Motor system fault diagnosis method and system

Abstract

This application belongs to the field of motor system fault diagnosis technology, and relates to a motor system fault diagnosis method and system. It collects core electrical data of the motor's three-phase current and voltage, as well as auxiliary data of ambient temperature, and performs timestamp alignment to eliminate phase deviation. Then, it performs signal purification processing to suppress noise and retain effective fault features. Subsequently, it extracts multi-dimensional features and filters core features through correlation evaluation, eliminating redundant features. A fault feature library is constructed and dynamically iterated based on historical data, expert experience, and new data to achieve dynamic tracking of fault modes. A hierarchical multi-model fusion architecture is used to progressively analyze core features, and the stability of the analysis results is ensured by combining result fusion and parameter adaptive optimization. The results are parsed through operating condition adaptation logic, interference is eliminated, and graded early warnings are executed. Finally, based on fault evolution characteristics, historical degradation data, and error correction logic, the remaining lifespan and confidence interval are estimated, filling the gap in the accuracy of lifespan prediction in existing technologies.

Description

Technical Field

[0001] This application belongs to the field of motor system fault diagnosis technology, and more specifically, relates to a motor system fault diagnosis method and system. Background Technology

[0002] As a core power equipment in industrial production, energy supply, and transportation, the stability of the motor's operating status directly determines the continuity, safety, and economy of the entire production system. In industrial settings, unplanned shutdowns of motor systems are a common problem that has long plagued enterprises. These shutdowns are mostly caused by two types of core faults: mechanical faults, such as bearing wear, rotor eccentricity, and loose couplings; and electrical faults, such as winding deterioration, insulation damage, and poor contact at terminals. Regardless of the type of fault, failure to address it promptly can lead to production disruptions and order delivery delays, as well as high costs associated with emergency repairs and spare parts replacements. In extreme cases, it can even cause equipment damage, personal injury, and other safety risks, resulting in irreparable losses for the company. With the increasing level of industrial intelligence, motors are developing towards higher power, higher speed, and greater integration, leading to increasingly complex operating conditions and diverse, concealed, and coupled fault modes. Some complex faults are also affected by power frequency / inverter fluctuations, giving rise to new fault scenarios such as abnormal process matching, further exacerbating the risk of unplanned downtime. Therefore, conducting research on motor system fault diagnosis technology to achieve early fault identification, accurate fault location, and remaining lifespan prediction is of significant practical importance for ensuring reliable equipment operation, reducing maintenance costs, and mitigating safety risks.

[0003] Currently, motor system fault diagnosis technology has gradually evolved from the traditional diagnosis mode based on human experience to an intelligent diagnosis mode based on multi-source data such as electrical signals, vibration signals, and environmental parameters. However, from the perspective of actual industrial application, existing technologies still have significant shortcomings and cannot meet the core needs of enterprises for timely fault warnings, accurate diagnosis, and reliable life prediction. Most existing methods either rely on single parameter monitoring, judging the equipment status only through single indicators such as temperature and vibration, and cannot comprehensively capture the coupled characteristics of faults—for example, the early stage of winding deterioration may be accompanied by small current fluctuations, which are difficult to detect early by monitoring only temperature; or they use simple threshold alarms. The existing mechanism, which only sets a fixed parameter range, cannot adapt to dynamic scenarios such as changes in motor load and switching of operating conditions, and it cannot effectively identify complex process matching faults caused by power frequency / variable frequency fluctuations. This directly leads to the difficulty of issuing early warnings for complex faults with existing technology. It can often only issue alarms when the fault has already appeared or is about to occur, leaving maintenance personnel with very little time to deal with it. At the same time, existing technology lacks in-depth analysis of the evolution of faults. Even if faults can be identified, it is difficult to accurately predict the remaining lifespan of the equipment. It cannot provide a scientific basis for enterprises to formulate preventive maintenance plans, and can only passively deal with fault downtime, making it difficult to fundamentally avoid the risk of unplanned downtime. In summary, existing motor system fault diagnosis methods are limited by single parameter monitoring, simple threshold mechanisms, and data processing deficiencies. They cannot provide effective early warnings for complex faults in advance, nor can they accurately predict the remaining lifespan of equipment. As a result, enterprises still face the risk of high-frequency unplanned downtime. They cannot meet the needs of motor systems under complex operating conditions for accurate early warning, efficient diagnosis, and full life cycle status assessment. This problem has become the core bottleneck restricting the application of intelligent motor diagnosis technology and helping enterprises achieve lean operation and maintenance. Summary of the Invention

[0004] This invention provides a method and system for diagnosing motor system faults, aiming to solve the technical problems of existing motor system fault diagnosis methods being limited by single parameter monitoring, simple threshold mechanisms and data processing defects, which make it impossible to provide effective early warning of complex faults (including process matching anomalies caused by power frequency / frequency conversion fluctuations) in advance, and difficult to accurately predict the remaining life of equipment, leading to enterprises facing the risk of high-frequency unplanned downtime.

[0005] On one hand, the present invention provides a method for diagnosing faults in a motor system, comprising the following steps: Data acquisition and synchronization processing: Collect core electrical data such as three-phase current and three-phase voltage of the motor and auxiliary data such as ambient temperature. Perform timestamp alignment processing on multi-source acquired data to eliminate feature distortion caused by phase deviation and ensure data timing consistency. Signal purification processing: Perform noise suppression and timing optimization processing on the timestamp-aligned data, retain the effective features related to faults in the data, and output standardized time-series data; Multidimensional feature extraction and screening: Extract multidimensional features from the cleaned time series data, screen core features and remove redundant features through feature correlation evaluation logic; Fault Feature Library Construction and Dynamic Iteration: An initial fault feature library is constructed based on historical data and expert experience. By combining newly added data and using similarity matching and time series analysis logic, the feature library is dynamically updated and fault mode tracking is achieved. Multi-model fusion analysis: A hierarchical model architecture is used to perform progressive analysis of core features. Through result fusion and parameter adaptive optimization logic, stable analysis results are output. Fault diagnosis and early warning generation: Parse and integrate the analysis results, combine the working condition adaptation logic to eliminate interference factors, output fault information and execute the hierarchical early warning strategy; Remaining lifetime estimation: Based on fault evolution characteristics and historical degradation data, the remaining lifetime and confidence interval of the equipment are output through condition assessment and error correction logic.

[0006] This invention collects core electrical data of the motor's three-phase current and voltage, along with auxiliary data of ambient temperature, and aligns them with timestamps to eliminate phase deviations and ensure data timing consistency. Signal purification then suppresses noise and retains effective fault features. Multidimensional features are extracted, and core features are screened through correlation evaluation, eliminating redundant features and comprehensively capturing fault coupling characteristics. A fault feature library is constructed and dynamically iterated based on historical data, expert experience, and new data to achieve dynamic tracking of fault modes and adapt to complex fault evolution and new fault scenarios. A hierarchical multi-model fusion architecture is used to progressively analyze core features, combining result fusion and adaptive parameter optimization to ensure the stability of analysis results, replacing simple threshold mechanisms and improving the accuracy of complex fault identification. Through operating condition adaptation logic analysis results, interference elimination, and hierarchical early warning, effective early warning of complex faults is achieved. Finally, based on fault evolution characteristics, historical degradation data, and error correction logic, the remaining lifespan and confidence interval are estimated, filling the gap in the accuracy of existing lifespan prediction technologies.

[0007] Preferably, the multi-dimensional features include time-domain, frequency-domain, and time-frequency-domain features. The time-domain features obtain feature parameters through signal amplitude statistical logic, the frequency-domain features analyze the fundamental wave, harmonic components, and distortion features through signal frequency conversion logic, and the time-frequency-domain features generate a time-frequency distribution map through multi-scale transformation logic to capture transient fault features. Then, through feature contribution evaluation logic, the correlation coefficient between each feature and the fault type is calculated, and the correlated features with correlation above the threshold are sorted and filtered, while redundant features with correlation below the threshold are removed, forming a core feature set with timestamps and operating condition labels.

[0008] Preferably, the feature contribution evaluation logic is as follows: Principal component analysis logic is used to reduce the dimensionality of the extracted multidimensional features, calculate the variance contribution rate of each feature, set a screening threshold based on the cumulative variance contribution rate, retain features with a cumulative contribution rate higher than the threshold as core features, and remove redundant features with a contribution rate lower than the threshold.

[0009] Preferably, the construction and dynamic iteration of the fault feature library includes the following steps: During the initial construction, the feature vectors extracted from historical fault data are standardized, and feature templates are stored hierarchically according to fault type to establish a mapping relationship between fault type and feature vector; During dynamic iteration, the matching degree between newly added features and existing templates in the library is compared through similarity calculation logic. For templates with matching degree higher than a preset threshold, weighted updates are performed to integrate the newly added feature information. New features with matching degree lower than a preset threshold are identified as new fault modes and new templates are added. Templates with similarity higher than the threshold are merged periodically through clustering logic. At the same time, the evolution of fault features over time is tracked through time series analysis to adapt to the feature changes of progressive faults.

[0010] Preferably, the hierarchical model architecture processes the faults in the order of anomaly detection, fault classification, and fault prediction. First, the anomaly detection model performs unsupervised learning to traverse and analyze the core feature set, delineates the normal feature distribution range, identifies abnormal data points and corresponding feature subsets that deviate from the normal feature distribution range, and outputs the anomaly labeling results and anomaly confidence. The anomaly labeling results and corresponding feature subsets are fed into the fault classification model. The fault classification model, based on feature matching logic, compares the anomaly features with various fault templates in the fault feature library one by one, calculates the feature matching degree, and outputs the corresponding fault type, type confidence and fault association features based on the feature matching degree. Input the fault type, fault association features and historical time series features into the prediction model, mine the change pattern of features over time through time series association, predict the fault development trend and occurrence probability, and output the probability prediction results and time series trend curve. The confidence level, type confidence level, and probability prediction results of the three types of models are assigned weights through weighted voting. The weight values ​​are dynamically adjusted based on the historical processing accuracy of each model. The higher the accuracy, the greater the weight. The results of each model are multiplied by their corresponding weights and then summed to obtain the comprehensive analysis score and corresponding results. The system uses adaptive parameter tuning logic to monitor the stability of the comprehensive analysis score and its deviation from historical results in real time. If the deviation exceeds the allowable range, it automatically optimizes the core parameters of each model and the weighted voting weight ratio, and iterates the calculation until the stability of the comprehensive analysis result meets the standard. Finally, it outputs a fusion analysis result that includes the location of the anomaly, the type of the fault, the probability of occurrence, and the time series trend.

[0011] Preferably, the fault diagnosis and early warning generation includes the following steps: First, the abnormal location, fault type, occurrence probability, and time series trend results output by the multi-model fusion are analyzed to extract the core fault diagnosis information and calculate the corresponding confidence level. When the confidence level reaches the set threshold, the fault type, severity, and associated causes are initially output to form the initial fault diagnosis results. Then, pseudo-fault correction is performed through process matching analysis to form a corrected fault diagnosis result. Based on the corrected fault diagnosis result, the comprehensive risk level is divided according to the fault type, severity and probability of occurrence. Different risk levels correspond to differentiated early warning response logic.

[0012] Preferably, the pseudo-fault correction includes the following steps: Based on preset process standard parameters, the deviation range of real-time power parameters, load parameters and benchmark parameters is compared. Combined with the fault correlation characteristics output by multi-model fusion, it is determined whether there is a causal relationship between parameter deviation and fault characteristics. That is, if the parameter deviation exceeds the allowable range but there is no corresponding fault characteristic to support it, or if the deviation range is within the allowable range of process fluctuation, it is determined to be a false fault signal. The corresponding items in the initial fault diagnosis results are marked, removed or corrected. At the same time, the parameter deviation details and the basis for the false fault judgment are recorded to form the corrected fault diagnosis results.

[0013] Preferably, the remaining lifetime estimation includes the following steps: By integrating historical degradation data with real-time fault evolution characteristics, a device degradation model is constructed using a state estimation algorithm. The model parameters are initialized and real-time feature data is input to obtain the basic life prediction result. By analyzing the correlation and impact intensity among multiple faults through fault coupling identification logic, correction factors are introduced to adjust the basic prediction results. Uncertainty modeling is employed to sample and analyze measurement noise and model error, quantify the prediction error, and output the remaining lifetime range at different confidence levels.

[0014] Preferably, the state estimation algorithm employs a particle filtering algorithm, combining historical health status data and initial parameters of the equipment to generate a particle set that conforms to the actual initial state distribution of the equipment. This particle set is then substituted into a preset equipment degradation model, and the state prediction value corresponding to each particle is obtained through model iteration. Real-time fault evolution feature data is then called to compare the deviation between the predicted value of each particle and the real-time feature data, and the particle weights are updated according to the magnitude of the deviation, with smaller deviations resulting in higher weight proportions. Subsequently, a resampling and removal operation is performed on weighted particles with weights below the preset value, retaining the effective particle set with weights above the preset value. Based on the statistical results of the effective particle set, the basic lifetime prediction result is output. During the fault coupling identification logic analysis, the fault association identification logic first iterates through the currently detected faults and historical fault records to distinguish between causal and parallel relationships between faults, quantifying the impact intensity of each fault on the equipment degradation rate. Based on the impact intensity, a correction factor weight is assigned to each fault, with higher weights for greater impact intensity. A weighted calculation yields a comprehensive correction factor. The basic lifespan prediction result is then superimposed with the comprehensive correction factor to correct the prediction deviation caused by multi-fault coupling, resulting in a preliminarily corrected lifespan value. The uncertainty modeling adopts Monte Carlo simulation, which performs multiple random samplings on potential error sources such as measurement noise, model parameter errors, and data fluctuations. Each sampling result is substituted into the above state estimation and coupling correction process to obtain multiple sets of lifetime prediction data. Statistical analysis is performed on multiple sets of data to fit the error distribution pattern. Based on the distribution pattern, numerical intervals corresponding to different confidence levels are defined, and finally, the remaining lifetime interval with confidence level label is output.

[0015] On the other hand, the present invention provides a motor system fault diagnosis system for implementing the motor system fault diagnosis method described herein, the system comprising: Data acquisition module: Used to collect core electrical data such as three-phase current and three-phase voltage of motor and auxiliary data such as ambient temperature, and has the functions of multi-source synchronous acquisition and timestamp alignment; Signal processing module: Communicates with the data acquisition module and is used to perform filtering and noise reduction, power frequency interference suppression, amplitude normalization and multi-scale noise reduction on the acquired electrical data, and output high-quality time-series data; Feature extraction module: Communicates with the signal processing module, integrates power feature analysis tools, and is used to extract time-domain, frequency-domain and multi-dimensional features. It filters core features through the feature contribution evaluation unit and generates a labeled multi-dimensional feature set. Fault Feature Library Module: Communicates with the feature extraction module, uses a database to store feature templates, and has functions such as template initialization, online updating, clustering and merging, and fault mode evolution tracking. It can dynamically adapt to changes in fault types. Algorithm model module: It communicates with the feature extraction module and the fault feature library module respectively, and deploys a three-layer algorithm model, a weighted voting fusion unit, and an adaptive parameter tuning unit to realize anomaly detection, fault classification and fault prediction. Diagnostic and early warning module: It communicates with the algorithm model module to parse the model fusion results, generate fault diagnosis information, execute multi-level early warning strategies, combine the process matching analysis unit to correct the results, and output early warning notifications and maintenance suggestions. The life prediction module is connected to the feature extraction module, the fault feature library module, and the diagnosis and early warning module. It uses a state estimation unit, a coupling effect modeling unit, and an uncertainty analysis unit to output the remaining life estimate and confidence interval.

[0016] The beneficial effects of the invention include: This invention collects core electrical data of the motor's three-phase current and voltage, along with auxiliary data of ambient temperature, and aligns them with timestamps to eliminate phase deviations and ensure data timing consistency. Signal purification then suppresses noise and retains effective fault features. Multidimensional features are extracted, and core features are screened through correlation evaluation, eliminating redundant features and comprehensively capturing fault coupling characteristics. A fault feature library is constructed and dynamically iterated based on historical data, expert experience, and new data to achieve dynamic tracking of fault modes and adapt to complex fault evolution and new fault scenarios. A hierarchical multi-model fusion architecture is used to progressively analyze core features, combining result fusion and adaptive parameter optimization to ensure the stability of analysis results, replacing simple threshold mechanisms and improving the accuracy of complex fault identification. Through operating condition adaptation logic analysis results, interference elimination, and hierarchical early warning, effective early warning of complex faults is achieved. Finally, based on fault evolution characteristics, historical degradation data, and error correction logic, the remaining lifespan and confidence interval are estimated, filling the gap in the accuracy of existing lifespan prediction technologies. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The above is an overall flowchart provided for an embodiment of the present invention.

[0019] Figure 2 The flowchart of step S2 provided in the embodiment of the present invention is shown.

[0020] Figure 3 The flowchart of step S4 provided in the embodiment of the present invention is shown.

[0021] Figure 4 A simplified system architecture diagram provided for an embodiment of the present invention. Detailed Implementation

[0022] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application. Example

[0023] See Figure 1As shown, this embodiment provides a method for diagnosing motor system faults, including the following steps: S1. Data Acquisition and Synchronization Processing High-precision Hall effect current and voltage sensors are used and installed at the input end of the motor power supply circuit to avoid data distortion caused by transmission line loss. The sampling rate is set to 12kHz to fully capture high-frequency components and transient changes. The motor's three-phase current and three-phase voltage signals are continuously collected. At the same time, environmental data is collected synchronously through an integrated temperature and humidity sensor as auxiliary data for subsequent elimination of environmental interference.

[0024] In this embodiment, a timestamp is added to each acquired signal through a hardware triggering mechanism. Based on the timestamp, the three-phase current, three-phase voltage and ambient temperature data are aligned to eliminate phase deviation caused by signal asynchrony and ensure that the multi-source data at the same time correspond one-to-one.

[0025] S2. Signal purification processing See Figure 2 As shown, in this embodiment, signal purification is achieved through a progressive processing flow, while retaining effective features related to the fault, avoiding feature loss caused by traditional filtering, as detailed below: First, Butterworth low-pass filtering is applied to the timestamp-aligned current and voltage signals, with a cutoff frequency of 2kHz. Utilizing the flat frequency response within the bandpass, high-frequency electromagnetic interference above 2kHz is removed while preserving the fundamental and low-order harmonics. Then, an adaptive filtering algorithm is used to suppress 50Hz power frequency interference. By dynamically adjusting the filtering parameters based on real-time power frequency signal characteristics, this method offers stronger adaptability to power frequency shifts caused by grid fluctuations compared to fixed-parameter filters. Specifically, in this embodiment, a least mean square adaptive filtering algorithm is used to suppress 50Hz power frequency interference. By dynamically updating the filter weight vector based on real-time amplitude and phase drift characteristics of the power frequency signal, precise adaptation to grid fluctuations is achieved. Preferably: Setting the core parameters of the filter: Filter order (Adaptable to power frequency and 2nd-3rd harmonic components, covering a fluctuation range of 49.5Hz-50.5Hz); Initial weight vector (Initial state has no filtering bias to avoid introducing additional interference); Step size factor (Balancing convergence speed and stability, calibrated with historical data, the value ranges from 0.005 to 0.02. Too large a value is prone to oscillation, while too small a value results in slow convergence.)

[0026] Construction of input and desired signals: Input vector Taken from the original current / voltage signal after timestamp alignment. At the current sampling time, For the first The sampled values ​​at time t, the vector elements are continuous Several sampling points are used to capture the timing characteristics of the power frequency signal; the desired signal Generated using a power frequency synchronization estimation strategy, based on real-time power grid frequency detection values. (Real-time acquisition via phase-locked loop (PLL)) Constructs a reference signal that is in phase and frequency with the power frequency interference, i.e. ,in: This is an estimated value of the power frequency interference amplitude (obtained by peak detection of the previous frame's signal and updated in real time). The sampling period; This is the phase compensation value (the phase of the power frequency signal is detected in real time by PLL and dynamically corrected to ensure that the reference signal is consistent with the actual interference phase).

[0027] Filtering Output and Error Calculation: The filtering output is calculated using the current weight vector, and then compared with the desired signal to obtain the error, which serves as the basis for weight adjustment; where the filtering output... (The estimated power frequency interference signal) is: ; In the formula: For the first Time of the first Weight values ​​of the first-order filter; Error signal (i.e., the prototype of the purified signal after eliminating power frequency interference) is The error value reflects the estimation accuracy of the current weight vector for power frequency interference. The larger the error, the greater the weight adjustment range.

[0028] Dynamic weight vector update: The weight vector is iteratively updated based on the error signal, so that the filtered output gradually approximates the actual power frequency interference. ; In the formula: This serves as the weight vector for the next sampling time, enabling dynamic parameter iteration. To adjust the coefficients, use the step size factor. Control the magnitude of weight updates at each step to avoid iterative oscillations; The weights are the product of the error and the input signal. This ensures that the direction of weight updates is consistent with the changes in the characteristics of the power frequency signal. That is, when the error is large, the weights are adjusted in the direction of reducing the error, and the amplitude and phase drift of the power frequency interference are tracked in real time.

[0029] Through the aforementioned dynamic adjustment logic, the filter weight vector is updated every frame (100ms), enabling rapid adaptation to power frequency offset (49.5Hz-50.5Hz) and amplitude fluctuations. Compared to fixed-parameter filters, it can stably improve the power frequency interference suppression ratio to over 40dB, while preserving the effective characteristics of the fundamental and low-order harmonics related to motor faults, providing a high-quality signal foundation for subsequent amplitude normalization and wavelet denoising. Compared to fixed-parameter filters, it exhibits stronger adaptability to power frequency offsets caused by power grid fluctuations, and can improve the power frequency interference suppression ratio to over 40dB.

[0030] Furthermore, in this embodiment, in order to eliminate the influence of amplitude fluctuations under different loads and operating conditions, the filtered signal is normalized and the signal amplitude is mapped to the [0,1] interval to ensure the comparability of subsequent features. For non-stationary signals such as current surges under variable frequency power supplies, wavelet threshold denoising optimization is introduced: the signal is decomposed into 5-level wavelet decomposition (using the db4 wavelet basis), and the signal is decomposed into high-frequency detail components and low-frequency trend components. The threshold is set to 0.02 (based on historical data calibration), and high-frequency components (corresponding to noise) exceeding the threshold are suppressed. Then, a clean signal is generated through wavelet reconstruction. Based on this, multi-scale analysis is used to balance noise removal and transient feature preservation, which significantly enhances the robustness of the signal under complex operating conditions and lays a high-quality data foundation for subsequent feature extraction.

[0031] S3. Multi-feature extraction and filtering In this embodiment, three types of features—time domain, frequency domain, and time-frequency domain—are extracted from the purified time-series data to comprehensively characterize the motor's operating status. Time-domain characteristics: obtained through signal amplitude statistics, including root mean square (RMS) value, peak factor, and waveform distortion rate; among which, the RMS value reflects the electrical load state of the motor, the peak factor is used to identify impact fluctuations caused by bearing wear, etc., and the waveform distortion rate characterizes waveform abnormalities caused by winding deterioration. The amplitude variation pattern is directly captured through these three types of characteristics.

[0032] Frequency domain characteristics: Perform a Fast Fourier Transform (FFT) on the purified signal to convert it into a frequency domain spectrum, and analyze the fundamental frequency (consistent with the power supply frequency), the 3rd / 5th lower harmonic components, and the total harmonic distortion rate; excessive lower harmonics are related to winding faults and power supply distortion, and the total harmonic distortion rate can comprehensively assess the degree of signal harmonic pollution, providing a basis for electrical fault diagnosis.

[0033] Time-frequency domain features: A time-frequency distribution map (horizontal axis is time, vertical axis is frequency, and highlight represents amplitude) is generated through continuous wavelet transform, which solves the problem that FFT cannot take into account both time and frequency resolution, and accurately captures transient fault features such as instantaneous current surges caused by bearing wear and voltage surges caused by frequency converter power supply switching.

[0034] All the features extracted above are labeled with corresponding timestamps and operating condition labels (such as 75% rated load, 25% light load, etc.) to form an initial multidimensional feature set.

[0035] Furthermore, in this embodiment, principal component analysis is introduced as the feature contribution evaluation logic, specifically: The initial multidimensional feature set is standardized to eliminate dimensional differences. Then, the covariance matrix of the standardized feature matrix is ​​calculated, and the eigenvalues ​​and eigenvectors are solved. The magnitude of the eigenvalues ​​corresponds to the contribution of the principal components. The cumulative contribution rate threshold is set to 90%. The top k principal components are selected to form a simplified core feature set, and redundant features with contribution rates below the threshold are removed. While retaining the core information, the feature dimension is reduced by more than 60%, which significantly reduces the computational cost of the subsequent model and improves the model convergence speed.

[0036] S4. Fault Feature Library Construction and Dynamic Iteration Initial build: See Figure 3 As shown, an initial database is constructed based on historical fault data (such as maintenance records and fault cases from the past 5 years) and the experience of domain experts. Specifically, historical operating data and corresponding fault records are collected, covering four types of faults: mechanical anomalies (bearing wear, rotor imbalance), electrical anomalies (winding deterioration, insulation degradation), power supply problems (power frequency / variable frequency fluctuations), and process matching anomalies. For each type of fault, the fault type, severity level (level 1-5, level 1 is minor, level 5 is severe), and operating condition label are manually labeled. Data processing steps S2 and S3 are performed on the historical data to obtain feature vectors. After standardization, feature templates are generated. Each template contains a standardized feature vector, fault type, severity level, timestamp, and operating condition label. These templates are stored hierarchically in a MySQL database according to fault type. The initial confidence weight of the templates is set to 1.0, establishing a precise mapping relationship between fault types and feature vectors.

[0037] Dynamic iteration: The library is dynamically updated through an online learning mechanism. Updates are triggered by fault events or a scheduled event every 7 days, as detailed below: When the system identifies a new fault (the model fusion confidence exceeds the threshold), it automatically extracts the core feature vector of the event, the associated fault type, the severity level, and the operating condition label, and standardizes it according to the mean and standard deviation of the existing library. The matching degree between the new feature vector and the existing templates in the library is calculated using Euclidean distance, as shown in the following formula: ; In the formula: Represents Euclidean distance; , These are the i-th components of the new vector and the old template, respectively; For example, in this embodiment, a similarity threshold is set. ; like (Match successful), the existing template is updated using a weighted average, where the feature vector update expression is as follows: ; In the formula: As a forgetting factor, it balances historical and new data; This represents the standardized feature vector of the updated fault feature template; This represents the standardized feature vector of existing fault templates in the library; This represents the core feature vector of newly added fault data after processing by S2 and S3 and standardization. The severity level update formula is: ; In the formula: This indicates the severity level of the fault after the update. As a rank weighting factor; This indicates the severity level of existing fault templates in the database; This indicates the severity level of the newly added fault data; at the same time, the template confidence weight is increased by 0.1. The higher the weight, the more verified the accuracy of the corresponding template matching, and the higher the priority. The weight ranges from 1.0 to 5.0, and will not increase further after the weight reaches 5.0.

[0038] like (Match failed) As an independent feature template, it is added to the fault feature library with an initial confidence weight of 1.0. At the same time, the corresponding fault type, severity level and operating condition label are labeled to enrich the coverage of the fault feature library.

[0039] Furthermore, in this embodiment, similar templates are merged using the K-means clustering algorithm at each predetermined time period (e.g., 30 days), and the number of clusters is [not specified]. N represents the total number of templates. Minimize the sum of squares within each cluster, take the weighted average eigenvector of each cluster as the new template, delete redundant templates, and avoid excessive expansion of the library.

[0040] Secondly, for progressive failures, time series analysis is introduced to track the trend of feature changes. For example, for progressive failure templates such as bearing wear, the feature vector components and timestamps of the most recent 50 updates are additionally stored to construct a time series queue. A linear regression model is used to fit the change law of features over time. For example, the expression is as follows: ; In the formula: The intercept term represents the baseline value of the j-th feature component at the initial timing, which is obtained by fitting the data from the most recent 50 feature template updates. The slope term represents the rate of change of the j-th feature per unit time, characterizing the rate of deterioration of the progressive fault. If b is positive, it means that the feature component increases with time; if b is negative, it means that the feature component decreases with time. The magnitude of its absolute value corresponds to the speed of fault deterioration. Represents the residual; This represents the evolution value of the j-th fault feature component over time, where j is the index of the feature vector component; This represents a time variable, with the baseline being the moment when the progressive fault characteristic was first detected. The unit is consistent with the motor running time statistics, and the value range covers the time span of the time series queue. Then, the parameters of the above model are calculated using the least squares method. , and residuals The model is stored as a template evolution parameter in the fault feature library, providing a core basis for the construction of the degradation model of the subsequent remaining lifetime prediction module. In this embodiment, the limitations of the static library are overcome, enabling the system to accurately adapt to the gradual changes in faults and improve the ability to identify early minor faults.

[0041] S5. Multi-model fusion analysis This embodiment employs a hierarchical model architecture encompassing anomaly detection, fault classification, and fault prediction. Weighted voting fusion of results and adaptive parameter tuning ensure stable and reliable analysis results. Specifically: Anomaly detection model: An unsupervised model is built based on the isolated forest algorithm. Taking the core feature set as input, 100 isolated trees are constructed to evaluate the degree of anomaly of the sample. Anomalies are easily isolated and thus obtain high anomaly scores. The model outputs anomaly labeling results and anomaly confidence, which serve as a primary filtering layer to quickly identify data points that deviate from the normal state. It does not rely on fault labels and is suitable for detecting unknown anomalies.

[0042] Fault classification model: A multi-classifier is built based on support vector machine and trained using historical fault data. The optimal classification hyperplane is found through RBF kernel function mapping. At the same time, transfer learning is introduced to reuse historical feature knowledge, which accelerates the convergence and generalization ability of the model. The subset of abnormal features output by anomaly detection is input into the fault classification model and matched one by one with the fault feature library template. The feature matching degree is calculated, and the fault type, type confidence (0-1 interval) and fault association features are output.

[0043] Fault prediction model: A time-series prediction model is constructed based on a long short-term memory network. The input fault type, associated features and historical time-series features are used to learn the changing patterns of features over time, determine the fault development trend and the probability of occurrence in the next 7-30 days, and output the probability prediction results and time-series trend curve. At the same time, the remaining lifespan of the equipment is initially estimated.

[0044] A weighted voting mechanism is then used to fuse the results of the three models. Weights are assigned based on the contribution of each model: anomaly detection model (0.3, primary indication), fault classification model (0.4, core diagnosis), and prediction model (0.3, prospective prediction). These weights can be dynamically adjusted based on the model's historical processing accuracy; higher accuracy results in a larger weight. The fusion confidence score is calculated using the following formula: ; In the formula: Indicates abnormal scores; Indicates classification confidence level; Indicates the probability of a failure occurring; like (Preset threshold) Output fusion analysis results, including anomaly location, fault type, occurrence probability, and time series trend.

[0045] Furthermore, in this embodiment, the deviation between the stability of the fusion result and the historical real result is monitored in real time through adaptive parameter tuning logic. If the deviation exceeds the allowable range of 5%, the model parameters are automatically optimized. For example, the number of hidden layer nodes of LSTM is dynamically adjusted (adjusted in the range of 8-64), and the gamma value of RBF kernel of SVM is adjusted (in the range of 0.01-10). At the same time, the weighted voting weight ratio is iteratively adjusted until the stability of the result meets the standard.

[0046] S6. Fault Diagnosis and Early Warning Generation Fault diagnosis and pseudo-fault correction: The analysis results are analyzed and the core information such as fault type, abnormal location and occurrence probability is extracted. The diagnostic confidence is calculated. When the confidence is greater than 0.8, the fault type, severity and related causes are initially output in combination with the fault feature library information to form the initial diagnosis result.

[0047] Furthermore, in this embodiment, in order to eliminate interference from non-equipment factors, pseudo-fault correction is performed through process compatibility analysis. The specific steps are as follows: Using preset process standard parameters as a benchmark (such as allowable range of power frequency deviation ±0.5Hz, allowable range of load fluctuation ±10%), compare the deviations of real-time power parameters, load parameters and benchmark parameters; combine fault correlation characteristics to determine causal relationship, that is, if the parameter deviation exceeds the range but there is no corresponding fault characteristic to support it, or the deviation is within the allowable range of process fluctuation, it is determined to be a false fault signal, and the corresponding items of the initial result are marked, removed or corrected; at the same time, the parameter deviation details and the basis for the false fault judgment are recorded to form the corrected diagnosis result.

[0048] Tiered early warning strategy: Based on the corrected diagnostic results, combined with the fault type, severity (standardized to a 0-1 range), probability of occurrence, and process compatibility index (Mismatch, 0 = no bias, 1 = biased), the comprehensive risk level is calculated: ; In the formula: ; Indicates the overall risk level; Indicates the probability of a failure occurring; Indicates the severity of the fault; Indicates the process compatibility index; according to Classify warning levels and implement differentiated responses, for example: Low-level warning: At that time, it is only recorded in the system log, without proactive notification; Intermediate warning: In such cases, notify maintenance personnel via industrial HMI interface or email, and suggest prioritizing the issue during planned maintenance. Advanced warning: If the fault occurs, immediately trigger the equipment shutdown procedure and urgently notify maintenance personnel via SMS or telephone, requesting immediate action to prevent the fault from escalating.

[0049] S7. Remaining life estimate In this embodiment, a particle filter algorithm is used to construct the device degradation model, as detailed below: Particle set initialization: Define health status variables (Including current values ​​of degradation features and degradation rate), 2000 particles are randomly selected from the initial degradation distribution of similar faults in the fault feature database. Initial weights ; State prediction: Constructing state transition equations based on historical degradation models ,in It is a linear degenerate transfer function. Gaussian process noise Predict each particle Always in good health; Indicates the state at the previous moment; Weight update: Combining real-time fault evolution characteristic observations By observing the likelihood function (follows a Gaussian distribution) , For the observation function, To observe the noise covariance, The weights are updated based on the current state. The smaller the deviation, the higher the weight percentage. After the update, normalization is performed to ensure that the sum of the weights is 1.

[0050] Resampling and lifetime calculation: Particles with a weight below 0.0001 are removed, and high-weight particles are replicated to prevent particle degradation; a device failure threshold is set. (From the fault feature library), predict the time from the current moment to the point where each effective particle reaches its destination. time Basic remaining life For the updated particle weights, Indicates the number of particles.

[0051] Multi-fault coupling correction: Since industrial motors often suffer from multi-fault coupling (such as bearing wear exacerbating winding deterioration), this embodiment introduces a correction factor to adjust the basic prediction results. Specifically, Granger causality tests are used to identify the direction and strength of coupling between faults and quantify the coupling influence coefficient. Calculate the comprehensive correction factor : ; ; In the formula: Indicates the number of faults; This represents the coupling influence coefficient of the m-th type of fault on the n-th type of fault; it is quantified by the Granger causality test. The test analyzes the causal relationship between the time series data of the two types of faults, outputs the coupling significance index, and then standardizes it into a 0-1 interval value—a value of 0 indicates that the m-th type of fault has no coupling influence on the n-th type of fault, a value of 1 indicates that the influence reaches its maximum value, and the intermediate values ​​are linearly distributed according to the influence intensity. , These represent the standardization severity of the m-th and n-th types of faults, respectively. This indicates the corrected lifespan.

[0052] Uncertainty Modeling: In this embodiment, Monte Carlo simulation is used to quantify uncertainties such as measurement noise, model error, and coupling coefficient error. Specifically: Random sampling was performed on the three types of error sources (measurement noise from...). Sampling (model parameters are sampled within the confidence interval) is performed, and the simulation is repeated 1000 times. The state estimation and coupling correction process described above is executed in each simulation to obtain 1000 sets of lifetime prediction data. The data are statistically analyzed to fit the error distribution pattern and output the lifetime interval (2.5% quantile - 97.5% quantile) at 95% confidence level. Finally, the remaining lifetime result with confidence level label and the equipment health report are generated.

[0053] Example 2 This embodiment provides a motor system fault diagnosis system, see [link / reference] Figure 4 As shown, it includes: The data acquisition module consists of a Hall effect current sensor, a voltage sensor, an integrated temperature and humidity sensor, and a high-speed data acquisition card. It features multi-source synchronous acquisition and timestamp alignment. The sensors are installed at the input end of the motor power supply circuit. The data acquisition card uses a PCIe interface (sampling rate up to 20kHz, supports 8 analog inputs, and achieves microsecond-level timestamp addition through a hardware trigger mechanism). It synchronously acquires three-phase current, three-phase voltage, and ambient temperature data. After preliminary encapsulation, the acquired data is transmitted to the signal processing module to ensure data integrity and timing consistency.

[0054] Signal processing module: Communicates with the data acquisition module and is deployed on an ARM Cortex-A9 embedded processor (1GHz clock speed, 2GB memory). It integrates filtering and denoising, power frequency interference suppression, amplitude normalization, and multi-scale denoising algorithms. At the software level, it uses C language programming to implement lightweight algorithm deployment, which can process the acquired data in real time and output purified standardized time-series data. It also has a data caching function (16GB cache capacity, which can store 72 hours of continuous data) to avoid data loss and provide stable input for the feature extraction module.

[0055] Feature extraction module: Communicates with the signal processing module and integrates the Power Characteristic Analysis (ESA) toolkit. At the software level, it implements time-domain, frequency-domain, and time-frequency-domain feature extraction algorithms based on Python, and also has a built-in feature contribution evaluation unit (PCA algorithm). After receiving the purification time series data, this module automatically extracts multi-dimensional features, filters core features through PCA, and generates a multi-dimensional feature set with timestamps and operating condition labels, which is then transmitted to the fault feature library module and the algorithm model module, respectively. At the same time, the original feature data is retained for subsequent traceability.

[0056] Fault Feature Library Module: Communicates with the feature extraction module, uses a MySQL 8.0 database (deployed on an industrial server), and features template initialization, online updates, cluster merging, and fault mode evolution tracking. The database stores feature templates hierarchically according to fault type, and achieves data interaction with other modules through a customized API interface. It supports weighted template updates, similar cluster merging, and evolution parameter storage, and can dynamically adapt to changes in fault type. It also has a data backup function (automatic daily backup, retaining 30 days of historical versions) to ensure data security.

[0057] Algorithm Model Module: Communicates with the feature extraction module and the fault feature library module respectively, and is deployed on edge computing devices or cloud platforms, supporting on-demand deployment; The algorithm model module has a built-in three-layer algorithm model (Isolated Forest, SVM, LSTM), and also integrates a weighted voting fusion unit and an adaptive parameter tuning unit: Real-time model inference is achieved through edge computing, the cloud platform is responsible for model parameter optimization and retraining, the weighted voting unit fuses model results according to preset weights, and the adaptive parameter tuning unit monitors model performance and optimizes parameters in real time to ensure stable and reliable analysis results.

[0058] Diagnostic and early warning module: Communicatively connected to the algorithm model module, it consists of an industrial HMI touchscreen, an alarm interface, and a notification module, integrating a process matching analysis unit. The diagnostic and early warning module analyzes the model fusion results, generates structured fault diagnosis information, and performs pseudo-fault correction through the process matching analysis unit. Based on the comprehensive risk level, it triggers graded early warnings, displays the warning information through the HMI interface, and simultaneously links equipment shutdown (advanced early warning) through the alarm interface (relay output), and pushes notifications via email and SMS modules, outputting maintenance suggestions.

[0059] The lifespan prediction module is communicatively connected to the feature extraction module, fault feature library module, and diagnostic early warning module. Deployed on an edge computing device, it incorporates a state estimation unit (particle filtering algorithm), a coupling effect modeling unit, and an uncertainty analysis unit (Monte Carlo simulation). This module integrates historical degradation data with real-time fault evolution characteristics, obtains a basic lifespan prediction result through particle filtering, corrects it through coupling effect modeling, quantifies the error through Monte Carlo simulation, and outputs a remaining lifespan estimate, a 95% confidence interval, and an equipment health report. The report can be displayed through the HMI interface or exported as a PDF, providing data support for predictive maintenance.

[0060] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method of diagnosing a fault in an electric motor system, characterized by, Includes the following steps: Data acquisition and synchronization processing: Collect core electrical data such as three-phase current and three-phase voltage of the motor and auxiliary data such as ambient temperature. Perform timestamp alignment processing on multi-source acquired data to eliminate feature distortion caused by phase deviation and ensure data timing consistency. Signal purification processing: Perform noise suppression and timing optimization processing on the timestamp-aligned data, retain the effective features related to faults in the data, and output standardized time-series data; Multidimensional feature extraction and screening: Extract multidimensional features from the cleaned time series data, screen core features and remove redundant features through feature correlation evaluation logic; Fault Feature Library Construction and Dynamic Iteration: An initial fault feature library is constructed based on historical data and expert experience. By combining newly added data and using similarity matching and time series analysis logic, the feature library is dynamically updated and fault mode tracking is achieved. Multi-model fusion analysis: A hierarchical model architecture is used to perform progressive analysis of core features. Through result fusion and parameter adaptive optimization logic, stable analysis results are output. Fault diagnosis and early warning generation: Parse and integrate the analysis results, combine the working condition adaptation logic to eliminate interference factors, output fault information and execute the hierarchical early warning strategy; Remaining lifetime estimation: Based on fault evolution characteristics and historical degradation data, the remaining lifetime and confidence interval of the equipment are output through condition assessment and error correction logic.

2. The motor system fault diagnosis method according to claim 1, characterized in that, The multi-dimensional features include time domain, frequency domain, and time-frequency domain features. The time domain features obtain feature parameters through signal amplitude statistical logic, the frequency domain features analyze the fundamental wave, harmonic components, and distortion features through signal frequency conversion logic, and the time-frequency domain features generate a time-frequency distribution map through multi-scale transformation logic to capture transient fault features. Then, through the feature contribution evaluation logic, the correlation coefficient between each feature and the fault type is calculated. The features with correlation above the threshold are sorted and filtered, and redundant features with correlation below the threshold are removed, forming a core feature set with timestamps and operating condition labels.

3. The motor system fault diagnosis method according to claim 2, characterized in that, The feature contribution evaluation logic is as follows: Principal component analysis logic is used to reduce the dimensionality of the extracted multidimensional features, calculate the variance contribution rate of each feature, set a screening threshold based on the cumulative variance contribution rate, retain features with a cumulative contribution rate higher than the threshold as core features, and remove redundant features with a contribution rate lower than the threshold.

4. The motor system fault diagnosis method according to claim 1, characterized in that, The construction and dynamic iteration of the fault feature library includes the following steps: During the initial construction, the feature vectors extracted from historical fault data are standardized, and feature templates are stored hierarchically according to fault type to establish a mapping relationship between fault type and feature vector; During dynamic iteration, the matching degree between newly added features and existing templates in the library is compared through similarity calculation logic. For templates with matching degree higher than a preset threshold, weighted updates are performed to integrate the newly added feature information. New features with matching degree lower than a preset threshold are identified as new fault modes and new templates are added. Templates with similarity higher than the threshold are merged periodically through clustering logic. At the same time, the evolution of fault features over time is tracked through time series analysis to adapt to the feature changes of progressive faults.

5. The motor system fault diagnosis method according to claim 1, characterized in that, The hierarchical model architecture processes faults in the order of anomaly detection, fault classification, and fault prediction. First, the anomaly detection model performs unsupervised learning to traverse and analyze the core feature set, delineates the normal feature distribution range, identifies abnormal data points that deviate from the normal feature distribution range and their corresponding feature subsets, and outputs the anomaly labeling results and anomaly confidence. The anomaly labeling results and corresponding feature subsets are fed into the fault classification model. The fault classification model, based on feature matching logic, compares the anomaly features with various fault templates in the fault feature library one by one, calculates the feature matching degree, and outputs the corresponding fault type, type confidence and fault association features based on the feature matching degree. Input the fault type, fault association features and historical time series features into the prediction model, mine the change pattern of features over time through time series association, predict the fault development trend and occurrence probability, and output the probability prediction results and time series trend curve. The confidence level, type confidence level, and probability prediction results of the three types of models are assigned weights through weighted voting. The weight values ​​are dynamically adjusted based on the historical processing accuracy of each model. The higher the accuracy, the greater the weight. The results of each model are multiplied by their corresponding weights and then summed to obtain the comprehensive analysis score and corresponding results. The system uses adaptive parameter tuning logic to monitor the stability of the comprehensive analysis score and its deviation from historical results in real time. If the deviation exceeds the allowable range, it automatically optimizes the core parameters of each model and the weighted voting weight ratio, and iterates the calculation until the stability of the comprehensive analysis result meets the standard. Finally, it outputs a fusion analysis result that includes the location of the anomaly, the type of the fault, the probability of occurrence, and the time series trend.

6. The motor system fault diagnosis method according to claim 1, characterized in that, The fault diagnosis and early warning generation includes the following steps: First, the abnormal location, fault type, occurrence probability, and time series trend results output by the multi-model fusion are analyzed to extract the core fault diagnosis information and calculate the corresponding confidence level. When the confidence level reaches the set threshold, the fault type, severity, and associated causes are initially output to form the initial fault diagnosis results. Then, pseudo-fault correction is performed through process matching analysis to form a corrected fault diagnosis result. Based on the corrected fault diagnosis result, the comprehensive risk level is divided according to the fault type, severity and probability of occurrence. Different risk levels correspond to differentiated early warning response logic.

7. The motor system fault diagnosis method according to claim 6, characterized in that, The pseudo-fault correction includes the following steps: Based on preset process standard parameters, the deviation range of real-time power parameters, load parameters and benchmark parameters is compared. Combined with the fault correlation characteristics output by multi-model fusion, it is determined whether there is a causal relationship between parameter deviation and fault characteristics. That is, if the parameter deviation exceeds the allowable range but there is no corresponding fault characteristic to support it, or if the deviation range is within the allowable range of process fluctuation, it is determined to be a false fault signal. The corresponding items in the initial fault diagnosis results are marked, removed or corrected. At the same time, the parameter deviation details and the basis for the false fault judgment are recorded to form the corrected fault diagnosis results.

8. The motor system fault diagnosis method according to claim 1, characterized in that, The remaining lifetime estimation includes the following steps: By integrating historical degradation data with real-time fault evolution characteristics, a device degradation model is constructed using a state estimation algorithm. The model parameters are initialized and real-time feature data is input to obtain the basic life prediction result. By analyzing the correlation and impact intensity among multiple faults through fault coupling identification logic, correction factors are introduced to adjust the basic prediction results. Uncertainty modeling is employed to sample and analyze measurement noise and model error, quantify the prediction error, and output the remaining lifetime range at different confidence levels.

9. The motor system fault diagnosis method according to claim 8, characterized in that, The state estimation algorithm employs a particle filtering algorithm, combining historical health status data and initial parameters of the equipment to generate a particle set that conforms to the actual initial state distribution of the equipment. This particle set is then substituted into a preset equipment degradation model, and the state prediction value corresponding to each particle is obtained through model iteration. Real-time fault evolution feature data is then called to compare the deviation between the predicted value of each particle and the real-time feature data, and the particle weights are updated according to the magnitude of the deviation, with smaller deviations resulting in higher weight proportions. Subsequently, a resampling and removal operation is performed on particles with weights below the preset value, retaining the effective particle set with weights above the preset value. Based on the statistical results of the effective particle set, the basic lifetime prediction result is output. During the fault coupling identification logic analysis, the fault association identification logic is first used to traverse the currently detected faults and historical fault records, distinguish the causal and parallel associations between faults, and quantify the impact intensity of each fault on the equipment degradation rate. Each fault is assigned a correction factor weight based on its impact intensity, with higher weights for greater impact intensity. The overall correction factor is obtained by weighted calculation. The baseline lifetime estimate is superimposed with a comprehensive correction factor to correct for prediction biases caused by multi-fault coupling, resulting in a preliminary corrected lifetime value. The uncertainty modeling adopts Monte Carlo simulation, which performs multiple random samplings on potential error sources such as measurement noise, model parameter errors, and data fluctuations. Each sampling result is substituted into the above state estimation and coupling correction process to obtain multiple sets of lifetime prediction data. Statistical analysis is performed on multiple sets of data to fit the error distribution pattern. Based on the distribution pattern, numerical intervals corresponding to different confidence levels are defined, and finally, the remaining lifetime interval with confidence level label is output.

10. A motor system fault diagnosis system, characterized in that, For implementing the motor system fault diagnosis method according to any one of claims 1 to 9, the system comprises: Data acquisition module: Used to collect core electrical data such as three-phase current and three-phase voltage of motor and auxiliary data such as ambient temperature, and has the functions of multi-source synchronous acquisition and timestamp alignment; Signal processing module: Communicates with the data acquisition module and is used to perform filtering and noise reduction, power frequency interference suppression, amplitude normalization and multi-scale noise reduction on the acquired electrical data, and output high-quality time-series data; Feature extraction module: Communicates with the signal processing module, integrates power feature analysis tools, and is used to extract time-domain, frequency-domain and multi-dimensional features. It filters core features through the feature contribution evaluation unit and generates a labeled multi-dimensional feature set. Fault Feature Library Module: Communicates with the feature extraction module, uses a database to store feature templates, and has functions such as template initialization, online updating, clustering and merging, and fault mode evolution tracking, which can dynamically adapt to changes in fault types; Algorithm model module: It communicates with the feature extraction module and the fault feature library module respectively, and deploys a three-layer algorithm model, a weighted voting fusion unit, and an adaptive parameter tuning unit to realize anomaly detection, fault classification and fault prediction. Diagnostic and early warning module: It communicates with the algorithm model module to parse the model fusion results, generate fault diagnosis information, execute multi-level early warning strategies, correct the results in combination with the process matching analysis unit, and output early warning notifications and maintenance suggestions. The lifetime prediction module is connected to the feature extraction module, the fault feature library module, and the diagnosis and early warning module. It uses a state estimation unit, a coupling effect modeling unit, and an uncertainty analysis unit to output the remaining lifetime estimate and confidence interval.

Images