Method and system for monitoring stress and strain of connector in real time

Abstract

The invention discloses a connector stress-strain real-time monitoring method and system, and relates to the technical field of structure health monitoring, and the method comprises the steps: decomposing real-time stress-strain data and pre-processed external environment data into a high-frequency component and a low-frequency component through wavelet multi-scale decomposition, and carrying out the dynamic weighting fusion into corrected stress-strain data; continuously monitoring the health indexes of the connector based on the corrected stress-strain data, analyzing the risk trend of the connector, and generating real-time monitoring feedback when a health index sequence is monitored to exceed a dynamic safety threshold value; according to the real-time monitoring feedback and the real-time stress-strain data, a health state knowledge graph is constructed, the fault type and the occurrence probability of the connector are predicted, and a fault prediction report is generated; and obtaining a personalized maintenance plan of the connector based on the fault prediction report in combination with historical fault data. According to the invention, accurate extraction and noise suppression of multi-scale signal features are realized, and the reliability and anti-interference capability of monitoring data are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structural health monitoring, in particular to a connector stress-strain real-time monitoring method and system. BACKGROUND

[0002] In the field of power connector reliability monitoring, real-time detection of stress-strain parameters is a key technical means for evaluating the health state of mechanical structures. Traditional monitoring methods mainly use resistance strain gauge sensors to directly measure structural deformation, and convert mechanical strain into electrical signals through a Wheatstone bridge for collection and analysis. This method is based on the linear assumption of Hooke's law, and calculates stress distribution through the calibrated strain-voltage relationship. In industrial field applications, a standardized detection process has been formed. At the same time, optical fiber grating sensing technology has been introduced into this field in recent years, which uses the linear relationship between wavelength shift and strain value to achieve distributed measurement, providing a new technical path for structural health monitoring.

[0003] When dealing with complex working conditions, the traditional method has difficulty in effectively distinguishing between real stress changes and measurement drift caused by environmental temperature and humidity fluctuations with a single sensor system, resulting in a decrease in monitoring accuracy under dynamic load conditions. In addition, the alarm mechanism with fixed thresholds lacks predictive judgment of equipment degradation trends, making it difficult to identify gradual failure characteristics in a timely manner. In particular, in industrial scenes with significant vibration interference, the existing technology has limited ability to capture high-frequency dynamic responses, and the spatial and temporal registration accuracy between different physical quantity measurements needs to be improved, which to some extent affects the integrity of the state evaluation. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a connector stress-strain real-time monitoring method to solve the problems of insufficient fusion of multi-source heterogeneous data and weak dynamic adaptability in monitoring technology.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present application provides a connector stress-strain real-time monitoring method, which comprises, collecting and preprocessing multi-physical quantity data and external environment data; extracting current and voltage signal sequences from the preprocessed multi-physical quantity data, and calculating real-time stress-strain data through an electrical model; based on the real-time stress-strain data and the preprocessed external environment data, decomposing into high-frequency components and low-frequency components through wavelet multi-scale decomposition, and dynamically weighting and fusing into corrected stress-strain data; Based on the corrected stress-strain data, the health indicators of the connector are continuously monitored, and the risk trend of the connector is analyzed, and when the health indicator sequence exceeds the dynamic safety threshold, real-time monitoring feedback is generated; According to the real-time monitoring feedback and the real-time stress-strain data, a health state knowledge graph is constructed, and the failure type and occurrence probability of the connector are predicted, and a failure prediction report is generated; Based on the failure prediction report, combined with historical failure data, a personalized maintenance plan for the connector is obtained.

[0007] As a preferred scheme of the connector stress-strain real-time monitoring method, the multi-physical quantity data includes strain signals, optical fiber signals, current signals and voltage signals; The external environment data includes temperature signals, humidity signals and vibration signals; The preprocessing includes denoising processing, standardization processing, time synchronization and data alignment.

[0008] As a preferred scheme of the connector stress-strain real-time monitoring method, the current and voltage signal sequences in the preprocessed multi-physical quantity data are extracted, and the real-time stress-strain data is calculated through an electrical model, and the steps are as follows, Separate the current signal sequence and the voltage signal sequence from the preprocessed multi-physical quantity data; Based on Ohm's law and Hooke's law, the mathematical relationship between current, voltage and stress-strain is established, and an electrical model is generated; The current signal sequence and the voltage signal sequence are input into the electrical model, and the initial stress-strain data is output; Compare the preprocessed optical fiber signal and the initial stress-strain data to obtain an offset value; The initial stress-strain data is corrected through offset compensation to generate real-time stress-strain data.

[0009] As a preferred scheme of the connector stress-strain real-time monitoring method, the real-time stress-strain data and the preprocessed external environment data are decomposed into high-frequency components and low-frequency components through wavelet multi-scale decomposition, and are dynamically weighted and fused into corrected stress-strain data, and the steps are as follows, The real-time stress-strain data and the preprocessed external environment data are decomposed into high-frequency components and low-frequency components through wavelet multi-scale decomposition, and are dynamically weighted and fused into corrected stress-strain data, and the steps are as follows, According to the high-frequency component set and the low-frequency component set, the environmental interference is identified, and the environmental interference characteristics are obtained; Based on the environmental interference characteristics, the dynamic weights of each component are calculated through an exponential function, and different weight distributions are performed according to different frequency components; The weighted high-frequency components and low-frequency components are fused into corrected stress-strain data.

[0010] As a preferred scheme of the connector stress-strain real-time monitoring method, wherein: based on the modified stress-strain data, the health index of the connector is continuously monitored, and the risk trend of the connector is analyzed, the steps are as follows, The modified stress-strain data is weighted and normalized to generate a health index sequence; Collect the historical health index sequence, and train the ARIMA model through the autocorrelation function; The health index sequence is input into the trained ARIMA model to predict the future short-term health index sequence and generate a risk trend slope.

[0011] As a preferred scheme of the connector stress-strain real-time monitoring method, wherein: when the health index sequence is monitored to exceed the dynamic safety threshold, real-time monitoring feedback is generated, the steps are as follows, The health index sequence is calculated by rolling statistics, and the risk trend slope is dynamically adjusted to generate a dynamic safety threshold; When the health index sequence exceeds the dynamic safety threshold, record the abnormal index value, the abnormal timestamp, the correlation parameter and the abnormal level, and generate real-time monitoring feedback.

[0012] As a preferred scheme of the connector stress-strain real-time monitoring method, wherein: according to the real-time monitoring feedback and the real-time stress-strain data, a health state knowledge graph is constructed, the steps are as follows, Align the timestamps of the real-time monitoring feedback and the real-time stress-strain data, and extract abnormal features and time sequence features respectively, and integrate them into a multi-dimensional feature vector sequence; Based on the multi-dimensional feature vector sequence, combined with the historical health index sequence, the health state nodes of the connector and the transformation relationship between the health states in the health state knowledge graph are defined in the health state knowledge graph, forming a health state knowledge graph framework; Map the multi-dimensional feature vector sequence to the health state knowledge graph framework, group the multi-dimensional feature vector sequence by clustering algorithm and associate the state nodes, construct the health state knowledge graph, and generate an embedding vector sequence.

[0013] As a preferred scheme of the connector stress-strain real-time monitoring method, wherein: the failure type and occurrence probability of the connector are predicted, and a failure prediction report is generated, the steps are as follows, According to the historical failure data, the parameters of the multilayer perceptron are adjusted through the loss function to generate a trained multilayer perceptron; The embedding vector sequence is input into the trained multilayer perceptron to output the failure type probability distribution and the occurrence probability value; The fault type probability distribution and the occurrence probability value are formatted, and a fault prediction report is obtained.

[0014] As a preferred scheme of the connector stress-strain real-time monitoring method, wherein: based on the fault prediction report, combined with historical fault data, a personalized maintenance plan of the connector is obtained, and the steps are as follows, Maintenance features and fault features in the fault prediction report and the historical fault data are extracted respectively, and a multi-dimensional maintenance feature vector is generated; Based on the multi-dimensional maintenance feature vector, a K-nearest neighbor algorithm is used to retrieve the maintenance plan of the similar fault case in the historical fault data, and a maintenance plan candidate set of the connector is generated; The maintenance plan candidate set of the connector is optimized by a genetic algorithm, and a personalized maintenance plan of the connector is generated.

[0015] In the second aspect, the application provides a connector stress-strain real-time monitoring system, comprising, A data acquisition module is used for acquiring and preprocessing multi-physical quantity data and external environment data; A real-time calculation module is used for extracting current and voltage signal sequences from the preprocessed multi-physical quantity data, and calculating real-time stress-strain data through an electrical model; A dynamic weighting module is used for decomposing the real-time stress-strain data and the preprocessed external environment data into high-frequency components and low-frequency components through wavelet multi-scale decomposition, and dynamically weighting and fusing them into corrected stress-strain data; A risk monitoring module is used for continuously monitoring the health index of the connector based on the corrected stress-strain data, and analyzing the risk trend of the connector, and generating a real-time monitoring feedback when the health index sequence exceeds a dynamic safety threshold; A fault prediction module is used for constructing a health state knowledge graph according to the real-time monitoring feedback and the real-time stress-strain data, and predicting the fault type and occurrence probability of the connector, and generating a fault prediction report; A fault maintenance module is used for obtaining a personalized maintenance plan of the connector based on the fault prediction report and combined with historical fault data.

[0016] The application has the following beneficial effects: through the wavelet multi-scale decomposition step, the real-time stress-strain data and the preprocessed external environment data are decomposed into high-frequency components and low-frequency components, the accurate extraction of multi-scale signal features and noise suppression are realized, and the reliability and anti-interference ability of the monitoring data are effectively improved. At the same time, through the health state knowledge graph construction step, the real-time monitoring feedback and the stress-strain data are mapped into a structured knowledge network, the deep semantic analysis of the health state and the early warning of the fault are realized. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative effort.

[0018] Fig. 1 Flow chart of connector stress-strain real-time monitoring method.

[0019] Fig. 2 Schematic diagram of connector stress-strain real-time monitoring system.

[0020] Fig. 3 Flow chart of data acquisition and preprocessing.

[0021] Fig. 4 Flow chart of health monitoring and maintenance plan generation. DETAILED DESCRIPTION

[0022] In order to make the above objectives, features and advantages of the present application more apparent and understandable, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0023] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced without the specific details, other than those described herein, and it is understood that the present application is not limited to the embodiments described herein and can be practiced with or without the other embodiments.

[0024] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. The "in one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment.

[0025] REFERENCE Figs. 1-4 For one embodiment of the present application, the embodiment provides a connector stress-strain real-time monitoring method, comprising the following steps: S1, collecting multi-physical quantity data and external environment data and preprocessing; The multi-physical quantity data includes strain signals, optical fiber signals, current signals and voltage signals; Further, the strain signals and optical fiber signals of the connector are collected by the resistance strain gauge sensor and the optical fiber Bragg grating sensor (the sampling frequency is usually 1000 Hz); the current signals and voltage signals of the connector are collected by the Hall current sensor and the voltage dividing circuit sensor (the sampling frequency is usually 2000 Hz).

[0026] The external environment data includes temperature signals, humidity signals, and vibration signals. Further, the temperature signals and humidity signals of the external environment are collected by a thermocouple temperature sensor and a capacitive humidity sensor (the sampling frequency is usually 100 Hz); the vibration signals of the external environment are collected by a piezoelectric vibration sensor (the sampling frequency is usually 5000 Hz).

[0027] The preprocessing includes denoising processing, standardization processing, time synchronization, and data alignment. Further, in a fixed-size sliding window (for example, 5 sampling points), for the signals at each time point in the multi-physical quantity data and the external environment data, the average value of adjacent data points in the window is calculated, and the average value of adjacent data points is taken to replace the current point, to eliminate random high-frequency noise; the multi-physical quantity data and the external environment data are subjected to Z-score standardization processing, and the steps are as follows: the arithmetic mean and the standard deviation of each signal are calculated, each data point is subtracted by the mean and divided by the standard deviation, and the standardized multi-physical quantity data and the external environment data are generated; a unified time point sequence is generated based on the time axis with the highest sampling frequency (for example, the vibration signal 5000 Hz), and linear interpolation is used to interpolate the data points with lower sampling frequency to the unified time point sequence; it is verified whether each time point contains data of all signals, and the missing values are filled by linear interpolation; abnormal values exceeding the standard deviation range are detected, and sliding window median filtering (window size of 5 points) is used for smoothing processing.

[0028] It should be noted that the standard deviation range is set according to the confidence interval of the normal distribution, for example, when the multi-physical quantity data and the external environment data follow an approximate normal distribution, about 99.7% of the data should fall within ±3 standard deviations of the mean.

[0029] S2, extracting the current and voltage signal sequences from the preprocessed multi-physical quantity data, and calculating real-time stress and strain data through an electrical model; Separating the current signal sequence and the voltage signal sequence from the preprocessed multi-physical quantity data; Further, the current signal sequence and the voltage signal sequence are extracted according to the multi-physical quantity signal labels (such as "current" and "voltage"), and the current signal sequence and the voltage signal sequence are detected point by point with a unified time axis to determine whether there are missing values and sampling intervals exceeding the set threshold, and the missing values are filled using linear interpolation.

[0030] Based on Ohm's law and Hooke's law, a mathematical relationship between current, voltage, and stress and strain is established to generate an electrical model. Further, based on the relationship between the resistance of the conductor, voltage and current described by Ohm's law, the relationship between stress and strain described by Hooke's law, and the relationship between the resistance change rate and strain described by the piezoresistive effect, the mathematical relationship between current, voltage and stress-strain is established, and the electrical model is generated; the current signal sequence, voltage signal sequence and fiber signal strain value are aligned point by point on the unified time axis, the strain value in the fiber signal is taken as the supervised reference input, and the initial stress-strain data calculated by the electrical model is compared, an error loss function with the mean square error of the two as the target is constructed, and the strain sensitivity coefficient and material elastic modulus correction coefficient in the electrical model are iteratively updated by using the gradient descent algorithm, and the training is stopped when the error change is less than the preset convergence threshold, and the calibrated electrical model is output.

[0031] It should be noted that the convergence threshold is set according to the convergence speed of the error change in the electrical model training and the signal noise level, and the exemplary value range is usually set to 1x10 -4 to 1x10 -6 .

[0032] The current signal sequence and the voltage signal sequence are input into the electrical model, and the initial stress-strain data is output. Further, the real-time resistance value at each time point in the current signal sequence and the voltage signal sequence is calculated by the Ohm's law expression, and the strain value and the stress value are calculated by the strain value calculation expression and the stress value calculation expression, and the strain values and stress values of all time points are integrated to generate the initial stress-strain data.

[0033] The strain value calculation expression is: ; Wherein, is the strain value at time point ; is the time point; is the index of the time point; is the strain sensitivity coefficient (the value range depends on the material type and structure, for example, the constantan alloy of the metal strain gauge is usually 2.0-2.2, and the silicon of the semiconductor is usually 100-170); is the real-time resistance value at time point ; is the initial resistance value.

[0034] It should be noted that the strain sensitivity coefficient is obtained by applying a known strain (e.g. 0-1000με) to the connector material under controlled loading conditions while recording the resistance change rate, calculating the ratio of strain to resistance change rate, and fitting the curve slope.

[0035] The stress value calculation expression is: ; wherein, is the stress value at time point ; is the material elastic modulus (value range depends on material type, for example, steel is usually 190-210 GPa).

[0036] It should be noted that the material elastic modulus is obtained by experimental calibration, specifically, the theoretical elastic modulus is obtained based on the material type, the stress is measured by applying a known strain in the experiment, and the theoretical value is corrected by linear fitting the slope of the stress-strain relationship curve to obtain the actual elastic modulus of the material under the current temperature and humidity and load environment.

[0037] By comparing the pre-processed optical fiber signal and the initial stress-strain data, the offset value is obtained; Further, on the unified time axis, the pre-processed optical fiber signal sequence and the initial stress-strain data sequence are correspondingly matched to ensure that the sampling times of the two signals are completely consistent; the time sequence change amount of the central wavelength is extracted from the pre-processed optical fiber signal, and the optical fiber signal strain value is obtained through the optical fiber strain calculation formula; the difference between the optical fiber signal strain value and the initial stress-strain data is calculated to obtain the offset value.

[0038] The initial stress-strain data is corrected by offset compensation to generate real-time stress-strain data; Further, the initial strain value in the initial stress-strain data at each time point and the offset value at the corresponding time point are read, the initial strain value and the offset value are added to obtain the corrected strain value; the corrected strain value is re-input into the stress value calculation expression to output the corrected stress value; the corrected strain value and the stress value at all time points are obtained and integrated into real-time stress-strain data.

[0039] S3, based on the real-time stress-strain data and the pre-processed external environment data, through wavelet multi-scale decomposition into high-frequency components and low-frequency components, and dynamically weighted fusion into corrected stress-strain data; The real-time stress-strain data and the pre-processed external environment data are subjected to wavelet multi-scale decomposition to generate a set of high-frequency components and a set of low-frequency components; Further, the real-time stress-strain data and the pre-processed external environment data are taken as input signals, db4 wavelet is selected as the mother wavelet function according to the principle of continuous wavelet transform, the characteristics of the input signals at different scales and times are localized and decomposed, and the change characteristics at different time scales are extracted; the number of decomposition layers is set according to the sampling frequency of the input signal, the signal is expanded in time and frequency dimensions at the same time, the decomposed input signal is obtained (for example, three layers of decomposition are set for the vibration signal with a sampling frequency of 5000 Hz), and the real-time stress-strain data and the pre-processed external environment data are point-by-point convoluted by the wavelet function to separate the high-frequency detail component and the low-frequency approximation component, the high-frequency detail component corresponds to the short-time rapid change characteristics in the decomposed input signal, and the low-frequency approximation component corresponds to the slow change long-term trend in the decomposed input signal; the wavelet decomposition is completed on the real-time stress-strain data and the pre-processed external environment data respectively, the high-frequency detail components and the low-frequency approximation components of the stress-strain signal, the temperature signal, the humidity signal and the vibration signal corresponding to the real-time stress-strain data at different scales are obtained, and the high-frequency component set and the low-frequency component set are integrated.

[0040] It should be noted that scale is the core variable in wavelet multi-scale, which is used to control the stretching ratio of wavelet basis function in time and frequency, and the scale is inversely proportional to the sampling frequency, which is usually determined by the sampling frequency of the input signal and the number of decomposition layers.

[0041] According to the high-frequency component set and the low-frequency component set, the environmental interference is identified, and the environmental interference characteristics are obtained. Further, for the temperature signal component, the humidity signal component and the vibration signal component in the high-frequency component set, the average energy change of each high-frequency component in the time sequence is calculated by the frequency energy analysis method, and the high-frequency interference characteristics are identified, for example, when the average energy change of the high-frequency component of the vibration signal increases to twice the original value in a short time, it is determined as transient mechanical impact interference; for the temperature signal component, the humidity signal component and the vibration signal component in the low-frequency component set, the average change rate and the fluctuation amplitude of the low-frequency component in the time sequence are calculated, when the signal has a persistent offset and a long-term cumulative trend, it is identified as a low-frequency trend interference characteristic, for example, when the mean value of the temperature signal component is monotonously increasing in a fixed time window (for example, 30 seconds) and the fluctuation amplitude is small, it is identified as temperature drift interference; the high-frequency interference characteristics and the low-frequency interference characteristics are time-weighted to generate environmental interference characteristics (the time weighting coefficients are determined according to the influence degree of each signal component on the stress-strain change, for example, the vibration signal has a large weight, and the humidity signal has a medium weight).

[0042] It should be noted that the frequency domain energy analysis method is a signal energy distribution evaluation method based on Fourier energy spectrum, which is used to calculate the energy proportion of signals in different frequency ranges. By comparing the energy proportion changes in different time periods, the change trend of high-frequency and low-frequency components in energy distribution is identified (for example, when the increase exceeds 50% of the reference value, it is determined that there is a transient mechanical impact or high-frequency vibration interference). The energy intensity is measured by performing frequency spectrum transformation on the high-frequency component and the low-frequency component respectively, and calculating the integral value of each signal power per unit time. High-frequency interference features include high-frequency energy mutation features, short-time amplitude sudden increase features and transient fluctuation features. Low-frequency interference features include low-frequency energy accumulation features, smooth trend offset features and long-term drift features.

[0043] Based on the environmental interference characteristics, the dynamic weight of each component is calculated by an exponential function, and different weight distribution is performed according to the components of different frequencies. Further, a nonlinear mapping relationship between environmental interference intensity and dynamic weight is established by an exponential function. The environmental interference intensity of each signal component at each time point is input into the exponential function, and the corresponding dynamic weight initial value is output. In the high-frequency component set, the change of the dynamic weight initial value shows an amplification effect on short-time fluctuation and transient energy mutation, so that the weight rapidly increases when the vibration impact or humidity rapidly changes. In the low-frequency component set, the change of the dynamic weight initial value shows a smoothing effect on long-term slowly varying trend, so that the weight slowly adjusts when the temperature drifts or the humidity accumulates, maintaining the stability of stress and strain estimation. When assigning weights to different frequency components, the influence proportion of the same source signal at different frequency levels is adjusted by the frequency band weighting coefficient, so that the component with high energy proportion is assigned a larger weight, and the component with low energy proportion is assigned a smaller weight, generating a dynamic weight sequence corresponding to the high-frequency component set and the low-frequency component set point by point on the time axis. It should be noted that the frequency band weighting coefficient is a proportional factor for measuring the degree of influence of different frequency components on stress and strain. The weight parameter is set according to the energy proportion and the degree of influence of each signal component, and the frequency band weighting coefficient is obtained after the energy proportion is normalized. The weighted high-frequency component and low-frequency component are fused into corrected stress and strain data. Further, the dynamic weight sequence is aligned with the corresponding energy amplitudes in the high-frequency component set and the low-frequency component set point by point in time, and the same scale is weighted and calculated, the energy amplitude of each time point is nonlinearly weighted by the dynamic weight sequence, the comprehensive interference influence of the external environment on the real-time stress-strain data is obtained, the environmental interference contribution is calculated, and the environmental interference contribution is decomposed into a high-frequency environmental correction term and a low-frequency environmental correction term through the wavelet reconstruction function in different scales; The high-frequency environmental correction term is subtracted from the high-frequency stress-strain signal, and the low-frequency environmental correction term is subtracted from the low-frequency stress-strain signal to obtain the high-frequency fusion component and the low-frequency fusion component after the environmental interference is suppressed; The high-frequency fusion component and the low-frequency fusion component of each scale are superimposed along the time axis through the wavelet reconstruction function to generate the corrected stress-strain data.

[0044] It should be noted that the energy amplitude refers to the local energy intensity obtained by the modulus square of the wavelet coefficient after the real-time stress-strain data and the preprocessed external environment data are decomposed by continuous wavelet transform; The wavelet reconstruction function is a function of recombining the components of each scale obtained by wavelet decomposition into a time domain signal. The inverse wavelet transform is performed on each scale of high-frequency detail component and low-frequency approximation component, and the overall structure of the signal is recovered in time sequence.

[0045] S4, based on the corrected stress-strain data, continuously monitor the health index of the connector, and analyze the risk trend of the connector, and generate real-time monitoring feedback when the health index sequence exceeds the dynamic safety threshold; The corrected stress-strain data is weighted and normalized to generate a health index sequence; Further, the continuous wavelet transform is performed on the corrected stress-strain data, the modulus square of the wavelet coefficient is read, and the energy amplitude sequence is generated point by point; The energy amplitude sequence is multiplied point by point with the corrected stress-strain data as a weighting factor to generate a weighted stress sequence and a weighted strain sequence; The mean and standard deviation of the weighted stress sequence and the weighted strain sequence are calculated based on the sliding statistical window, and Z-score normalization is performed to obtain a normalized weighted stress sequence and a normalized weighted strain sequence; The normalized weighted stress sequence and the normalized weighted strain sequence are fused to generate a health index sequence.

[0046] It should be noted that the wavelet coefficient is a set of numerical values obtained by convoluting the mother wavelet function at different scales and different time positions with the multi-physical quantity signal, which is used to quantify the energy size of the multi-physical quantity signal.

[0047] The historical health index sequence is collected, and an ARIMA model is trained through an autocorrelation function; Further, the health index data of multiple time periods within the time window is continuously collected to generate a historical health index sequence, the historical health index sequence is subjected to difference and stationary processing, the degradation of the device performance and the load of the environmental operation are eliminated by calculating the change amount between adjacent time points, and a stationary health index sequence is obtained; the stationary health index sequence is input into an autocorrelation function and a partial autocorrelation function, the correlation coefficients and the partial correlation coefficients under different lag orders are calculated, a lag order correlation distribution is output, the order parameters of the ARIMA model are set according to the lag order correlation distribution, the autoregressive order (for example, p=1, the difference order d=1, and the moving average order q=1) is set, and the ARIMA model order parameters are generated; the stationary health index sequence and the ARIMA model order parameters are input into the ARIMA model for training, and the steps are as follows: the autoregressive coefficients, the moving average coefficients and the constant term are calculated by the maximum likelihood estimation method, the residual error is calculated, and the white noise test is performed; when the autocorrelation coefficients of the residual error sequence are close to zero at all lag orders, the training is completed, and the trained ARIMA model is output.

[0048] It should be noted that the maximum likelihood estimation method is a parameter estimation method, specifically, the maximum likelihood estimation method measures the possibility of observing the current data under different parameter values by constructing a likelihood function, and finds the parameter value that maximizes the possibility of the current data by using an optimization algorithm; the autocorrelation function is used to measure the linear correlation degree of the time sequence under different lag orders, to judge the periodicity and trend continuity of the data; the partial autocorrelation function is used to exclude the influence of the intermediate lag term and calculate the direct correlation degree of the time delay step.

[0049] The health index sequence is input into the trained ARIMA model to predict the future short-term health index sequence, and a risk trend slope is generated. Further, the health index sequence is input into the trained ARIMA model, a short-term predicted health index sequence difference domain is iteratively generated in a one-step rolling manner, the short-term predicted health index sequence difference domain is subjected to inverse difference reconstruction, is accumulated in time sequence and aligned with the benchmark, and a short-term predicted health index sequence is generated; the short-term predicted health index sequence and the corresponding time index are subjected to least squares linear fitting in the prediction interval, and the slope of the fitted straight line is taken as the risk trend slope, which is used to quantify the rising or falling speed of the health index sequence in the short term.

[0050] The health index sequence is subjected to rolling statistical calculation, and a dynamic safety threshold is generated by combining the risk trend slope. Further, based on the health index sequence, the rolling mean and the rolling standard deviation corresponding to each time point are calculated in a fixed-length sliding window (for example, the window length is 60 sampling points) to generate a rolling statistic sequence; a nonlinear mapping relationship of the threshold gain coefficient is established according to the sign of the risk trend slope, and the risk trend slope is mapped to the threshold gain coefficient; The threshold gain coefficient expression is: ; Among them, is the threshold gain coefficient of time point t; is the adjustment coefficient; is the risk trend slope of time point t; The rolling standard deviation is multiplied by the corresponding threshold gain coefficient at each time point to form a dynamic standard deviation sequence, with the threshold gain coefficient sequence as the weight; when the risk trend slope is positive (the health index rises), the threshold gain coefficient is greater than 1, and the rolling standard deviation is amplified; when the risk trend slope is negative (the health index decreases), the threshold gain coefficient is less than 1, and the rolling standard deviation is compressed; The rolling mean and the dynamic standard deviation are aligned point by point in time, and a linear combination operation is performed to generate a dynamic safety threshold. The exemplary value range of the dynamic safety threshold is usually 0.6-1.2. When the risk trend slope is positive, the dynamic safety threshold moves up to expand the safety range to avoid false positives. When the risk trend slope is negative, the dynamic safety threshold moves down to narrow the safety range to strengthen monitoring. The dynamic safety threshold expression is set as: ; Among them, is the dynamic safety threshold of time point ; is the rolling mean; is the safety relaxation coefficient; is the dynamic standard deviation; It should be noted that the adjustment coefficient is set based on the trend change sensitivity and time response demand of the health index sequence. The exemplary value range is usually between 0.05 and 0.3. The safety relaxation coefficient is set based on the statistical distribution characteristics of the historical health index sequence and the risk tolerance. The safety relaxation coefficient is set by performing statistics on the historical health index data in the training stage. The exemplary value range is usually between 1.2 and 2.5.

[0051] When the health index sequence exceeds the dynamic safety threshold, the abnormal index value, the abnormal timestamp, the associated parameters and the abnormal level are recorded to generate real-time monitoring feedback. Further, at the same time point, the health indicator sequence is compared with the dynamic safety threshold point by point, and the positions where the health indicator sequence is greater than the dynamic safety threshold are marked to form an out-of-bound marking sequence. The values at the time points marked in the out-of-bound marking sequence are taken as abnormal indicator values, and the time indexes corresponding to the time points are taken as abnormal time stamps. Meanwhile, the stress value and strain value at the corresponding time point are extracted from the corrected stress-strain data, and the rolling mean and dynamic standard deviation at the corresponding time point are extracted from the rolling mean and dynamic standard deviation. The scalar values at the corresponding time point are extracted from the risk trend slope, temperature signal, humidity signal, and vibration signal, and are integrated into correlation parameters, including rolling mean, dynamic standard deviation, risk trend slope, stress value, strain value, temperature parameter, humidity parameter, and vibration parameter. The difference between the health indicator sequence and the dynamic safety threshold at the same time point sequence is calculated, and the intensity score is obtained in combination with the dynamic standard deviation sequence. The comprehensive abnormal score is calculated by weighting the intensity score and the abnormal time stamp. When the comprehensive abnormal score is greater than the high abnormal threshold, it is marked as high-level abnormality. When the comprehensive abnormal score is between the medium abnormal threshold and the high abnormal threshold, it is marked as medium-level abnormality. When the comprehensive abnormal score is less than the medium abnormal threshold, it is marked as low-level abnormality, and an abnormal level sequence is generated. The abnormal indicator value, abnormal time stamp, correlation parameter, and abnormal level are integrated into real-time monitoring feedback.

[0052] It should be noted that the high abnormal threshold and the medium abnormal threshold are calculated based on the historical health indicator sequence to obtain the historical comprehensive abnormal score in the corresponding time period, and the medium threshold and the high threshold are set according to the confidence level of the distribution of the historical comprehensive abnormal score (for example, the medium threshold corresponds to the 95th percentile position of the score distribution, and the high threshold corresponds to the 99th percentile position of the score distribution). The example value range of the medium threshold is 1.5-2.0, and the example value range of the high threshold is 2.5-3.0.

[0053] S5, according to the real-time monitoring feedback and the real-time stress-strain data, a health state knowledge graph is constructed, and the failure type and occurrence probability of the connector are predicted to generate a failure prediction report; The time stamps of the real-time monitoring feedback and the real-time stress-strain data are aligned, and the abnormal features and time sequence features are extracted respectively to form a multi-dimensional feature vector sequence. Further, the abnormal time stamps in the real-time monitoring feedback and the time indexes in the real-time stress-strain data are read, the time points are aligned through inner connection, the abnormal indicator values, correlation parameters, and abnormal levels in the real-time monitoring feedback are extracted, and an abnormal feature sequence is generated by combination. The stress value and strain value are read point by point from the real-time stress-strain data, and the first-order difference, sliding mean, and sliding standard deviation of the stress value and strain value are obtained in a fixed sliding window (for example, the sliding window length is 60 sampling points) to generate a time sequence feature sequence. The abnormal feature sequence and the time sequence feature sequence are concatenated point by point into a single vector under the same time stamp, and are stacked in time order to form a multi-dimensional feature vector sequence.

[0054] Based on the multi-dimensional feature vector sequence, the transformation relationship edge between the health state node of the connector and the health state in the health state knowledge graph is defined in combination with the historical health index sequence, and a health state knowledge graph framework is formed. Further, based on the multi-dimensional feature vector sequence, in combination with the historical health index sequence, the health state of each time point of the multi-dimensional feature vector sequence is determined, and the steps are as follows: when the abnormal level sequence is a high-level abnormality and the stress value reaches the material elastic safety limit threshold, it is marked as “overload”; when the abnormal level sequence is a medium-level abnormality, it is marked as “early warning”; when the abnormal level sequence is a low-level abnormality and the value of the health index sequence does not continuously exceed the dynamic safety threshold (for example, 3 times of not continuously exceeding the dynamic safety threshold), it is marked as “normal”; when the abnormal level sequence reaches the material elastic safety limit threshold multiple times (for example, 3 times or more) in a time window and is accompanied by a slow accumulation of strain value, it is marked as “fatigue evolution”, and a health state label sequence is generated; the multi-dimensional feature vector sequence is grouped and aggregated according to the health state label sequence, the center vector and the distribution range of each health state in the feature space are calculated, and a health state node set described by the state name, the center vector and the statistical boundary is formed; the adjacent time point transfer statistics of the health state label sequence is performed, the transfer number and the transfer probability from any health state to another health state are obtained, and the average stay duration and the transfer trigger condition are recorded, and a health state transformation relationship edge set is generated; the health state node set and the health state transformation relationship edge set are uniformly indexed according to the node identification and the time sequence, and assembled into a health state knowledge graph framework containing the health state node attribute (center vector and statistical boundary) and the transformation relationship edge attribute (transfer probability, average stay duration, trigger condition) between health states.

[0055] It should be noted that the material elastic safety limit threshold is usually set to 60% to 85% of the material elastic safety limit, which is the maximum safe stress value that the connector material can withstand in the elastic stage, and is set based on the material mechanics performance parameters and in combination with the environmental correction coefficient.

[0056] The multi-dimensional feature vector sequence is mapped into the health state knowledge graph framework, the multi-dimensional feature vector sequence is grouped and associated with the state node through a clustering algorithm, the health state knowledge graph is constructed, and an embedded vector sequence is generated. Further, the number of health state node sets is read as the number of clusters, the multi-dimensional feature vector sequence is clustered using K-Means to generate a cluster label sequence and a cluster center set; the cluster center set and the center vector in the health state node set are calculated one by one to find the center vector of the health state node with the smallest Euclidean distance for each cluster center, and a corresponding mapping relationship is established, and finally a mapping relationship of "cluster center→health state node" is formed, and the mapping relationship is applied to the cluster label sequence to generate a state association sequence; the state association sequence is read at adjacent time points, the occurrence times of each pair of adjacent states are accumulated and updated with the transition probability and the average stay time in the health state transition relationship edge set, and a health state knowledge graph is generated; Node2Vec random walk is performed on the health state knowledge graph to obtain a vector representation of each health state node, and the health state node vectors corresponding to each time point in the state association sequence are arranged in time sequence to generate an embedding vector sequence.

[0057] According to historical fault data, the parameters of the multi-layer perceptron are adjusted through the loss function to generate a trained multi-layer perceptron; Further, the embedding vector sequence is aligned with the historical fault data point by point, a training sample set and a validation sample set are constructed with the embedding vector as the feature and the fault type and occurrence label in the historical fault data as the supervision label, a full connection structure of the input layer, the hidden layer and the output layer of the multi-layer perceptron is established, a cross-entropy loss function is used as the optimization objective, and the Adam optimization algorithm is used to perform forward calculation of the cross-entropy loss function value in small batch iteration and reverse propagation according to the gradient of the loss function to update the weights and biases of the multi-layer perceptron; weight decay is used to suppress overfitting, and the cycle is repeated according to the cycle until the validation sample set loss function no longer decreases to trigger stopping, and a trained multi-layer perceptron is generated.

[0058] It should be noted that the historical fault data is derived from the historical monitoring records and maintenance records of the connector during long-term operation, and the historical fault data includes time index, fault type, occurrence label, operating environment parameters, historical repair and maintenance information, and stress-strain reference value.

[0059] The embedding vector sequence is input into the trained multi-layer perceptron to output a fault type probability distribution and an occurrence probability value; Further, the embedding vector sequence is input into the trained multi-layer perceptron at the same time point sequence in sequence, forward calculation is performed, and normalized probability transformation is performed at the output layer to obtain a fault type probability distribution corresponding to each time point, the class corresponding to the maximum probability is read as the fault type at each point on the fault type probability distribution, and the corresponding probability value is read as the fault occurrence probability value.

[0060] formatting the fault type probability distribution and the occurrence probability value, and obtaining a fault prediction report; Further, the fault type probability distribution and the occurrence probability value are point by point corresponding and uniformly formatted, the fault type probability distribution at each time point is converted into percentage form, and the fault type corresponding to the maximum probability is extracted, and the occurrence probability value at the corresponding time point is recorded together to form a fault prediction report.

[0061] S6, based on the fault prediction report, combining historical fault data, obtaining the individual maintenance plan of the connector; Respectively extract the maintenance features and fault features in the fault prediction report and the historical fault data, and generate a multi-dimensional maintenance feature vector; Further, the maintenance and maintenance information (including maintenance action, maintenance time, downtime, component replacement record and cost) in the fault prediction report and the historical fault data are matched according to time index, and the fault features (fault type, occurrence probability value and fault type probability distribution) are numerically valued, and the maintenance features (maintenance action, maintenance time, downtime, component replacement record and cost) are encoded and standardized, All features are concatenated in sequence at the same time point to generate a multi-dimensional maintenance feature vector.

[0062] Based on the multi-dimensional maintenance feature vector, the K nearest neighbor algorithm is used to retrieve the maintenance plan of the similar fault case in the historical fault data, and a maintenance plan candidate set of the connector is generated; Further, the K nearest neighbor algorithm is used to calculate the Euclidean distance between the multi-dimensional maintenance feature vector and the historical multi-dimensional maintenance feature vector library in the feature space, and the nearest neighbor samples with the smallest distance (for example, the number of neighbors is 5) are selected under the standardization measurement, The maintenance action, maintenance time, downtime, component replacement record and cost record corresponding to the nearest neighbor sample are extracted as the maintenance plan of the similar fault case, and the maintenance plan of the similar fault case is integrated in ascending order of time index and distance, Output the maintenance plan candidate set of the connector.

[0063] It should be noted that the historical multi-dimensional maintenance feature vector library is generated by uniformly encoding and standardizing the fault features and maintenance features contained in the historical fault data.

[0064] The maintenance plan candidate set of the connector is optimized by genetic algorithm, and the individual maintenance plan of the connector is generated.

[0065] Further, the entries of the maintenance plan candidate set of each connector are encoded into chromosomes, the maintenance actions, maintenance time length, downtime length, component replacement records and cost records are taken as gene contents, a multi-objective fitness function containing cost minimization, downtime length minimization and risk reduction amplitude maximization is established, and is synthesized into a single fitness score in a weighted manner; the chromosome population is constantly updated by selection, crossover and mutation operations to evolve iteratively in the population, constraint checking is performed on each generation population to ensure that the maintenance time length does not exceed the available maintenance window, the iteration is terminated when the fitness improvement of continuous generations is less than a preset fitness threshold or reaches an upper limit of generation number, the chromosome with the highest fitness is output and decoded into a maintenance action list, execution order, predicted maintenance time length, predicted downtime length and predicted cost, and a personalized maintenance plan of the connector is generated.

[0066] It should be noted that the available maintenance window refers to the time period during which the connector is allowed to perform maintenance operations in the operation cycle, which is set based on the operation plan; the fitness threshold is set based on the normalized scale of the multi-objective function and the actual optimization sensitivity, and the exemplary value range is generally 1x10 -4 ~ 1x10 -2 ; the upper limit of generation number is determined empirically according to the dimension of the multi-dimensional maintenance feature vector and the population size, and the exemplary value range is generally 100-300 generations.

[0067] The embodiment also provides a connector stress-strain real-time monitoring system, comprising: a data acquisition module for acquiring and preprocessing multi-physical quantity data and external environment data; a real-time calculation module for extracting current and voltage signal sequences from the preprocessed multi-physical quantity data and calculating real-time stress-strain data through an electrical model; a dynamic weighting module for decomposing the real-time stress-strain data and the preprocessed external environment data into high-frequency components and low-frequency components through wavelet multi-scale decomposition, and dynamically weighting and fusing them into corrected stress-strain data; a risk monitoring module for continuously monitoring the health indicators of the connector based on the corrected stress-strain data, and analyzing the risk trend of the connector, and generating a real-time monitoring feedback when the health indicator sequence exceeds a dynamic safety threshold; a fault prediction module for constructing a health state knowledge graph according to the real-time monitoring feedback and the real-time stress-strain data, and predicting the fault type and occurrence probability of the connector, and generating a fault prediction report; a fault maintenance module for obtaining a personalized maintenance plan of the connector based on the fault prediction report and combining historical fault data.

[0068] To sum up, the present application realizes the accurate extraction of multi-scale signal features and noise suppression by decomposing real-time stress-strain data and preprocessed external environment data into high-frequency components and low-frequency components through the wavelet multi-scale decomposition step, effectively improving the reliability and anti-interference ability of the monitoring data. At the same time, the real-time monitoring feedback and stress-strain data are mapped into a structured knowledge network through the step of constructing a health state knowledge graph, realizing the deep semantic analysis of the health state and the early warning of the fault.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A connector stress-strain real-time monitoring method, characterized by: The method comprises the following steps: Collecting and preprocessing multi-physical quantity data and external environment data; Extracting current and voltage signal sequences from the preprocessed multi-physical quantity data, and calculating real-time stress and strain data through an electrical model; Based on the real-time stress and strain data and the preprocessed external environment data, the high-frequency components and low-frequency components are decomposed by wavelet multi-scale decomposition, and the modified stress and strain data are dynamically weighted and fused; Based on the modified stress and strain data, the health indicators of the connector are continuously monitored, and the risk trend of the connector is analyzed, and real-time monitoring feedback is generated when the health indicator sequence exceeds the dynamic safety threshold; According to the real-time monitoring feedback and the real-time stress and strain data, a health status knowledge graph is constructed, and the failure type and occurrence probability of the connector are predicted, and a failure prediction report is generated; Based on the failure prediction report, combined with historical failure data, the individual maintenance plan of the connector is obtained.

2. The connector strain real-time monitoring method of claim 1, wherein: The multi-physical quantity data includes strain signals, optical fiber signals, current signals and voltage signals; The external environment data includes temperature signals, humidity signals and vibration signals; The preprocessing includes denoising, standardization, time synchronization and data alignment.

3. The connector strain real-time monitoring method of claim 2, wherein: The steps of extracting current and voltage signal sequences from the preprocessed multi-physical quantity data and calculating real-time stress and strain data through an electrical model are as follows: Separate the current signal sequence and the voltage signal sequence from the preprocessed multi-physical quantity data; Based on Ohm's law and Hooke's law, the mathematical relationship between current, voltage and stress and strain is established, and an electrical model is generated; Input the current signal sequence and the voltage signal sequence into the electrical model to output initial stress and strain data; Compare the preprocessed optical fiber signal with the initial stress and strain data to obtain an offset value; Correct the initial stress and strain data through offset compensation to generate real-time stress and strain data.

4. The connector strain real-time monitoring method of claim 3, wherein: The steps of decomposing the real-time stress and strain data and the preprocessed external environment data into high-frequency components and low-frequency components by wavelet multi-scale decomposition, and dynamically weighting and fusing them into modified stress and strain data are as follows: Wavelet multi-scale decomposition is performed on the real-time stress and strain data and the preprocessed external environment data to generate a high-frequency component set and a low-frequency component set; Identify environmental interference and obtain environmental interference characteristics according to the high-frequency component set and the low-frequency component set; Based on the environmental interference characteristics, calculate the dynamic weight of each component by an exponential function, and assign different weights to components of different frequencies; Fuse the weighted high-frequency components and low-frequency components into modified stress and strain data.

5. The connector strain real-time monitoring method of claim 4, wherein: The steps of continuously monitoring the health indicators of the connector based on the modified stress and strain data, and analyzing the risk trend of the connector are as follows: Weighted normalization is performed on the modified stress and strain data to generate a health indicator sequence; Collect historical health indicator sequences and train an ARIMA model through autocorrelation function; Input the health indicator sequence into the trained ARIMA model to predict the future short-term health indicator sequence and generate a risk trend slope.

6. The connector strain real-time monitoring method of claim 5, wherein: The steps of generating real-time monitoring feedback when the health indicator sequence exceeds the dynamic safety threshold are as follows: Rolling statistics calculation is performed on the health indicator sequence, and the dynamic safety threshold is dynamically adjusted in combination with the risk trend slope. When the health index sequence exceeds the dynamic safety threshold, record the abnormal index value, abnormal timestamp, correlation parameter and abnormal level, and generate real-time monitoring feedback.

7. The connector strain real-time monitoring method of claim 6, wherein: The health state knowledge graph is constructed based on the real-time monitoring feedback and real-time stress-strain data, and the steps are as follows, Align the timestamps of the real-time monitoring feedback and the real-time stress-strain data, and extract abnormal features and time series features respectively, and integrate them into a multi-dimensional feature vector sequence; Based on the multi-dimensional feature vector sequence, the health state nodes of the connector and the transformation relationship between the health states are defined in the health state knowledge graph based on the historical health index sequence, and a health state knowledge graph framework is formed; The multi-dimensional feature vector sequence is mapped into the health state knowledge graph framework, and the multi-dimensional feature vector sequence is grouped and associated with the state nodes by clustering algorithm to construct the health state knowledge graph, and an embedding vector sequence is generated.

8. The connector strain real-time monitoring method of claim 7, wherein: The fault type and occurrence probability of the connector are predicted, and a fault prediction report is generated, and the steps are as follows, According to the historical fault data, the multilayer perceptron is adjusted by the loss function to generate the trained multilayer perceptron; The embedding vector sequence is input into the trained multilayer perceptron, and the fault type probability distribution and occurrence probability value are output; The fault type probability distribution and occurrence probability value are formatted, and the fault prediction report is obtained.

9. The connector strain real-time monitoring method of claim 8, wherein: Based on the fault prediction report, the individual maintenance plan of the connector is obtained based on the historical fault data, and the steps are as follows, Respectively extract the maintenance features and fault features in the fault prediction report and the historical fault data to generate a multi-dimensional maintenance feature vector; Based on the multi-dimensional maintenance feature vector, the K nearest neighbor algorithm is used to retrieve the maintenance plan of the similar fault case in the historical fault data to generate a maintenance plan candidate set of the connector; The maintenance plan candidate set of the connector is optimized by genetic algorithm, and the individual maintenance plan of the connector is generated.

10. A connector stress-strain real-time monitoring system based on the connector stress-strain real-time monitoring method according to any one of claims 1-9, characterized in that: It includes, The data acquisition module is used for collecting multi-physical quantity data and external environment data and pre-processing; The real-time calculation module is used for extracting the current and voltage signal sequence in the pre-processed multi-physical quantity data, and calculating the real-time stress-strain data by the electrical model; The dynamic weighting module is used for decomposing the real-time stress-strain data and the pre-processed external environment data into high-frequency components and low-frequency components by wavelet multi-scale decomposition, and dynamically weighting and fusing them into corrected stress-strain data; The risk monitoring module is used for continuously monitoring the health index of the connector based on the corrected stress-strain data, and analyzing the risk trend of the connector, and generating real-time monitoring feedback when the health index sequence exceeds the dynamic safety threshold; The fault prediction module is used for constructing the health state knowledge graph based on the real-time monitoring feedback and the real-time stress-strain data, and predicting the fault type and occurrence probability of the connector, and generating a fault prediction report; The fault maintenance module is used for obtaining the individual maintenance plan of the connector based on the fault prediction report and the historical fault data.

Images