship to generate loss trend increment heteroscedasticity features; Step S243: performing a change-point lognormal distribution analysis on the loss increment time series relationship based on the loss trend increment heteroscedasticity feature to obtain change-point lognormal distribution data; Step S244: performing time series incremental regression analysis on the change point log-normal distribution data to obtain time series incremental data of braking force loss.

8. The elevator failure prediction method according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: normalizing the braking force loss time series increment data to obtain normalized braking force loss time series increment data; Step S32: performing incremental gradient nonlinear induction on the normalized data of the braking force loss time series increment to obtain loss incremental gradient nonlinear induction data; Step S33: constructing an elevator fault prediction model based on the loss increment gradient nonlinear induction data based on the K nearest neighbor algorithm to obtain an elevator fault prediction model; Step S34: Send the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

9. The elevator failure prediction method according to claim 8, characterized in that: Step S32 includes the following steps: Step S321: Drawing a loss time series increment curve for the normalized braking force loss time series increment data to generate a loss time series increment curve; Step S322: performing a logarithmic difference calculation of the amplitude change ratio of adjacent segments on the loss time series incremental curve to obtain the logarithmic difference of the amplitude change ratio; Step S323: performing time series reverse slip accumulation on the loss time series increment curve according to the logarithmic difference of the amplitude change ratio to obtain loss time series reverse slip accumulation data; Step S324: performing incremental gradient nonlinear induction according to the loss time series reverse slip accumulation data to obtain loss incremental gradient nonlinear induction data.

10. An elevator fault prediction system, characterized in that: For executing the elevator fault prediction method according to claim 1, the elevator fault prediction system comprises: The vibration frequency disorder structure analysis module is used to deploy a vibration sensor on the elevator brake to collect the running vibration signal, and then perform frequency domain conversion to obtain the running vibration cleaning frequency domain data; the vibration frequency disorder structure analysis is performed on the running vibration cleaning frequency domain data to obtain the vibration frequency disorder structure data; The braking force incremental loss analysis module is used to analyze the abnormal vibration frequency intensity based on the vibration frequency disordered structure data to obtain abnormal vibration frequency intensity data; based on the abnormal vibration frequency intensity data, the elevator braking force incremental loss analysis is performed to obtain the braking force loss time series incremental data; The fault prediction model construction module is used to perform incremental gradient nonlinear induction on the braking force loss time series incremental data to obtain loss incremental gradient nonlinear induction data; construct an elevator fault prediction model based on the loss incremental gradient nonlinear induction data based on the K nearest neighbor algorithm to obtain an elevator fault prediction model; and send the elevator fault prediction model to the terminal to execute the elevator fault prediction method.

Images