Mechanical and electrical equipment whole life cycle management method and system based on internet of things

Abstract

The application relates to the technical field of industrial Internet of Things, and discloses a full-life-cycle management method and system for electromechanical equipment based on the Internet of Things, which comprises the following steps: acquiring a load sequence and a temperature sequence, denoising the pure temperature sequence and the pure load sequence, and identifying a temperature mutation interval; extracting a load frequency feature and an operation section sequence according to the pure load sequence and the temperature mutation interval, and constructing a loss feature vector; performing generalized cross-correlation analysis on the pure load sequence and the pure temperature sequence to obtain a synchronization offset sequence, and quantifying the loss rate index; using the loss feature vector and the loss rate index to obtain a loss coefficient through a prediction model, calculating and correcting a final performance deviation vector; performing secondary clustering on the final performance deviation vector to obtain a refined trend subcluster; filtering the pure wear evolution sequence by removing the influences of the load and the temperature according to the refined trend subcluster, the pure load sequence and the pure temperature sequence, and predicting a remaining life curve. The method improves the full-life-cycle management efficiency.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) technology, and in particular to a method and system for the full lifecycle management of electromechanical equipment based on IoT. Background Technology

[0002] Currently, with the widespread adoption of IoT technology, the full lifecycle management of electromechanical equipment requires real-time data sensing to address the challenge of dynamically changing internal wear. During equipment operation, the wear coefficient constantly changes due to load fluctuations and environmental factors, making it difficult to accurately separate true wear characteristics from conventional monitoring. In online mode, continuous collection of dynamic data and intelligent analysis of wear trends are essential. This necessitates IoT chips to achieve real-time data sensing, combined with big data processing technology to deeply analyze massive amounts of information, thereby identifying the dominant factors of internal wear and achieving accurate assessment of equipment performance degradation.

[0003] In one existing technology, firstly, based on the design manual and historical operation and maintenance experience provided by the equipment manufacturer, fixed alarm thresholds and shutdown maintenance cycles are preset for various key parameters of the equipment (such as vibration amplitude, bearing temperature, current fluctuations, etc.). During operation, routine data collection and recording rely on basic sensors installed on the equipment. When the predetermined operating time or cycle is reached, the equipment is completely shut down. While the equipment is shut down, technicians use portable testing instruments (such as handheld vibration analyzers or infrared thermal imagers) to manually measure key parts of the equipment at fixed points, collecting data in static or offline states. Subsequently, the collected data is compared one by one with the previously set fixed thresholds. If any measured value exceeds the preset threshold range, it is determined that there is an abnormality or wear in the corresponding part of the equipment, and a maintenance report is generated accordingly, arranging for component replacement or maintenance. The entire process relies on preset, unchanging threshold standards and fixed time cycle triggers. Fixed thresholds cannot match dynamically changing actual wear, and periodic shutdown inspections are difficult to capture the state evolution under continuous operating conditions, leading to delayed or redundant maintenance decisions.

[0004] Therefore, existing technologies suffer from low efficiency in full lifecycle management. Summary of the Invention

[0005] This invention provides a method and system for the full lifecycle management of electromechanical equipment based on the Internet of Things, so as to improve the efficiency of full lifecycle management.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for full lifecycle management of electromechanical equipment based on the Internet of Things, comprising:

[0007] Obtain the load sequence and temperature sequence of the electromechanical equipment to obtain the original dataset;

[0008] Based on the original dataset, denoising and smoothing processes are performed to obtain a clean temperature sequence and a clean load sequence. Then, based on the clean temperature sequence, abrupt change intervals are identified to obtain temperature abrupt change intervals.

[0009] Based on the pure load sequence and the temperature jump interval, the loss features are initially extracted to obtain the load frequency features and the operating segment sequence. Then, a loss feature vector is constructed based on the load frequency features, the temperature jump interval, and the operating segment sequence.

[0010] Generalized cross-correlation analysis is performed on the pure load sequence and the pure temperature sequence to obtain the synchronization offset sequence, and the internal loss rate is quantified based on the synchronization offset sequence to obtain the loss rate index.

[0011] Based on the loss feature vector and the loss rate index, the loss coefficient is predicted using a pre-built loss coefficient prediction model to obtain the loss coefficient.

[0012] Based on the loss coefficient, a performance deviation vector containing load loss deviation, temperature response deviation, and duration accumulation deviation is calculated, and the duration accumulation deviation is periodically corrected based on the load loss deviation and the temperature response deviation to obtain the final performance deviation vector.

[0013] Based on the final performance deviation vector, evolution pattern clustering is performed to obtain loss evolution trend clusters, and loss acceleration analysis and secondary clustering are performed based on the loss evolution trend clusters to obtain refined trend sub-clusters.

[0014] Based on the refined trend sub-cluster, the pure load sequence, and the pure temperature sequence, the initial wear sequence, the load impact component, and the temperature response component are separated. The influence of the load impact on the initial wear sequence is filtered out based on the load impact component to obtain the impact-filtered wear sequence.

[0015] Based on the impact-filtered wear sequence and the temperature response component, the influence of temperature cycle changes is filtered out to obtain a pure wear evolution sequence. Based on the pure wear evolution sequence, lifetime prediction is performed to obtain the remaining lifetime curve.

[0016] Secondly, the present invention provides an IoT-based electromechanical equipment lifecycle management system, comprising:

[0017] The data acquisition module is used to acquire the load sequence and temperature sequence of electromechanical equipment to obtain the raw dataset;

[0018] The denoising module is used to perform denoising and smoothing processing on the original dataset to obtain a clean temperature sequence and a clean load sequence, and to identify temperature mutation intervals based on the clean temperature sequence.

[0019] The loss feature construction module is used to perform preliminary loss feature extraction based on the pure load sequence and the temperature change interval to obtain load frequency features and operating segment sequence, and construct a loss feature vector based on the load frequency features, the temperature change interval and the operating segment sequence.

[0020] The loss rate quantization module is used to perform generalized cross-correlation analysis based on the pure load sequence and the pure temperature sequence to obtain a synchronization offset sequence, and to perform internal loss rate quantization based on the synchronization offset sequence to obtain a loss rate index.

[0021] The loss coefficient prediction module is used to predict the loss coefficient based on the loss feature vector and the loss rate index using a pre-built loss coefficient prediction model, and obtain the loss coefficient.

[0022] The performance deviation analysis module is used to calculate a performance deviation vector containing load loss deviation, temperature response deviation and time accumulation deviation based on the loss coefficient, and to periodically correct the time accumulation deviation based on the load loss deviation and the temperature response deviation to obtain the final performance deviation vector.

[0023] The evolution trend analysis module is used to perform evolution pattern clustering and grouping based on the final performance deviation vector to obtain loss evolution trend clusters, and to perform loss acceleration analysis and secondary clustering based on the loss evolution trend clusters to obtain refined trend subclusters.

[0024] The loss analysis module is used to separate the initial wear sequence, load impact component and temperature response component based on the refined trend sub-cluster, the pure load sequence and the pure temperature sequence, and to filter out the influence of load impact on the initial wear sequence based on the load impact component to obtain the impact-filtered wear sequence.

[0025] The life prediction module is used to filter out the influence of temperature cycle changes based on the impact-filtered wear sequence and the temperature response component to obtain a pure wear evolution sequence, and to perform life prediction based on the pure wear evolution sequence to obtain the remaining life curve.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] (1) This invention uses first-order low-pass filtering and sliding window gradient analysis to denoise and identify abrupt change intervals of the original load and temperature sequences. It combines short-time Fourier transform and running segment segmentation to construct a multi-dimensional loss feature vector, thereby realizing refined feature extraction of the operating status of electromechanical equipment and providing highly discriminative feature input for subsequent loss analysis.

[0028] (2) This invention uses a generalized cross-correlation analysis method to quantify the synchronous shift of load and temperature sequences, and combines a neural network to construct a loss coefficient prediction model. Through periodic correction of performance deviation vector and DBSCAN clustering, it achieves accurate capture and pattern grouping of the dynamic evolution trend of equipment loss, thereby improving the timeliness and accuracy of condition assessment.

[0029] (3) Based on STL decomposition and local weighted regression, this invention filters out the effects of load impact and temperature cycle, extracts the pure wear evolution sequence, and outputs the remaining life curve through the life prediction model, so as to realize the accurate prediction of the remaining life of electromechanical equipment and support preventive maintenance and full life cycle management decision-making. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the process for the Internet of Things-based full lifecycle management method for electromechanical equipment provided in the first embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the structure of the Internet of Things-based electromechanical equipment lifecycle management system provided in the second embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Reference Figure 1 The first embodiment of the present invention provides a method for full lifecycle management of electromechanical equipment based on the Internet of Things, including the following steps:

[0034] S11, Obtain the load sequence and temperature sequence of the electromechanical equipment to obtain the original dataset;

[0035] S12, based on the original dataset, perform denoising and smoothing processing to obtain a clean temperature sequence and a clean load sequence, and identify temperature mutation intervals based on the clean temperature sequence.

[0036] S13, Based on the pure load sequence and the temperature jump interval, perform preliminary loss feature extraction to obtain load frequency features and operating segment sequence, and construct a loss feature vector based on the load frequency features, the temperature jump interval and the operating segment sequence;

[0037] S14. Perform generalized cross-correlation analysis on the pure load sequence and the pure temperature sequence to obtain the synchronization offset sequence, and quantify the internal loss rate based on the synchronization offset sequence to obtain the loss rate index.

[0038] S15, Based on the loss feature vector and the loss rate index, the loss coefficient is predicted using a pre-built loss coefficient prediction model to obtain the loss coefficient.

[0039] S16. Based on the loss coefficient, a performance deviation vector including load loss deviation, temperature response deviation and duration accumulation deviation is calculated, and the duration accumulation deviation is periodically corrected based on the load loss deviation and the temperature response deviation to obtain the final performance deviation vector.

[0040] S17. Based on the final performance deviation vector, perform evolution pattern clustering to obtain loss evolution trend clusters, and perform loss acceleration analysis and secondary clustering based on the loss evolution trend clusters to obtain refined trend sub-clusters.

[0041] S18, based on the refined trend sub-cluster, the pure load sequence, and the pure temperature sequence, separate the initial wear sequence, the load impact component, and the temperature response component, and filter out the influence of the load impact on the initial wear sequence based on the load impact component to obtain the impact-filtered wear sequence;

[0042] S19. Based on the impact-filtered wear sequence and the temperature response component, the influence of temperature cycle changes is filtered out to obtain a pure wear evolution sequence. Based on the pure wear evolution sequence, lifetime prediction is performed to obtain the remaining lifetime curve.

[0043] In step S11, the load sequence and temperature sequence of the electromechanical equipment are obtained to obtain the original dataset.

[0044] Specifically, IoT sensors deployed in critical parts of electromechanical equipment (such as motor main bearing housings and gearbox housings) are responsible for collecting load and temperature signals. Load sensors utilize current transformers or power transmitters, while temperature sensors employ thermocouples or infrared temperature measurement units. All sensors operate at a fixed sampling frequency, set according to the Nyquist theorem. The highest signal frequency is determined through spectral analysis of historical data, and the sampling frequency is set to at least twice this highest frequency. The analog signals output by the sensors are converted from analog to digital and then uploaded to a cloud server via the MQTT protocol. The server aligns the data according to timestamps to form the original dataset.

[0045] In step S12, based on the original dataset, denoising and smoothing processing is performed to obtain a clean temperature sequence and a clean load sequence. Then, based on the clean temperature sequence, abrupt change intervals are identified to obtain temperature abrupt change intervals, including:

[0046] Based on the original dataset, a first-order low-pass filtering algorithm is used for denoising, and the data is smoothed by linear interpolation based on the denoising result to obtain a first-filter load sequence and a pure temperature sequence.

[0047] Based on the pure temperature sequence, the temperature gradient is calculated using the sliding window method to obtain the temperature gradient sequence;

[0048] The time points when the temperature gradient sequence exceeds a preset gradient threshold are marked as abrupt change times. The consecutive abrupt change times are merged to obtain a temperature abrupt change interval that includes temperature values ​​and time points.

[0049] The first-filtered load sequence corresponding to the mutation time is extracted, and a second filtering is performed using a first-order low-pass filtering algorithm to obtain a clean load sequence.

[0050] Specifically, firstly, first-order low-pass filtering is applied to both the load value sequence and the temperature value sequence to achieve noise reduction. The filtering process is performed independently for each data point in the sequence. For the current data point to be processed, the calculation of its filtered output value depends on the original input value of that point and the filtered output value of the previous data point. The specific calculation process involves multiplying the current original input value by a filter coefficient, then multiplying the filtered output value of the previous data point by the difference between this filter coefficient and the original input value, and finally adding these two products together. The sum is the filtered output value of the current point. Determining the filter coefficient requires analyzing the noise spectrum characteristics in historical data. Specifically, a historical load value sequence and a corresponding historical temperature value sequence under stable operating conditions are selected. Fast Fourier Transforms are performed on both sequences to obtain their respective spectra. Energy peaks with frequencies higher than the normal operating frequency range of the equipment (determined by the equipment design parameters and historical operating spectrum analysis) are identified in each spectrum. The arithmetic mean of the frequencies corresponding to all identified peaks is calculated. This average is used as the cutoff frequency reference value, and the final filter coefficient value is calculated from this cutoff frequency reference value based on the digital design principles of the first-order low-pass filter. After filtering and denoising, linear interpolation smoothing is performed on the processed load sequence and temperature sequence. This process inserts new data points between two adjacent original data points at a fixed time interval (e.g., half of the original sampling interval). The value of each newly inserted data point is calculated from the values ​​of the two original data points before and after it through a linear relationship. After this process, a filtered load sequence and a clean temperature sequence are obtained.

[0051] Subsequently, using the clean temperature sequence obtained in the previous step as input, the temperature gradient sequence is calculated using the sliding window method. A fixed-length sliding window (e.g., covering several original sampling points) needs to be set, starting from the beginning of the sequence and sliding forward one data point at a time. For each window position, the difference between the temperature value at the end of the window and the temperature value at the beginning of the window is calculated. This difference is then divided by the time difference between the two times, and the quotient is used as the temperature gradient value at the center of the window. This process is repeated throughout the entire sequence to obtain the temperature gradient sequence. This step requires a preset temperature gradient threshold to determine whether the gradient is significant. This threshold is obtained by collecting multiple historical clean temperature sequences under normal conditions, calculating the temperature gradient value for each sequence segment using the method described above, and taking the absolute value of all calculated temperature gradient values. The 99.5th percentile of this absolute value sequence is then set as the temperature gradient threshold.

[0052] Next, the absolute value of each data point in the calculated temperature gradient sequence is compared with the aforementioned temperature gradient threshold. When the absolute value of the gradient at a data point is greater than the threshold, the corresponding time is recorded. A fixed-duration time interval is extracted before and after this time, centered on the threshold. This fixed duration is based on the thermal inertia time constant of the device, which can be obtained by analyzing the response duration of step changes in historical temperature data (e.g., taking the average of such durations). Simultaneously, the timestamps corresponding to all points in the temperature gradient sequence where the absolute gradient value exceeds the temperature gradient threshold are marked as abrupt change moments. These abrupt change moments then need to be merged to form continuous abrupt change intervals, requiring a preset minimum merging interval parameter. This parameter is determined based on the minimum achievable duration of a step change in device temperature. This can be achieved by analyzing historical temperature data, identifying all obvious temperature step change events, and statistically analyzing the duration of each event from start to end. The minimum of these durations is taken as the minimum merging interval. If the time difference between any two abrupt change moments is less than this minimum merging interval, these two moments and all consecutive moments between them are merged, considered as a continuous abrupt change interval. For each merged mutation interval, its start timestamp, end timestamp, and all corresponding temperature values ​​in the pure temperature sequence within that interval are recorded. This information together constitutes the temperature mutation interval output.

[0053] After performing the above operations on all points where the gradient exceeds the limit, multiple potentially overlapping time periods will be obtained. For each such time period, a subsequence of load data within the exact same time interval is found in the first-filtered load sequence, and this subsequence is then subjected to a second filtering process using the aforementioned first-order low-pass filtering algorithm. This second filtering also uses the first-order low-pass filtering algorithm, but with a smaller filter coefficient than the first filtering (e.g., 0.75 times the first filter coefficient) to achieve a more stringent smoothing effect. The filtered output serves as the clean load data corresponding to that time period. This process is repeated for all time periods where the temperature gradient exceeds the limit, and the processed data is then concatenated chronologically with the data directly taken from the first-filtered load sequence for time periods where the temperature gradient does not exceed the limit, ultimately forming a complete clean load sequence.

[0054] In step S13, based on the pure load sequence and the temperature abrupt change interval, preliminary loss features are extracted to obtain load frequency features and operating segment sequences. A loss feature vector is then constructed based on the load frequency features, the temperature abrupt change interval, and the operating segment sequences, including:

[0055] Based on the pure load sequence, time-frequency decomposition is performed using short-time Fourier transform to obtain the load frequency characteristics;

[0056] Obtain continuous operation segment records of the electromechanical equipment, and divide the continuous operation segment records into intervals according to the temperature change intervals to obtain an operation segment sequence containing stable segments and change intervals;

[0057] Based on the load frequency characteristics, find the frequency corresponding to the peak value to obtain the main frequency, and calculate the energy percentage of the main frequency to obtain the main frequency energy percentage;

[0058] Based on the temperature abrupt change range, the mean abrupt change amplitude and the standard deviation of the abrupt change amplitude are calculated. Based on the running segment sequence, the average stable running time of the stable segment is calculated by weighted average method. The frequency of the abrupt change segment within a preset period is statistically analyzed to obtain the abrupt change frequency.

[0059] The main frequency energy ratio, the mean of the mutation amplitude, the standard deviation of the mutation amplitude, the mutation frequency, and the average stable running time are used as multi-dimensional feature vectors. Based on the multi-dimensional feature vectors, the features are standardized using the Z-Score standardization method to obtain the loss feature vector.

[0060] Specifically, a short-time Fourier transform is first performed on the clean load sequence to obtain the load frequency characteristics. The specific implementation process is as follows: the clean load sequence is divided into multiple consecutive time frames in chronological order. The duration of each frame is determined by analyzing historical load sequences, identifying their periodic fluctuation patterns, and statistically calculating the period length. The median of all statistically obtained period lengths is taken as the duration of each frame. An overlap is set between adjacent time frames, with the overlap length set to half the frame duration. A Hanning window function is applied to the load data points within each frame. The coefficients of this window function are calculated based on the relative position of each data point within the frame's time range, using a cosine function to transform this relative position. A fast Fourier transform algorithm is applied to each frame of data after windowing. This algorithm converts the time-domain data into frequency-domain data, obtaining a sequence of spectral coefficients in complex form. In the spectrum of each frame, the amplitude values ​​corresponding to all frequency points are traversed, and the frequency point with the largest amplitude value is found and recorded as the dominant frequency of that frame. Next, the square of the amplitude corresponding to the dominant frequency point is calculated, and then the sum of the squares of the amplitudes corresponding to all frequency points in the frame is calculated. The former is divided by the latter, and the quotient is recorded as the dominant frequency energy proportion of the frame. A dominant frequency energy proportion threshold is preset for subsequent feature construction. This threshold is determined by analyzing historical normal operation data. Specifically, multiple historical clean load sequences under normal equipment conditions are collected, and the above short-time Fourier transform process is performed on each sequence to obtain the dominant frequency energy proportion values ​​of all frames. The 50th percentile of these proportion values ​​is calculated and set as the dominant frequency energy proportion threshold. Finally, the time center point of each frame, the dominant frequency value corresponding to the frame, and the dominant frequency energy proportion value of the frame are combined into a data unit. The data units of all frames are arranged in chronological order to form the load frequency feature output.

[0061] Simultaneously, continuous operation segment records of electromechanical equipment are retrieved from the operation log database of the equipment monitoring system. These records exist as a sequence of Boolean values, where each data point corresponds to a timestamp. A true value indicates the equipment is running at that moment, while a false value indicates it is stopped. The operation record is segmented based on identified temperature abrupt change intervals. Each temperature abrupt change interval includes a series of start and end times, defining the time period in which a significant temperature change occurs. The time range of the operation record is traversed, and time periods that are completely outside any temperature abrupt change interval and where the equipment is continuously running for a duration exceeding a preset minimum stable segment duration are marked as stable segments. The minimum stable segment duration is determined by statistically analyzing all historical intervals from shutdown to startup or from operation to shutdown, and taking the tenths of these interval values ​​as the minimum stable segment duration. The portion of the operation record that overlaps temporally with any temperature abrupt change interval is marked as an abrupt change segment. All identified stable and abrupt change segments are arranged in chronological order of their start times to form an operation segment sequence. Each segment in the sequence records its start and end timestamps, as well as its segment type.

[0062] A loss feature vector is constructed based on load frequency characteristics, temperature abrupt change intervals, and operating segment sequences. From the load frequency characteristics, a complete analysis period is defined as one-quarter of the typical maintenance cycle of the equipment, determined by the equipment maintenance manual or historical average maintenance intervals. Within this analysis period, the main frequency energy percentage of all frames is extracted, and the arithmetic mean of these values ​​is calculated and recorded as the average main frequency energy percentage. From the temperature abrupt change intervals, for each interval, the absolute value of the difference between the start and end times of the interval in the pure temperature sequence is calculated to obtain the temperature abrupt change amplitude for each interval. The arithmetic mean of the abrupt change amplitudes of all intervals is calculated to obtain the average abrupt change amplitude, and the standard deviation of these abrupt change amplitudes is also calculated. From the operating segment sequence, all records of type stable segment are extracted, the duration of each stable segment is calculated, and the arithmetic mean of these durations is calculated to obtain the average stable operating time. The number of times records of type abrupt change segment appear within the same analysis period is counted, and this number is divided by the length of the analysis period to obtain the abrupt change frequency. The five calculated values—mean of main frequency energy percentage, mean of mutation amplitude, standard deviation of mutation amplitude, average stable running time, and mutation frequency—are arranged in this order to form an initial multidimensional feature vector.

[0063] The initial multidimensional feature vector is Z-score standardized. This process requires a historical feature dataset, which is constructed by collecting historical clean load sequences, historical temperature abrupt change intervals, and historical operating segment records from multiple analysis periods of the device, and calculating the initial multidimensional feature vector for each period using the same process described above. The arithmetic mean and standard deviation of all sample values ​​for each feature item (i.e., mean of main frequency energy percentage, mean of abrupt change amplitude, standard deviation of abrupt change amplitude, average stable operating time, and abrupt change frequency) in this historical dataset are calculated. For the current initial multidimensional feature vector to be processed, each value is subtracted from the historical arithmetic mean of the corresponding feature item, and the difference is divided by the historical standard deviation of that feature item. The resulting standardized feature values ​​are then used to construct the final loss feature vector in their original order.

[0064] In step S14, a generalized cross-correlation analysis is performed on the pure load sequence and the pure temperature sequence to obtain a synchronization offset sequence. Then, the internal loss rate is quantified based on the synchronization offset sequence to obtain a loss rate index, including:

[0065] The pure load sequence and the pure temperature sequence are divided into multiple overlapping windows through a preset sliding window, and the cross power spectrum is calculated by fast Fourier transform.

[0066] Based on the cross-power spectrum, whitening is performed using a preset phase transformation weighting function, and a generalized cross-correlation function sequence with time shift as the independent variable and cross-correlation value as the dependent variable is obtained through inverse Fourier transform.

[0067] The peak value of the generalized cross-correlation function sequence is extracted as the optimal time shift, and the optimal time shift corresponding to all the overlapping windows is combined to obtain the synchronization offset sequence.

[0068] The time period corresponding to the synchronization offset sequence exceeding the preset offset threshold is marked as an abnormal time period to obtain the preliminary abnormal time period;

[0069] Calculate the cumulative duration of the initial abnormal time period, divide the cumulative duration by the continuous operation segment record, and obtain the loss rate index.

[0070] Specifically, a fixed-time sliding window needs to be set for analysis. The window length is determined by analyzing historical clean load sequences, identifying recurring load fluctuation patterns with significant amplitudes, measuring the period length of these patterns, and taking the arithmetic mean of all measured period lengths. Twice this average is set as the sliding window length. An overlapping portion is set between adjacent analysis windows, with the overlap length set to three-quarters of the aforementioned window length. Based on this window division rule, the clean load sequence and clean temperature sequence are divided into a series of time-aligned data segments, forming multiple pairs of window data. Each pair of data contains load and temperature segments within the same time period.

[0071] For each pair of data windows, the spectra of the two data segments need to be calculated separately. The Fast Fourier Transform (FFT) algorithm is applied to process the load and temperature segments separately. This algorithm converts the time-domain data into a complex representation in the frequency domain, obtaining the load spectrum and temperature spectrum. Next, the conjugate of the complex number corresponding to each frequency point in the load spectrum is calculated, and then multiplied point-by-point with the complex number at the same frequency point in the temperature spectrum to obtain the cross-power spectrum of the window. Subsequently, the cross-power spectrum is whitened to enhance anti-interference capability using a phase transform weighting function. This function is constructed by first calculating the amplitude (i.e., the modulus of the complex number) corresponding to each frequency point in the cross-power spectrum, then taking the reciprocal of these amplitudes, and finally exponentiating each reciprocal. The exponent is calibrated to negative one based on historical data analysis. Specifically, the calibration method involves collecting the cross-power spectrum of multiple data segments under known good conditions and analyzing their amplitude statistical distribution to determine the exponent value that optimizes noise suppression. After whitening, the cross-power spectrum is transformed back to the time domain using the inverse fast Fourier transform algorithm, resulting in a function sequence with time shift as the independent variable and cross-correlation value as the dependent variable, called the generalized cross-correlation function sequence. Within this sequence, the point with the largest cross-correlation value is searched and located; the time shift corresponding to this point is considered the optimal time shift for that analysis window. This process is repeated for all window pairs, calculating the optimal time shift for each window. These optimal time shifts are then arranged according to the time order of the center points of each window, forming a synchronization offset sequence.

[0072] An offset threshold is preset to identify anomalies from the synchronization offset sequence. The threshold is determined by collecting multiple historical clean load and temperature sequences generated by the equipment during a recognized fault-free operation phase, calculating the corresponding historical synchronization offset sequence according to the generalized cross-correlation analysis method described above, calculating the absolute value of each data point in the historical sequence to obtain an absolute value sequence, and taking the 95th percentile value of the absolute value sequence as the offset threshold.

[0073] In the currently calculated synchronization offset sequence, the absolute value of each data point is compared with the aforementioned offset threshold. If the absolute value of a point is greater than the threshold, the corresponding time point is marked as an outlier. Subsequently, outliers that occur consecutively in time are merged, and each consecutive time period formed after merging is recorded as a preliminary outlier time period. The start and end times of each preliminary outlier time period must be recorded.

[0074] From the continuous operation segment records obtained in step S13 (which clearly define all time periods during which the equipment is in operation), the total duration of the equipment's operation is calculated. Next, the sum of the durations of all initial abnormal time periods is calculated to obtain the cumulative duration. Finally, the cumulative duration is divided by the total equipment operation time, and the resulting quotient is defined as the loss rate index.

[0075] In step S15, the loss coefficient prediction model construction process includes:

[0076] Obtain historical loss feature vectors, historical loss rate indices, and historical loss coefficients;

[0077] The historical loss feature vector and historical loss rate index are input into the input layer of the initially constructed neural network model for training, and the predicted loss coefficients output by the output layer of the neural network model are obtained.

[0078] Substitute the predicted loss coefficient and the historical loss coefficient into the loss function to calculate the loss value;

[0079] The gradient of the output layer of the neural network model is calculated based on the loss value, and the gradient is passed forward layer by layer through the chain rule to calculate the gradient of the parameters of each layer and obtain the gradient data.

[0080] Based on the gradient data and the preset learning rate, the parameters of each layer of the neural network model are updated using the gradient descent method.

[0081] The parameters of each layer are iteratively updated until the number of training iterations of the neural network model is greater than the preset number of iterations, or the loss value of the neural network model is less than the preset loss threshold. At this point, the training is considered complete, and a loss coefficient prediction model that has been trained is obtained.

[0082] Specifically, a loss coefficient prediction model is constructed based on a historical dataset. This dataset contains three parts: historical loss feature vectors, historical loss rate indices, and corresponding historical loss coefficients. The historical loss feature vectors are obtained by processing historically collected and stored equipment pure load sequences and historical temperature abrupt change intervals strictly according to the procedure described in step S13. The historical loss rate indices are calculated by processing historical pure load sequences and historical pure temperature sequences from the same time period strictly according to the procedure described in step S14. The historical loss coefficients originate from physical inspection records during past equipment downtime maintenance. Technicians use measuring tools (such as calipers to measure bearing clearance or thickness gauges to measure seal wear) to actually measure key components. The measured physical wear is mapped to a scalar value between zero and one through a preset linear normalization function. This scalar value is the historical loss coefficient, which is dimensionless and characterizes the actual wear degree of the component at the corresponding historical moment.

[0083] In this invention, the wear coefficient is a dimensionless comprehensive index with a value range of [0, 1], used to characterize the overall wear degree of the equipment at the current moment. The measured wear coefficient, i.e., the historical wear coefficient, is mapped to the [0,1] interval through a linear normalization function. For each key component, a maximum allowable threshold for its wear is preset, such as the design limit value of bearing clearance. The actual wear is measured during maintenance, and the historical wear coefficient is the minimum value between the actual wear divided by the maximum threshold and 1. For equipment containing multiple components, the wear coefficients of each component can be weighted and averaged, or a representative component can be designated as the basis for the overall machine wear coefficient. In this embodiment, the wear of the core component of the equipment, such as the bearing, is preferably used as the representative of the overall machine wear coefficient. The predicted wear coefficient is output in real time by the wear coefficient prediction model when the equipment is running online. The model takes the real-time collected wear feature vector and wear rate index as input, and outputs it after neural network calculation, which is used to characterize the assessment of the wear state of the equipment at the current moment.

[0084] This loss coefficient prediction model employs a three-layer feedforward fully connected neural network structure. The input layer dimension is the loss feature vector dimension plus 1 (loss rate index), the number of neurons in the hidden layer is 150% of the input layer dimension, the activation function is ReLU, and the output layer consists of linear neurons. The connection weights between all neurons in the network model are initialized before training begins using a truncated normal distribution random number generator. The random numbers generated by this generator have a mean of zero and a standard deviation of 0.01. The bias parameters of all neurons are initialized to zero.

[0085] The model training process uses a batch gradient descent strategy. In each iteration, a batch of training samples is randomly selected from the historical dataset. Each sample contains a historical loss feature vector and a historical loss rate index. These two data points are concatenated to form a one-dimensional array, which serves as the input to the neural network. The data propagates layer by layer through the network. In each layer, the output value of each neuron is equal to the sum of all input values ​​received by that neuron multiplied by their corresponding connection weights, plus the neuron's bias value. This result is then input into the neuron's activation function to calculate the final output. After propagation through the input and hidden layers, the predicted loss coefficient for that sample is finally obtained at the output layer. For the entire batch of training samples, the loss function is defined as mean squared error, and the result is the loss value under the current model parameters.

[0086] Next, backpropagation is performed to optimize the model parameters. Using automatic differentiation, starting from the end of the computation graph (i.e., the loss value), the partial derivatives of the loss value with respect to the output layer parameters (weights and biases) of the neural network are calculated according to the chain rule; this is the gradient. Then, this gradient information is propagated backward along the network structure, calculating the gradients of the loss value with respect to the hidden layer parameters and the loss value with respect to the input layer parameters, ultimately obtaining the gradient data for each trainable parameter in the model.

[0087] Parameter updates employ the Adam optimization algorithm, with the learning rate determined through grid search. Training termination conditions are reached when the preset maximum number of iterations is reached (twice the average number of historical convergent iterations) or when the loss falls below a preset threshold (the 20th percentile of historical final loss values). After training, the weight and bias parameters of all layers in the neural network model are saved. This set of parameters constitutes the final trained loss coefficient prediction model. This model can accept a new loss feature vector and loss rate index, calculated in real-time in steps S13 and S14, as input and output a predicted loss coefficient value.

[0088] In step S16, a performance deviation vector comprising load loss deviation, temperature response deviation, and duration accumulation deviation is calculated based on the loss coefficient. The duration accumulation deviation is then periodically corrected based on the load loss deviation and the temperature response deviation to obtain the final performance deviation vector, including:

[0089] Calculate the difference between the loss coefficient and the preset loss coefficient benchmark value to obtain a performance deviation vector that includes load loss deviation, temperature response deviation and time accumulation deviation;

[0090] Based on the load loss deviation and the temperature response deviation, the time point of the peak occurrence is identified, and the average time difference between adjacent peaks is calculated to obtain the load loss cycle and the temperature response cycle.

[0091] Based on the load loss deviation and the temperature response deviation, the cross-correlation function value is calculated using the cross-correlation function formula as the delay time. Based on the delay time, the load loss period, and the temperature response period, the characteristic period value is calculated using the preset coupling period formula.

[0092] The cumulative time deviation is segmented using the characteristic periodic value as the period, and the first-order forward difference value of two adjacent segments is calculated to obtain the cumulative rate sequence.

[0093] When the characteristic period value exceeds the preset period threshold, the duration accumulation deviation is corrected by linear weighted summation according to the cumulative rate sequence, and the performance deviation vector is updated to obtain the final performance deviation vector. When the characteristic period value does not exceed the preset period threshold, the performance deviation vector is output as the final performance deviation vector.

[0094] Specifically, three preset benchmark values ​​need to be determined first: load loss benchmark value, temperature response benchmark value, and cumulative duration benchmark value. The load loss benchmark value is obtained by extracting all historical loss coefficients recorded in the equipment's historical operation database during the first typical maintenance cycle after commissioning, under stable load conditions, and calculating the arithmetic mean of these historical loss coefficients. The temperature response benchmark value is obtained by extracting all historical loss coefficients recorded in the same historical database during the first typical maintenance cycle, under stable temperature conditions, and calculating the arithmetic mean of these historical loss coefficients. The cumulative duration benchmark value is obtained by extracting all historical loss coefficients recorded in the same historical database during the initial short period after the equipment is put into operation in a completely new state, and calculating the arithmetic mean of these historical loss coefficients.

[0095] Next, the current input loss coefficient is calculated and then subtracted from the three reference values ​​to obtain the load loss deviation, temperature response deviation, and time cumulative deviation. These three scalar values ​​form a vector, which is called the initial performance deviation vector.

[0096] In subsequent processing, load loss deviation and temperature response deviation are treated as two independent time series (although the initial values ​​are scalars, they form a sequence over consecutive analysis periods). To extract periodic features from these two series, peak points in each series need to be identified. A sliding window method is used for identification. The length of the sliding window is set based on the sampling frequency of the original data and the minimum action interval time that the device can recognize. Specifically, the number of data points covered by the window is equal to the sampling frequency multiplied by the minimum action interval time. For each data point in the sequence, a sliding window is constructed centered on that point. If the value of the center point is simultaneously greater than the values ​​of the two data points before and after it, the center point is marked as a peak point. In the load loss deviation sequence, all marked peak points are identified, and the arithmetic mean of the time interval between every two adjacent peak points is calculated. This average is defined as the load loss period. The same peak identification and adjacent peak time interval calculation process is performed on the temperature response deviation sequence, and the resulting arithmetic mean is defined as the temperature response period.

[0097] To quantify the time lag between the effects of load and temperature, it is necessary to calculate the delay time between the load loss deviation sequence and the temperature response deviation sequence. The delay time is determined by first defining a search range for the maximum time shift, typically covering the time length of plus or minus one load loss cycle. Then, for each candidate time shift within the search range, the load loss deviation sequence is shifted along the time axis by that shift, and the arithmetic mean of the products of the shifted sequence and the temperature response deviation sequence over the overlapping time region is calculated. This process is repeated for all candidate time shifts, and the one that maximizes this arithmetic mean is defined as the delay time.

[0098] Next, a characteristic value representing the period of load-temperature coupling needs to be calculated, called the characteristic period value. In this embodiment, to simplify the calculation and ensure coverage of the longest period of load-temperature coupling, the characteristic period value is taken as the larger of the load loss period and the temperature response period. This value ensures that when the cumulative time deviation is subsequently corrected in segments, the segment length is not less than any original period, thereby effectively capturing the macroscopic impact of the coupling effect. In another embodiment, the two can also be weighted and averaged according to the delay time, but this embodiment prefers to use the maximum value to balance conservatism and calculation simplicity.

[0099] Then, using the calculated characteristic period value as a fixed segment length, the cumulative time deviation sequence (composed of scalar values ​​of cumulative time deviation at consecutive time points) is divided at equal intervals. The division begins at the start of the sequence, and the length of each segment is equal to the characteristic period value. If the remaining portion at the end of the sequence is less than the length of a complete characteristic period value, it is discarded. For each segment, the arithmetic mean of all cumulative time deviation data within that segment is calculated as the representative value for that segment. Next, following the segmentation order, the difference between each segment's representative value and the representative value of the previous segment is calculated, i.e., the subsequent segment's representative value minus the previous segment's representative value. This yields a difference sequence, called the cumulative rate sequence.

[0100] This step requires a preset periodic threshold to determine whether the periodic coupling effect is significant. This threshold is determined by collecting characteristic periodic values ​​calculated over multiple complete historical analysis periods, forming a historical characteristic periodic value sequence. The 75th percentile of this sequence is then calculated and set as the periodic threshold. The currently calculated characteristic periodic value is compared with this periodic threshold. If the current characteristic periodic value is greater than the periodic threshold, a significant periodic coupling effect is considered to exist, and the cumulative time bias needs to be corrected.

[0101] The correction process employs a linear weighted sum method. First, the standard deviation of the accumulated rate sequence obtained above is calculated. Next, a weight sequence with the same length as the accumulated rate sequence is constructed. The weight value at each position in this weight sequence is equal to the previously calculated standard deviation divided by the absolute value of the difference at the corresponding position in the accumulated rate sequence. Then, all weights are normalized so that the sum of all weights is one. Next, this weight sequence is used to perform a weighted summation operation on the original duration accumulated deviation sequence. Specifically, the product of each weight in the weight sequence and the corresponding data point in the duration accumulated deviation sequence is calculated, and all product results are summed to obtain a corrected scalar value. This corrected scalar value replaces the original duration accumulated deviation value. Finally, the corrected duration accumulated deviation value is combined with the uncorrected original load loss deviation value and temperature response deviation value to form a new vector, which is the final performance deviation vector. If the currently calculated characteristic period value is not greater than (less than or equal to) a preset period threshold, the above correction operation is not performed, and the initial performance deviation vector is directly output as the final performance deviation vector.

[0102] In step S17, based on the final performance deviation vector, evolution pattern clustering is performed to obtain loss evolution trend clusters. Then, based on these loss evolution trend clusters, accelerated loss analysis and secondary clustering are performed to obtain refined trend sub-clusters, including:

[0103] Based on the final performance deviation vector, the evolution pattern is clustered and grouped using the DBSCAN clustering algorithm to obtain loss evolution trend clusters containing load components, temperature components, and duration components.

[0104] Based on the loss evolution trend cluster, the variance contribution rates of the load component, the temperature component, and the duration component are calculated to obtain the contribution rates of the load component, the temperature component, and the duration component.

[0105] The contribution rates of the load component and the temperature component are compared with preset load contribution rate thresholds and preset temperature contribution rate thresholds, respectively, and the time periods when both exceed the thresholds are extracted to obtain the contribution overlap period.

[0106] Based on the overlapping contribution period, the rising slope of the contribution rate of the duration component is calculated, and the overlapping contribution period corresponding to the rising slope exceeding the preset slope threshold is marked as the loss acceleration period, thus obtaining the loss acceleration period.

[0107] Based on the loss acceleration period and the loss evolution trend cluster, a secondary clustering of the acceleration stage is performed using a hierarchical clustering algorithm to obtain refined trend sub-clusters.

[0108] Specifically, the three components of the final performance deviation vector at each time point are used as three-dimensional spatial points. DBSCAN clustering is employed, with the neighborhood radius set at the 15th percentile of the historical Euclidean distance and the minimum number of points set at 5% of the total data points. During clustering, each data point is traversed, and the number of other data points contained within a spatial range centered on that point and with a neighborhood radius threshold as the radius is counted. If this number is greater than or equal to the preset minimum number of points threshold, the point is marked as a core point. All mutually reachable core points and other points within their neighborhoods (including other core points and boundary points within the neighborhood of a core point but not meeting the core point criteria) are merged into the same cluster. Any point not belonging to the neighborhood of any core point is marked as a noise point and excluded. Each identified cluster contains data points corresponding to one or more consecutive time periods on the original time axis. All time periods belonging to the same cluster are merged, and the load component sequence value, temperature component sequence value, and duration component sequence value are extracted within these merged time periods. The set of these three sequence values ​​is defined as a loss evolution trend cluster.

[0109] For each loss evolution trend cluster obtained above, it is necessary to calculate the variance contribution rate of its three internal components, that is, the proportion of component variance to total variance.

[0110] This step pre-sets two thresholds for filtering key time periods: the load contribution rate threshold and the temperature contribution rate threshold. The load contribution rate threshold is determined by collecting the load component contribution rate values ​​of all loss evolution trend clusters generated in historical data analysis and taking the 75th percentile value as the load contribution rate threshold. The temperature contribution rate threshold is determined in the same way, taking the 75th percentile value as the temperature contribution rate threshold. For each original time period corresponding to a loss evolution trend cluster, it is checked whether its calculated load component contribution rate exceeds the load contribution rate threshold, and simultaneously, whether its temperature component contribution rate exceeds the temperature contribution rate threshold. If both conditions are met, the original time period corresponding to that loss evolution trend cluster is extracted; these time periods are called contribution overlap periods.

[0111] Within each overlapping contribution period, it is necessary to analyze the trend of the contribution rate of the duration component over time. Time is used as the independent variable, and the contribution rates of the duration component arranged chronologically within that period are used as the dependent variable. A straight line is fitted using the least squares method. The purpose of this method is to find a straight line that minimizes the sum of the squares of the vertical distances from all contribution rate values ​​to this line. The slope of this fitted line is calculated, and this slope value is defined as the rising slope of the duration component contribution rate within the overlapping contribution period. A slope threshold is preset to determine whether the upward trend is significant. This threshold is determined by collecting the rising slope values ​​calculated for all overlapping contribution periods in historical data analysis, taking their absolute values ​​to form a historical absolute rising slope value sequence, calculating the 90th percentile value of this sequence, and setting this value as the slope threshold. The absolute value of the rising slope calculated for the current overlapping contribution period is compared with the slope threshold. If the current absolute value is greater than the slope threshold, the current overlapping contribution period is marked as a period of accelerated loss.

[0112] Based on all the loss acceleration periods identified in the above process, more refined clustering analysis is needed within these periods. Specifically, from the final performance deviation vector, all data points falling within these loss acceleration periods are extracted (each point contains three values: load component, temperature component, and duration component). This set of data points constitutes a new dataset. A hierarchical clustering algorithm is then used for secondary clustering on this dataset. First, the Euclidean distance between any two data points in the new dataset is calculated, forming a symmetric distance matrix. The clustering process uses the WARD method as the inter-cluster connectivity criterion. This method calculates the increment of the sum of intra-cluster variances caused by all possible merging methods at each step of cluster merging and selects the two clusters that minimize this increment for merging. By observing the abrupt change points (i.e., elbow points) in the rate of change of inter-cluster distance as the number of clusters decreases during the cluster merging process, the optimal number of clusters is determined. Based on the determined number of clusters, the hierarchical clustering dendrogram is cut, dividing the dataset into a specified number of sub-clusters, each of which is a refined trend sub-cluster.

[0113] In step S18, based on the refined trend sub-cluster, the pure load sequence, and the pure temperature sequence, the initial wear sequence, the load impact component, and the temperature response component are separated. The influence of the load impact component on the initial wear sequence is filtered out to obtain the impact-filtered wear sequence, including:

[0114] Extract the pure load sequence and the pure temperature sequence corresponding to the refined trend sub-cluster, and separate the load impact component representing short-term sudden changes, the temperature response component representing periodic changes, and the baseline change component representing long-term slow changes using the STL decomposition method. Use the baseline change component as the initial wear sequence.

[0115] Based on the load impact component, identify the peak value and record the time point, and extract the local amplitude of the initial wear sequence based on the time point;

[0116] The time points corresponding to the local amplitude exceeding the preset amplitude threshold are marked as load impact induced deviation times, thus obtaining a set of induced deviation times;

[0117] Based on the set of induced deviation times, the abnormal fluctuations of load impact-induced deviations in the initial wear sequence are removed by a local weighted regression algorithm to obtain the impact-filtered wear sequence.

[0118] Specifically, firstly, based on the specific start and end time points defined for each refined trend sub-cluster, data segments within the corresponding time intervals are extracted from the global pure load sequence and pure temperature sequence, which are respectively called sub-cluster load sequence segments and sub-cluster temperature sequence segments.

[0119] A seasonal trend decomposition method was applied to the sub-cluster load sequence fragments, decomposing them into three components. This decomposition was achieved through an iterative weighted local regression process. First, two key periodic parameters were set: the seasonal period length and the trend smoothing window length. The seasonal period length was set based on the periodic variation of temperature, determined by performing spectral analysis on historical pure temperature sequences to identify the frequency component with the highest energy; the reciprocal of this reciprocal is the seasonal period length. The trend smoothing window length was set based on the slowness of the equipment wear trend change, determined by analyzing historical wear data and calculating the time lag corresponding to the autocorrelation function decaying to half its initial value; this time lag value was set as the trend smoothing window length.

[0120] Before the decomposition process begins, the temperature response component is initialized as a zero-value sequence of the same length as the sub-cluster load sequence segment, and the baseline change component is initialized as a zero-value sequence of the same length. Then, an outer loop is entered, with a fixed number of iterations (two). Each outer loop contains an inner loop, with a fixed number of iterations (one). In the inner loop, the detrended sequence is first calculated, i.e., the original sub-cluster load sequence segment is subtracted from the current baseline change component. Next, periodic subsequence smoothing is performed on the detrended sequence to update the temperature response component. Specifically, the detrended sequence is divided into multiple consecutive subsequences according to the seasonal cycle length, with each subsequence containing data points at the same phase position. Locally weighted regression smoothing is performed independently on each phase subsequence. The local weighted regression uses a quadratic polynomial as the local model, and the weight function uses a cubic decay function, meaning that the farther a data point is from the regression center point, the smaller its weight; the weight value is inversely proportional to the cube of the distance. After smoothing, the smoothed values ​​of each phase subsequence are obtained. These smoothed values ​​are then recombined into a complete sequence according to the original phase order. This complete sequence is then smoothed by a moving average with a sliding window length equal to the seasonal cycle length, finally yielding the updated temperature response components.

[0121] Then, the deseasonal sequence is calculated, which is the original subcluster load sequence fragment minus the updated temperature response component. The deseasonal sequence is then trend-smoothed to update the baseline change component. Specifically, locally weighted regression is used to globally smooth the deseasonal sequence. The window length of the locally weighted regression is equal to the pre-set trend-smoothing window length. A linear polynomial is used as the local model, and the weighting function also uses a cubic decay function. After smoothing, the updated baseline change component is obtained.

[0122] After the inner loop completes, the load impact component is calculated, which is the original sub-cluster load sequence fragment minus the current temperature response component and baseline change component. Then, the residual sequence is calculated, which is the difference between the original sub-cluster load sequence fragment and the sum of the current three components. Based on the residual sequence, a robustness weight is calculated for each data point, used in the next outer loop's locally weighted regression. The robustness weight is calculated by first calculating the absolute value of the median of the residual sequence, then dividing the absolute value of each residual by the median to obtain a ratio, and finally mapping this ratio to a weight value between zero and one using a bisquare function; data points with larger absolute residual values ​​have smaller weight values. The specific mapping rule for the bisquare function is as follows: a preset adjustment constant is set, which is a fixed value greater than zero, determined according to the standard method of robust statistics. During calculation, the ratio is first divided by the adjustment constant to obtain a standardized residual. If the absolute value of this standardized residual is less than or equal to one, its corresponding weight is equal to one minus the square of the absolute value of the standardized residual, and then the difference is squared. If the absolute value of this standardized residual is greater than one, its corresponding weight is directly set to zero. The calculated robustness weight is multiplied by the distance weight in the next local weighted regression to obtain the new composite weight. After two outer loops, the final load impact component, temperature response component, and baseline change component are obtained. The baseline change component is used as the initial wear sequence reflecting the wear trend of this sub-cluster.

[0123] To identify the anomalous impact of load shocks on the initial wear sequence, it is necessary to locate peak points within the load shock components. Specifically, this involves calculating the first-order difference sequence of the load shock component sequence, i.e., the sequence of each value minus the previous value. Simultaneously, an impact threshold is set. This threshold is determined by collecting multiple load shock components from historical data analysis, identifying all local peak points, and calculating their absolute values ​​to form a set of historical impact peak absolute values. The 85th percentile of this set is then calculated and set as the impact threshold. In the load shock component sequence, if a data point's value is greater than the values ​​of its two preceding and two following data points, and also exceeds the preset impact threshold, this point is recorded as a peak point. The specific time points corresponding to all peak points are recorded. Based on these time points, identical moments are found in the initial wear sequence, and the value at that moment is read. Using that moment as the center, five data points before and after it are taken to form a neighborhood, and the arithmetic mean of all values ​​within this neighborhood is calculated. Calculate the absolute value of the difference between the value at that moment and the arithmetic mean, and define this absolute value as the local amplitude at that moment. A preset amplitude threshold is used to determine whether the local deviation is significant. This threshold is determined by analyzing all calculated local amplitudes from historical data. Calculate the 90th percentile of this set of local amplitudes and set this value as the amplitude threshold. Mark the corresponding time points where all local amplitudes are greater than the amplitude threshold; the set of these marked time points is called the induced deviation time set.

[0124] Based on the induced deviation time set, a local weighted regression smoothing process is performed on the initial wear sequence to eliminate abnormal fluctuations induced by load impact. For each time point in the initial wear sequence, a smoothing window is dynamically determined with that point as the center. The basic principle for the window size is to ensure that the window contains at least ten valid data points. For each data point within the window, the weight of that point in the regression with respect to the center point is calculated based on the difference between its timestamp and the timestamp of the center point.

[0125] The specific process of weight calculation is as follows: For each data point within the smoothing window, calculate the absolute value of the difference between its timestamp and the timestamp of the window center point to obtain the absolute value of the time difference for that point; next, find the maximum value of the absolute value of the time difference for all data points within the smoothing window, called the maximum time difference of the window; then, divide the absolute value of the time difference for each data point by the maximum time difference of the window to obtain a standardized distance value between zero and one; then subtract the standardized distance value for that point from one to obtain an intermediate value; perform a cube operation on this intermediate value, that is, calculate the result of multiplying the intermediate value by itself three times; the final result is the unnormalized original weight value for that data point; after completing the above calculations for all data points within the window, sum the original weight values ​​corresponding to all data points to obtain the total weight value; finally, divide the original weight value of each data point by the total weight value to obtain its final normalized weight value.

[0126] Within each window, a local polynomial curve (usually a first or second-order polynomial) is fitted using weighted least squares. The corresponding value of the center point on the fitted curve is then used as the value of the smoothed sequence at that point. After traversing all points of the initial wear sequence, a new sequence is obtained, called the impact filtration wear sequence.

[0127] In step S19, based on the impact-filtered wear sequence and the temperature response component, the influence of temperature cycle changes is filtered out to obtain a pure wear evolution sequence. Based on this pure wear evolution sequence, lifetime prediction is performed to obtain the remaining lifetime curve, including:

[0128] Based on the temperature response components, the time points of the peaks and troughs are extracted to obtain the temperature peak and trough time set;

[0129] Based on the impact filter wear sequence and the temperature peak and valley time set, the amplitude difference between the peak time and the valley time is calculated to obtain the temperature influence sequence;

[0130] Based on the temperature influence sequence, the impact filtering wear sequence is corrected for temperature influence using a pre-built residual correction model to obtain a pure wear evolution sequence;

[0131] Based on the pure wear evolution sequence, the remaining life curve is obtained by predicting the remaining life using a pre-built life prediction model.

[0132] Specifically, to quantify the impact of periodic temperature changes, it is necessary to analyze the temperature response components to extract their peak and trough times. First, the first-order difference sequence of the temperature response component sequence is calculated. The zero-crossing points where the values ​​in this difference sequence change from positive to negative are identified. These points correspond to the locations where the first derivative of the original sequence is zero and the second derivative is negative; these are the peak points, and their times are recorded. Simultaneously, the zero-crossing points where the values ​​in the first-order difference sequence change from negative to positive are identified. These points correspond to the locations where the first derivative of the original sequence is zero and the second derivative is positive; these are the trough points, and their times are recorded. All peak and trough times are combined to form a set called the temperature peak-trough time set.

[0133] By combining the impact filtration wear sequence with the temperature peak-valley time set, the influence of periodic temperature fluctuations on wear observations is calculated. Specifically, in the temperature peak-valley time set, a peak time is selected sequentially in time, and the first valley time immediately following that peak is found. In the impact filtration wear sequence, the corresponding values ​​for these two times are read, and the value of the peak time is subtracted from the value of the valley time to obtain a difference. This calculation is repeated for all sequentially matched peak-valley pairs, and the resulting differences are arranged in chronological order into a new sequence, called the temperature influence sequence.

[0134] To eliminate the influence of temperature periodicity, a pre-built residual correction model is needed to correct the impact filter wear sequence. This residual correction model is a multiple linear regression model. Its construction process is as follows: Impact filter wear sequence fragments and their corresponding temperature influence sequence fragments from multiple similar devices during historical stable wear phases (i.e., phases without severe failures or accelerated wear) are collected as training data. For each set of training data, the temperature influence sequence is used as the first independent variable, the value of each value in the temperature influence sequence corresponding to the previous moment (i.e., its first-order lag sequence) is used as the second independent variable, and the value of the impact filter wear sequence is used as the dependent variable. The least squares method is used to fit all training data, solving for a set of regression coefficients, including the intercept term, the coefficient of the temperature influence term, and the coefficient of its first-order lag term, thus completing the model construction. When applying this model, the currently analyzed temperature influence sequence and its first-order lag sequence are input into the model. The model calculates a predicted value sequence based on the stored regression coefficients, which represents the wear component estimated by the periodic temperature changes. Then, each value in the impact-filtered wear sequence is subtracted from the corresponding value predicted by the model, and the resulting difference sequence is the pure wear evolution sequence.

[0135] It should be noted that when constructing the residual correction model, the impact filter wear sequence is used as the dependent variable, and the temperature influence sequence at the current and historical moments is used as the independent variable. The purpose is to establish a local linear relationship describing how temperature periodic fluctuations affect wear observations. This model focuses on the perturbation pattern of temperature changes on wear values ​​within a local time window. By introducing a first-order lag term of the temperature influence sequence, the model can capture the short-term memory effect of temperature influence, such as the temperature effect delay caused by the heat capacity of the equipment. In practical applications, the model will be trained and applied within a sliding time window, for example, covering 2-3 temperature cycles, to ensure that the model's local linearity assumption holds. By subtracting the temperature perturbation value predicted by the model from the impact filter wear sequence, the periodic fluctuations that are highly correlated with temperature can be effectively removed, thus obtaining a pure wear evolution sequence dominated by irreversible physical wear.

[0136] Finally, the remaining lifespan of the equipment is predicted based on pure wear evolution sequences. This function is implemented by a pre-built lifespan prediction model using a long short-term memory (LSTM) neural network structure. The model is constructed by collecting a sufficient number of complete pure wear evolution sequences of the same model of equipment from the start of operation until reaching the failure criterion (e.g., wear exceeding a set threshold) as a training sample library. The input layer of the model is designed to receive sequence segments of fixed length, the length of which is based on the average duration of the wear acceleration phase in historical data, taking the statistical average of this duration. The network structure contains two cascaded LSM layers, each containing twice the length of the input sequence segment. The network output layer is a fully connected layer that receives the output of the last memory unit and maps it to a scalar value representing the predicted remaining operating time required for the equipment to reach the failure threshold after the input sequence segment. Using the prepared historical training samples, with mean squared error as the loss function, an adaptive moment estimation algorithm is used as the optimizer to iteratively train the network parameters until the model converges. When applying the trained model, the pure wear evolution sequence of the current device is divided into continuous overlapping segments according to the set segment length, and then input into the model in sequence. The model outputs a remaining lifetime prediction value for each segment. All prediction values ​​are connected in chronological order to form a remaining lifetime curve.

[0137] Reference Figure 2 The second embodiment of the present invention provides an IoT-based electromechanical equipment lifecycle management system, including:

[0138] The data acquisition module is used to acquire the load sequence and temperature sequence of electromechanical equipment to obtain the raw dataset;

[0139] The denoising module is used to perform denoising and smoothing processing on the original dataset to obtain a clean temperature sequence and a clean load sequence, and to identify temperature mutation intervals based on the clean temperature sequence.

[0140] The loss feature construction module is used to perform preliminary loss feature extraction based on the pure load sequence and the temperature change interval to obtain load frequency features and operating segment sequence, and construct a loss feature vector based on the load frequency features, the temperature change interval and the operating segment sequence.

[0141] The loss rate quantization module is used to perform generalized cross-correlation analysis based on the pure load sequence and the pure temperature sequence to obtain a synchronization offset sequence, and to perform internal loss rate quantization based on the synchronization offset sequence to obtain a loss rate index.

[0142] The loss coefficient prediction module is used to predict the loss coefficient based on the loss feature vector and the loss rate index using a pre-built loss coefficient prediction model, and obtain the loss coefficient.

[0143] The performance deviation analysis module is used to calculate a performance deviation vector containing load loss deviation, temperature response deviation and time accumulation deviation based on the loss coefficient, and to periodically correct the time accumulation deviation based on the load loss deviation and the temperature response deviation to obtain the final performance deviation vector.

[0144] The evolution trend analysis module is used to perform evolution pattern clustering and grouping based on the final performance deviation vector to obtain loss evolution trend clusters, and to perform loss acceleration analysis and secondary clustering based on the loss evolution trend clusters to obtain refined trend subclusters.

[0145] The loss analysis module is used to separate the initial wear sequence, load impact component and temperature response component based on the refined trend sub-cluster, the pure load sequence and the pure temperature sequence, and to filter out the influence of load impact on the initial wear sequence based on the load impact component to obtain the impact-filtered wear sequence.

[0146] The life prediction module is used to filter out the influence of temperature cycle changes based on the impact-filtered wear sequence and the temperature response component to obtain a pure wear evolution sequence, and to perform life prediction based on the pure wear evolution sequence to obtain the remaining life curve.

[0147] It should be noted that the IoT-based electromechanical equipment lifecycle management system provided in this embodiment of the invention is used to execute all process steps of the IoT-based electromechanical equipment lifecycle management method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0148] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0149] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for full lifecycle management of electromechanical equipment based on the Internet of Things, characterized in that, include: Obtain the load sequence and temperature sequence of the electromechanical equipment to obtain the original dataset; Based on the original dataset, denoising and smoothing processes are performed to obtain a clean temperature sequence and a clean load sequence. Then, based on the clean temperature sequence, abrupt change intervals are identified to obtain temperature abrupt change intervals. Based on the pure load sequence and the temperature jump interval, the loss features are initially extracted to obtain the load frequency features and the operating segment sequence. Then, a loss feature vector is constructed based on the load frequency features, the temperature jump interval, and the operating segment sequence. Generalized cross-correlation analysis is performed on the pure load sequence and the pure temperature sequence to obtain the synchronization offset sequence, and the internal loss rate is quantified based on the synchronization offset sequence to obtain the loss rate index. Based on the loss feature vector and the loss rate index, the loss coefficient is predicted using a pre-built loss coefficient prediction model to obtain the loss coefficient. Based on the loss coefficient, a performance deviation vector containing load loss deviation, temperature response deviation, and duration accumulation deviation is calculated, and the duration accumulation deviation is periodically corrected based on the load loss deviation and the temperature response deviation to obtain the final performance deviation vector. Based on the final performance deviation vector, evolution pattern clustering is performed to obtain loss evolution trend clusters, and loss acceleration analysis and secondary clustering are performed based on the loss evolution trend clusters to obtain refined trend sub-clusters. Based on the refined trend sub-cluster, the pure load sequence, and the pure temperature sequence, the initial wear sequence, the load impact component, and the temperature response component are separated. The influence of the load impact on the initial wear sequence is filtered out based on the load impact component to obtain the impact-filtered wear sequence. Based on the impact-filtered wear sequence and the temperature response component, the influence of temperature cycle changes is filtered out to obtain a pure wear evolution sequence. Based on the pure wear evolution sequence, lifetime prediction is performed to obtain the remaining lifetime curve.

2. The method for full lifecycle management of electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The process involves denoising and smoothing the original dataset to obtain a clean temperature sequence and a clean load sequence, and then identifying temperature abrupt change intervals based on the clean temperature sequence, including: Based on the original dataset, a first-order low-pass filtering algorithm is used for denoising, and the data is smoothed by linear interpolation based on the denoising result to obtain a first-filter load sequence and a pure temperature sequence. Based on the pure temperature sequence, the temperature gradient is calculated using the sliding window method to obtain the temperature gradient sequence; The time points when the temperature gradient sequence exceeds a preset gradient threshold are marked as abrupt change times. The consecutive abrupt change times are merged to obtain a temperature abrupt change interval that includes temperature values ​​and time points. The first-filtered load sequence corresponding to the mutation time is extracted, and a second filtering is performed using a first-order low-pass filtering algorithm to obtain a clean load sequence.

3. The method for full lifecycle management of electromechanical equipment based on the Internet of Things according to claim 2, characterized in that, The step involves preliminary extraction of loss features based on the pure load sequence and the temperature abrupt change range to obtain load frequency features and operating segment sequences, and construction of a loss feature vector based on the load frequency features, the temperature abrupt change range, and the operating segment sequences, including: Based on the pure load sequence, time-frequency decomposition is performed using short-time Fourier transform to obtain the load frequency characteristics; Obtain continuous operation segment records of the electromechanical equipment, and divide the continuous operation segment records into intervals according to the temperature change intervals to obtain an operation segment sequence containing stable segments and change intervals; Based on the load frequency characteristics, find the frequency corresponding to the peak value to obtain the main frequency, and calculate the energy percentage of the main frequency to obtain the main frequency energy percentage; Based on the temperature abrupt change range, the mean abrupt change amplitude and the standard deviation of the abrupt change amplitude are calculated. Based on the running segment sequence, the average stable running time of the stable segment is calculated by weighted average method. The frequency of the abrupt change segment within a preset period is statistically analyzed to obtain the abrupt change frequency. The main frequency energy ratio, the mean of the mutation amplitude, the standard deviation of the mutation amplitude, the mutation frequency, and the average stable running time are used as multi-dimensional feature vectors. Based on the multi-dimensional feature vectors, the features are standardized using the Z-Score standardization method to obtain the loss feature vector.

4. The method for full lifecycle management of electromechanical equipment based on the Internet of Things according to claim 3, characterized in that, The step involves performing generalized cross-correlation analysis based on the pure load sequence and the pure temperature sequence to obtain a synchronization offset sequence, and then quantifying the internal loss rate based on the synchronization offset sequence to obtain a loss rate index, including: The pure load sequence and the pure temperature sequence are divided into multiple overlapping windows through a preset sliding window, and the cross power spectrum is calculated by fast Fourier transform. Based on the cross-power spectrum, whitening is performed using a preset phase transformation weighting function, and a generalized cross-correlation function sequence with time shift as the independent variable and cross-correlation value as the dependent variable is obtained through inverse Fourier transform. The peak value of the generalized cross-correlation function sequence is extracted as the optimal time shift, and the optimal time shift corresponding to all the overlapping windows is combined to obtain the synchronization offset sequence. The time period corresponding to the synchronization offset sequence exceeding the preset offset threshold is marked as an abnormal time period to obtain the preliminary abnormal time period; Calculate the cumulative duration of the initial abnormal time period, divide the cumulative duration by the continuous operation segment record, and obtain the loss rate index.

5. The method for full lifecycle management of electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The process of constructing the loss coefficient prediction model includes: Obtain historical loss feature vectors, historical loss rate indices, and historical loss coefficients; The historical loss feature vector and historical loss rate index are input into the input layer of the initially constructed neural network model for training, and the predicted loss coefficients output by the output layer of the neural network model are obtained. Substitute the predicted loss coefficient and the historical loss coefficient into the loss function to calculate the loss value; The gradient of the output layer of the neural network model is calculated based on the loss value, and the gradient is passed forward layer by layer through the chain rule to calculate the gradient of the parameters of each layer and obtain the gradient data. Based on the gradient data and the preset learning rate, the parameters of each layer of the neural network model are updated using the gradient descent method. The parameters of each layer are iteratively updated until the number of training iterations of the neural network model is greater than the preset number of iterations, or the loss value of the neural network model is less than the preset loss threshold. At this point, the training is considered complete, and a loss coefficient prediction model that has been trained is obtained.

6. The method for full lifecycle management of electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The process involves calculating a performance deviation vector based on the loss coefficient, which includes load loss deviation, temperature response deviation, and cumulative duration deviation. The cumulative duration deviation is then periodically corrected based on the load loss deviation and the temperature response deviation to obtain the final performance deviation vector, which includes: Calculate the difference between the loss coefficient and the preset loss coefficient benchmark value to obtain a performance deviation vector that includes load loss deviation, temperature response deviation and time accumulation deviation; Based on the load loss deviation and the temperature response deviation, the time point of the peak occurrence is identified, and the average time difference between adjacent peaks is calculated to obtain the load loss cycle and the temperature response cycle. Based on the load loss deviation and the temperature response deviation, the cross-correlation function value is calculated using the cross-correlation function formula as the delay time. Based on the delay time, the load loss period, and the temperature response period, the characteristic period value is calculated using the preset coupling period formula. The cumulative time deviation is segmented using the characteristic periodic value as the period, and the first-order forward difference value of two adjacent segments is calculated to obtain the cumulative rate sequence. When the characteristic period value exceeds the preset period threshold, the duration accumulation deviation is corrected by linear weighted summation according to the cumulative rate sequence, and the performance deviation vector is updated to obtain the final performance deviation vector. When the characteristic period value does not exceed the preset period threshold, the performance deviation vector is output as the final performance deviation vector.

7. The method for full lifecycle management of electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The process involves clustering evolution patterns based on the final performance deviation vector to obtain loss evolution trend clusters, and then performing accelerated loss analysis and secondary clustering based on these loss evolution trend clusters to obtain refined trend sub-clusters, including: Based on the final performance deviation vector, the evolution pattern is clustered and grouped using the DBSCAN clustering algorithm to obtain loss evolution trend clusters containing load components, temperature components, and duration components. Based on the loss evolution trend cluster, the variance contribution rates of the load component, the temperature component, and the duration component are calculated to obtain the contribution rates of the load component, the temperature component, and the duration component. The contribution rates of the load component and the temperature component are compared with preset load contribution rate thresholds and preset temperature contribution rate thresholds, respectively, and the time periods when both exceed the thresholds are extracted to obtain the contribution overlap period. Based on the overlapping contribution period, the rising slope of the contribution rate of the duration component is calculated, and the overlapping contribution period corresponding to the rising slope exceeding the preset slope threshold is marked as the loss acceleration period, thus obtaining the loss acceleration period. Based on the loss acceleration period and the loss evolution trend cluster, a secondary clustering of the acceleration stage is performed using a hierarchical clustering algorithm to obtain refined trend sub-clusters.

8. The method for full lifecycle management of electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The process involves separating the initial wear sequence, load impact component, and temperature response component based on the refined trend sub-cluster, the pure load sequence, and the pure temperature sequence, and filtering out the influence of load impact on the initial wear sequence based on the load impact component to obtain the impact-filtered wear sequence, including: Extract the pure load sequence and the pure temperature sequence corresponding to the refined trend sub-cluster, and separate the load impact component representing short-term sudden changes, the temperature response component representing periodic changes, and the baseline change component representing long-term slow changes using the STL decomposition method. Use the baseline change component as the initial wear sequence. Based on the load impact component, identify the peak value and record the time point, and extract the local amplitude of the initial wear sequence based on the time point; The time points corresponding to the local amplitude exceeding the preset amplitude threshold are marked as load impact induced deviation times, thus obtaining a set of induced deviation times; Based on the set of induced deviation times, the abnormal fluctuations of load impact-induced deviations in the initial wear sequence are removed by a local weighted regression algorithm to obtain the impact-filtered wear sequence.

9. The method for full lifecycle management of electromechanical equipment based on the Internet of Things according to claim 8, characterized in that, The process involves filtering out the influence of temperature cycle changes based on the impact-filtered wear sequence and the temperature response component to obtain a pure wear evolution sequence, and then performing lifetime prediction based on the pure wear evolution sequence to obtain the remaining lifetime curve, including: Based on the temperature response components, the time points of the peaks and troughs are extracted to obtain the temperature peak and trough time set; Based on the impact filter wear sequence and the temperature peak and valley time set, the amplitude difference between the peak time and the valley time is calculated to obtain the temperature influence sequence; Based on the temperature influence sequence, the impact filtering wear sequence is corrected for temperature influence using a pre-built residual correction model to obtain a pure wear evolution sequence; Based on the pure wear evolution sequence, the remaining life curve is obtained by predicting the remaining life using a pre-built life prediction model.

10. A lifecycle management system for electromechanical equipment based on the Internet of Things, characterized in that, include: The data acquisition module is used to acquire the load sequence and temperature sequence of electromechanical equipment to obtain the raw dataset; The denoising module is used to perform denoising and smoothing processing on the original dataset to obtain a clean temperature sequence and a clean load sequence, and to identify temperature mutation intervals based on the clean temperature sequence. The loss feature construction module is used to perform preliminary loss feature extraction based on the pure load sequence and the temperature change interval to obtain load frequency features and operating segment sequence, and construct a loss feature vector based on the load frequency features, the temperature change interval and the operating segment sequence. The loss rate quantization module is used to perform generalized cross-correlation analysis based on the pure load sequence and the pure temperature sequence to obtain a synchronization offset sequence, and to perform internal loss rate quantization based on the synchronization offset sequence to obtain a loss rate index. The loss coefficient prediction module is used to predict the loss coefficient based on the loss feature vector and the loss rate index using a pre-built loss coefficient prediction model, and obtain the loss coefficient. The performance deviation analysis module is used to calculate a performance deviation vector containing load loss deviation, temperature response deviation and time accumulation deviation based on the loss coefficient, and to periodically correct the time accumulation deviation based on the load loss deviation and the temperature response deviation to obtain the final performance deviation vector. The evolution trend analysis module is used to perform evolution pattern clustering and grouping based on the final performance deviation vector to obtain loss evolution trend clusters, and to perform loss acceleration analysis and secondary clustering based on the loss evolution trend clusters to obtain refined trend subclusters. The loss analysis module is used to separate the initial wear sequence, load impact component and temperature response component based on the refined trend sub-cluster, the pure load sequence and the pure temperature sequence, and to filter out the influence of load impact on the initial wear sequence based on the load impact component to obtain the impact-filtered wear sequence. The life prediction module is used to filter out the influence of temperature cycle changes based on the impact-filtered wear sequence and the temperature response component to obtain a pure wear evolution sequence, and to perform life prediction based on the pure wear evolution sequence to obtain the remaining life curve.

Images