General purpose generator electrical monitoring system with fault self-diagnosis function

Abstract

This application relates to the field of electrical monitoring technology and provides a general-purpose generator electrical monitoring system with fault self-diagnosis function. By judging the coherence coefficient and phase difference between the current harmonic components and the bearing vibration frequency band, and using a coupling mode library to mark fault coupling identifiers, it deeply explores the potential correlation between electrical and mechanical parameters, accurately identifies electrical and mechanical coupling faults, and significantly improves the ability to identify complex faults. At the same time, it dynamically adjusts the diagnostic threshold by combining load rate and winding temperature, enabling the system to better adapt to different generator operating conditions and improve the accuracy and reliability of diagnosis. Furthermore, it analyzes the fault causal chain and locates the fault source through a causal inference algorithm. Compared with existing fault tracing methods that lack a systematic approach, it can more comprehensively and deeply analyze the causes and processes of fault occurrence, quickly and accurately locate the fault source, and facilitate maintenance personnel to take timely and targeted measures to improve the safety and reliability of generator operation.

Description

Technical Field

[0001] This application relates to the field of electrical monitoring technology, and in particular to a general-purpose generator electrical monitoring system with fault self-diagnosis function. Background Technology

[0002] In modern power systems, general-purpose generators are core equipment, and their operational stability is crucial to the safety of the power grid. With the rapid development of the power industry, existing generator electrical monitoring technologies are also constantly evolving. Early monitoring systems mostly adopted single-parameter monitoring methods, such as monitoring only electrical parameters like current and voltage, or mechanical parameters like vibration and temperature. This method cannot fully reflect the operating status of the generator, makes it difficult to discover potential correlations between parameters, and results in low accuracy in fault diagnosis.

[0003] With advancements in sensor and data processing technologies, multi-parameter monitoring systems have gradually become mainstream. These systems can simultaneously acquire multiple operating parameters of a generator and analyze the data using simple threshold comparisons or empirical formulas. However, existing technologies still have many shortcomings. On the one hand, traditional systems lack in-depth analysis of the coupling relationship between electrical and mechanical parameters, making it difficult to accurately identify complex faults caused by this coupling. For example, when abnormal current harmonics and abnormal bearing vibration occur simultaneously, it is difficult to determine their correlation and the root cause of the fault. On the other hand, thresholds are mostly fixed values ​​or adjusted based on a single factor, failing to adapt to changes in generator operating conditions such as different load rates and winding temperatures, easily leading to misjudgments or missed diagnoses. Most existing technologies have not solved the problem of how to achieve fault isolation and fault tracing based on the coupling analysis of motor electrical and mechanical parameters. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides a general-purpose generator electrical monitoring system with fault self-diagnosis function. The system includes: a coupling sensing module, a fault diagnosis module, and a fault tracing module.

[0005] The coupling sensing module is used to acquire the current harmonic amplitude, bearing vibration velocity, load rate and winding temperature in real time, and simultaneously extract the current harmonic components and bearing vibration frequency band. In response to the judgment of the coherence coefficient and phase difference between the current harmonic components and the bearing vibration frequency band, the fault coupling identifier is marked by the coupling mode library.

[0006] The fault diagnosis module is used to extract the diagnosis threshold. After receiving the fault coupling identifier, it combines the load rate and winding temperature to dynamically adjust the diagnosis threshold. During the adjustment process, it filters the current harmonic amplitude and bearing vibration velocity that are greater than the diagnosis threshold. It separates electrical faults from mechanical faults based on a dual-channel network. At the same time, it reconstructs the target current harmonics and target bearing vibrations through the dual-channel network, and generates a fault contribution matrix by comparing them with the current harmonic amplitude and bearing vibration velocity respectively.

[0007] The fault tracing module is used to receive the fault contribution matrix, analyze the fault causal chain and locate the fault source through the causal inference algorithm, and update the diagnostic threshold and coupling mode library in response.

[0008] As an optional implementation, the marking logic of the fault coupling identifier includes:

[0009] Configure a coherence coefficient threshold, filter associated frequency points with a coherence coefficient greater than the coherence coefficient threshold, and calculate the phase difference between the current harmonic component and the bearing vibration frequency band at the associated frequency points. Then, use a clustering algorithm to divide the frequency into multiple phase difference intervals based on the magnitude of the phase difference, with each phase difference interval representing a fault coupling mode.

[0010] Establish a coupling mode library, map phase difference intervals to fault coupling modes, and assign fault coupling identifiers to each fault coupling mode to form an identifier mapping table;

[0011] Based on the phase difference interval to which the current phase difference belongs, find the fault coupling identifier in the identifier mapping table and mark the associated frequency point.

[0012] As an optional implementation, the coherence coefficient determination sub-logic includes:

[0013] Real-time acquisition of current harmonic amplitude and bearing vibration velocity, and simultaneous extraction of current harmonic components and bearing vibration frequency band;

[0014] The current harmonic components and bearing vibration frequency bands are divided into data segments according to time intervals. The dimensional features of the current harmonic components and bearing vibration frequency bands in each data segment are extracted by fast Fourier transform, and the dimensional features are reduced by principal component analysis.

[0015] Based on the dimensionality features after dimensionality reduction, the coherence coefficients of the current harmonic components and the bearing vibration frequency band at each frequency point are calculated to form a coherence coefficient sequence. Then, outlier detection is performed on the coherence coefficient sequence using a clustering algorithm.

[0016] As an optional implementation, the logic for generating the fault contribution matrix includes:

[0017] The contribution weights of electrical and mechanical fault features to fault diagnosis are determined based on the attention mechanism.

[0018] The difference between the target current harmonic and the current harmonic amplitude, and the difference between the target bearing vibration and the bearing vibration velocity, are combined with contribution weights to determine the fault contribution of electrical fault characteristics and mechanical fault characteristics to fault diagnosis.

[0019] The fault contribution values ​​of electrical fault features and mechanical fault features are arranged in order to generate a fault contribution matrix.

[0020] As an optional implementation, the diagnostic threshold adjustment sub-logic includes:

[0021] Historical operating data of the generator were obtained, and the data distribution of current harmonic amplitude and bearing vibration velocity under normal operating conditions were statistically analyzed to serve as the initial diagnostic threshold.

[0022] After receiving the fault coupling identifier, the fault coupling mode corresponding to the fault coupling identifier is obtained from the coupling mode library, and the changing trends of current harmonic amplitude and bearing vibration velocity under the fault coupling mode are statistically analyzed to determine the adjustment weight of the initial diagnostic threshold.

[0023] The load rate and winding temperature are acquired in real time, and the influence factors of load rate and winding temperature on current harmonic amplitude and bearing vibration velocity are analyzed by support vector regression algorithm.

[0024] The adjustment coefficient is obtained by weighting and summing the adjustment weights of the initial diagnostic threshold and the influencing factors, and then combined with the initial diagnostic threshold to adjust the diagnostic threshold.

[0025] As an optional implementation, the separation sub-logic for electrical and mechanical faults includes:

[0026] Data points with current harmonic amplitudes greater than the diagnostic threshold are selected and marked as electrical candidate fault points, and data points with bearing vibration velocity greater than the diagnostic threshold are selected and marked as mechanical candidate fault points;

[0027] A dual-channel network consisting of physical and data channels is constructed. The physical channel enhances the features of electrical and mechanical candidate fault points to obtain electrical and mechanical fault features.

[0028] The data channel extracts the correlation features between the current harmonic amplitude and the bearing vibration velocity under the fault coupling mode from the coupling mode library based on the fault coupling identifier, and performs real-time correction of the correlation features in combination with the load rate and winding temperature to generate fault feature templates.

[0029] The cosine similarity between electrical fault features, mechanical fault features and fault feature templates is calculated separately to determine the probability of electrical faults and the probability of mechanical faults, and a decision threshold is configured to separate electrical faults and mechanical faults.

[0030] As an optional implementation, the reconstruction sub-logic for the target current harmonics and target bearing vibration includes:

[0031] The generator's current harmonic amplitude and bearing vibration velocity are acquired in real time through the physical channel;

[0032] The historical current harmonic amplitude and historical bearing vibration velocity of the generator under normal operating conditions are retrieved from the data channel and classified and stored according to the combination of load rate and winding temperature to form a feature mapping relationship.

[0033] The current harmonic amplitude and bearing vibration velocity obtained from the physical channel are matched in the feature mapping relationship with the load rate and winding temperature to screen normal operating condition features, and the normal operating condition features are fed back to the physical channel to reconstruct the target current harmonics and target bearing vibration.

[0034] The reconstructed target current harmonics and target bearing vibrations are transmitted to the physical channel for error verification to determine whether to re-screen normal operating condition characteristics and reconstruct the target current harmonics and target bearing vibrations.

[0035] As an optional implementation, the logic for updating the diagnostic threshold includes:

[0036] Analyze the changing trends of current harmonic amplitude and bearing vibration velocity at the fault source, and determine the threshold offset of the changes in current harmonic amplitude and bearing vibration velocity caused by the fault source.

[0037] The diagnostic threshold is updated by using a reinforcement learning algorithm with the false positive rate and false negative rate as reward functions, combined with the threshold offset.

[0038] The effectiveness of the updated diagnostic thresholds was evaluated using cross-validation to determine whether further updates to the diagnostic thresholds were necessary.

[0039] As an optional implementation, the fault source location sub-logic includes:

[0040] The fault contribution matrix is ​​received, and the causal relationship between current harmonic amplitude, bearing vibration velocity, load rate and winding temperature is automatically searched through the causal inference algorithm to generate a causal network.

[0041] The causal network is transformed into a Bayesian network, where each node represents a fault feature and each edge represents a causal relationship. The prior probability distribution of each node is set according to the fault contribution matrix.

[0042] By using Bayesian inference to determine the posterior probability of each fault feature as a fault source, potential fault sources are screened out, and the confidence level of each potential fault source is determined to locate the fault source.

[0043] As an optional implementation, the update logic of the coupling mode library includes:

[0044] The DBSCAN algorithm is used to cluster abnormal data with unmatched fault coupling patterns, and the clustering regions of new fault coupling patterns are identified based on data density.

[0045] The mapping relationship between phase difference interval division and fault coupling identifier in the new fault coupling mode is dynamically adjusted by using the gradient descent algorithm to update the coupling mode library.

[0046] Compared with existing technologies, the beneficial effects of this application are as follows: By judging the coherence coefficient and phase difference between the current harmonic components and the bearing vibration frequency band, and using the coupling mode library to mark fault coupling identifiers, it is possible to deeply explore the potential correlation between electrical and mechanical parameters, accurately identify electrical and mechanical coupling faults, and significantly improve the identification capability of complex faults compared with traditional single-parameter or simple correlation analysis methods, avoiding missed faults; at the same time, by dynamically adjusting the diagnostic threshold based on load rate and winding temperature, it overcomes the limitations of traditional fixed threshold or single-factor adjustment, enabling the system to better adapt to different generator operating conditions, effectively reducing the false judgment rate, and improving the accuracy and reliability of diagnosis; and by analyzing the fault causal chain and locating the fault source through causal inference algorithms, compared with existing fault tracing methods that lack systematicity, it can more comprehensively and deeply analyze the causes and processes of fault occurrence, quickly and accurately locate the fault source, facilitate maintenance personnel to take timely and targeted measures, shorten fault handling time, reduce downtime losses, and improve the safety and reliability of generator operation. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments 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. Wherein:

[0048] Figure 1 This is a system flowchart of a general-purpose generator electrical monitoring system with fault self-diagnosis function provided in an embodiment of this application;

[0049] Figure 2 This is a logic diagram of the fault coupling identifier for a general generator electrical monitoring system with fault self-diagnosis function provided in an embodiment of this application.

[0050] Figure 3 The logic diagram for separating electrical and mechanical faults in a general-purpose generator electrical monitoring system with self-diagnosis function provided in this application embodiment;

[0051] Figure 4 This is a logic diagram for locating the fault source of a general-purpose generator electrical monitoring system with self-diagnosis function provided in an embodiment of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0053] Example

[0054] like Figure 1 The diagram shown is a system flowchart of a general-purpose generator electrical monitoring system with fault self-diagnosis function provided in this application embodiment. The system includes a coupling sensing module, a fault diagnosis module, and a fault tracing module.

[0055] The coupling sensing module is used to acquire current harmonic amplitude, bearing vibration velocity, load rate and winding temperature in real time, and simultaneously extract current harmonic components and bearing vibration frequency band. In response, it judges the coherence coefficient and phase difference between the current harmonic components and the bearing vibration frequency band, so as to mark the fault coupling identifier through the coupling mode library.

[0056] Furthermore, the decision logic for the coherence coefficient includes:

[0057] Real-time acquisition of current harmonic amplitude and bearing vibration velocity, and simultaneous extraction of current harmonic components and bearing vibration frequency band;

[0058] The current harmonic components and bearing vibration frequency bands are divided into data segments according to time intervals. The dimensional features of the current harmonic components and bearing vibration frequency bands in each data segment are extracted by fast Fourier transform, and the dimensional features are reduced by principal component analysis.

[0059] Based on the dimensionality features after dimensionality reduction, the coherence coefficients of the current harmonic components and the bearing vibration frequency band at each frequency point are calculated to form a coherence coefficient sequence. Then, outlier detection is performed on the coherence coefficient sequence using a clustering algorithm.

[0060] During generator operation, faults in the electrical and mechanical systems manifest as abnormal coupling between current harmonics and bearing vibration. Real-time synchronous acquisition of both types of signals is necessary to capture the early time-frequency domain correlation characteristics of the fault. Electrical and vibration signals are acquired using current and vibration sensors to monitor current harmonic amplitude and bearing vibration velocity. The sampling times of both sensors are uniformly calibrated to ensure no timestamp errors exist. The current signal is then bandpass filtered to retain fault-related harmonic frequency bands, including integer multiples of the fundamental frequency. Wavelet packet decomposition is performed on the vibration signal to extract the bearing fault-sensitive frequency bands. Frequency domain separation is achieved using a filter bank, thereby extracting the current harmonic components and bearing vibration frequency bands. This ensures strict time-domain synchronization of the electrical and vibration signals and highlights fault characteristics in the frequency domain, providing time-consistent and characteristically clear raw data for subsequent coherence analysis. The synchronized signal provides standardized input for data segmentation and feature dimensionality reduction, ensuring that subsequent feature extraction accurately reflects the coupling relationship between signals.

[0061] The original signal has temporal continuity, resulting in a large amount of data and correlations between multidimensional features. Therefore, segmentation, discretization, and dimensionality reduction are necessary to eliminate redundant information and improve computational efficiency. A sliding window mechanism is used to segment the signal into multiple data segments. The window length is set according to the time-varying characteristics of the fault features, and the window overlap rate is set to 50% to ensure data continuity. A fast Fourier transform is used to perform a time-frequency transformation on each data segment to extract the amplitude characteristics of the current harmonic components at each frequency point and the energy distribution characteristics of the bearing vibration frequency band, thus extracting the dimensional features. Principal component analysis is then used to calculate the covariance matrix of the dimensional features. Principal components are selected based on the cumulative variance contribution rate to compress high-dimensional features into a low-dimensional space, eliminating linear correlations between dimensional features. This transforms the continuous signal into discrete data segments, reducing computational complexity. Dimensionality reduction preserves key fault-sensitive features and avoids the interference of the curse of dimensionality on subsequent analysis.

[0062] The coherence coefficient quantifies the coupling strength between current and vibration signals at a specific frequency point, while the abnormal coherence coefficient indicates the coupling state of early faults. Atypical coupling modes are identified through clustering algorithms. Based on the dimensionality-reduced feature vectors, the cross-power spectral density at each frequency point is calculated using the Welch method. The coherence coefficient is obtained through the ratio of the cross-power spectrum to the auto-power spectrum, forming a sequence reflecting the coupling strength at each frequency point. This coherence coefficient sequence is then analyzed using the DBSCAN clustering algorithm. By setting the neighborhood radius and minimum sample size parameters, data points are divided into core clusters and outliers. Data points within the core cluster satisfy the high-density distribution condition, representing the coupling mode of normal operation or typical faults. Data points deviating from the core cluster are identified as outliers, corresponding to coupling characteristics of unsteady states or early faults. This enables quantitative analysis of the coupling strength between current and vibration signals, identifying anomalous coupling characteristics of potential faults through a data-driven approach, providing a basis for fault mode classification. The selected anomalous coherence coefficient frequency points and corresponding characteristics provide a candidate set for fault coupling identification, narrowing the scope of subsequent phase difference analysis.

[0063] Specifically, such as Figure 2 As shown, the marking logic for fault coupling identification includes:

[0064] Configure a coherence coefficient threshold, filter associated frequency points with a coherence coefficient greater than the coherence coefficient threshold, and calculate the phase difference between the current harmonic component and the bearing vibration frequency band at the associated frequency points. Then, use a clustering algorithm to divide the frequency into multiple phase difference intervals based on the magnitude of the phase difference, with each phase difference interval representing a fault coupling mode.

[0065] Establish a coupling mode library, map phase difference intervals to fault coupling modes, and assign fault coupling identifiers to each fault coupling mode to form an identifier mapping table;

[0066] Based on the phase difference interval to which the current phase difference belongs, find the fault coupling identifier in the identifier mapping table and mark the associated frequency point.

[0067] The coherence coefficient characterizes the signal coupling strength, while the phase difference reflects the phase relationship between signals. The combination of these two factors uniquely identifies the fault coupling mode. This requires threshold screening and clustering to divide typical phase difference intervals. A coherence coefficient threshold is determined through historical fault data statistics. Frequency points with coherence coefficients greater than the threshold are selected as associated frequency points to ensure the analysis focuses on strongly coupled characteristic frequencies. The phase difference between the current signal and the vibration signal is then calculated for these associated frequency points. Instantaneous phase is obtained using a phase correlation algorithm, and time averaging is used to eliminate random noise. Next, the phase differences of the associated frequency points are classified using the K-means clustering algorithm. The number of clusters is determined based on prior knowledge of fault coupling modes. The continuous phase difference space is divided into several discrete phase difference intervals based on the magnitude of the phase difference, with each interval corresponding to a typical fault coupling mode. Threshold screening focuses on strongly coupled features, and phase difference clustering discretizes the continuous feature space, reducing the complexity of fault coupling mode identification and laying the foundation for standardized descriptions of fault coupling modes. The divided phase difference intervals provide classification units for establishing a coupling mode library, ensuring a one-to-one correspondence between fault coupling modes and feature parameters.

[0068] A standardized mapping relationship needs to be established between abstract phase difference intervals and fault coupling modes. This allows for rapid retrieval and invocation of modes through digital identifiers. A database structure is constructed to store descriptions of phase difference intervals and fault coupling modes, along with their corresponding fault coupling identifiers. A key-value pair mapping relationship is used, and each fault coupling mode is assigned a unique identifier. These identifiers follow an encoding rule—a letter prefix followed by a numerical sequence—ensuring the uniqueness and scalability of the fault coupling identifiers. A data dictionary maintains the correspondence between fault coupling identifiers and fault coupling modes, supporting dynamic updates. This forms a standardized coupling mode library, enabling rapid mapping from signal characteristics to fault coupling modes. This provides prior knowledge support for the fault diagnosis module, improving diagnostic efficiency. The identifier mapping table provides a query basis for real-time phase difference matching and is a key data interface connecting signal characteristics and fault diagnosis.

[0069] During generator operation, current signal characteristics need to be converted into fault coupling identifiers in real time to provide direct input for subsequent fault diagnosis and achieve real-time early warning of faults. For the phase difference calculated in real time, the phase difference interval to which it belongs is determined by binary search, and the corresponding fault coupling identifier is queried in the identifier mapping table. At the same time, the fault coupling identifier is bound to the associated frequency point, and the marking result is transmitted to the fault diagnosis module in real time through the data bus or message queue to ensure the timeliness of identifier transmission. This realizes the real-time digital marking of fault coupling features, providing the fault diagnosis module with tagged feature input. The marked fault coupling identifier directly drives the threshold adjustment and fault separation of the fault diagnosis module, and is a key interaction node between the perception layer and the diagnosis layer.

[0070] The fault diagnosis module is used to extract diagnostic thresholds. After receiving the fault coupling identifier, it combines the load rate and winding temperature to dynamically adjust the diagnostic thresholds. During the adjustment process, it filters out current harmonic amplitudes and bearing vibration velocities that are greater than the diagnostic thresholds. It separates electrical faults from mechanical faults based on a dual-channel network. At the same time, it reconstructs the target current harmonics and target bearing vibrations through the dual-channel network, and generates a fault contribution matrix by comparing them with the current harmonic amplitudes and bearing vibration velocities, respectively.

[0071] Furthermore, the adjustment sub-logic for the diagnostic threshold includes:

[0072] Historical operating data of the generator were obtained, and the data distribution of current harmonic amplitude and bearing vibration velocity under normal operating conditions were statistically analyzed to serve as the initial diagnostic threshold.

[0073] After receiving the fault coupling identifier, the fault coupling mode corresponding to the fault coupling identifier is obtained from the coupling mode library, and the changing trends of current harmonic amplitude and bearing vibration velocity under the fault coupling mode are statistically analyzed to determine the adjustment weight of the initial diagnostic threshold.

[0074] The load rate and winding temperature are acquired in real time, and the influence factors of load rate and winding temperature on current harmonic amplitude and bearing vibration velocity are analyzed by support vector regression algorithm.

[0075] The adjustment coefficient is obtained by weighting and summing the adjustment weights of the initial diagnostic threshold and the influencing factors, and then combined with the initial diagnostic threshold to adjust the diagnostic threshold.

[0076] To distinguish between normal operating conditions and fault conditions, an initial diagnostic threshold needs to be established based on the generator's historical operating data. This initial diagnostic threshold serves as the basis for subsequent dynamic adjustments, ensuring the statistical reliability of the initial judgment. Operating data of the generator under normal operating conditions is retrieved from the database, including current harmonic amplitudes and bearing vibration velocities under different load rates and winding temperatures. The current harmonic amplitudes and bearing vibration velocities are cleaned to remove outliers caused by sensor anomalies or short-term interference. The cleaned data is then analyzed using probabilistic statistical methods, including fitting probability distribution curves to determine the initial diagnostic threshold based on confidence principles. This ensures that the vast majority of data points during normal operation fall within the range of the initial diagnostic threshold. This establishes an initial diagnostic threshold that conforms to the actual operating characteristics of the generator, providing a reasonable reference for subsequent dynamic adjustments and reducing the probability of misjudgment in the initial state. The initial diagnostic threshold provides benchmark data for the calculation of subsequent adjustment weights and the analysis of operating condition influence factors, ensuring that the dynamic adjustment process has a clear starting point and reference standard.

[0077] Different fault coupling modes have varying degrees of impact on current harmonic amplitude and bearing vibration velocity. Therefore, the adjustment weight of the initial diagnostic threshold needs to be determined based on historical fault data statistics of the fault coupling modes. This allows the initial diagnostic threshold to adaptively shift according to the characteristic changes of specific fault coupling modes. When the system receives a fault coupling identifier, it retrieves the corresponding historical fault data from the coupling mode library, including records of changes in current harmonic amplitude and bearing vibration velocity when similar faults occur. Time-series analysis is performed on this historical fault data to extract characteristic parameters, including amplitude change rate and trend slope. The average change amplitude of current harmonic amplitude and bearing vibration velocity under this fault coupling mode is calculated using statistical methods and compared with the initial diagnostic threshold to determine the corresponding adjustment ratio, which serves as the adjustment weight. This allows the initial diagnostic threshold to be adjusted specifically for the characteristic changes of particular fault coupling modes, avoiding false alarms or missed alarms caused by a uniform threshold. It also improves the sensitivity of the initial diagnostic threshold to the target fault. The adjustment weight provides parameters for the fault coupling mode dimension in subsequent threshold calculations based on combined influence factors, ensuring that the diagnostic threshold adjustment considers the combined effects of fault type and real-time operating conditions.

[0078] The generator's load rate and winding temperature directly affect the normal fluctuation range of electrical and vibration signals. It is necessary to quantify the influence of operating condition parameters on signal characteristics to achieve adaptive adjustment of diagnostic thresholds according to operating conditions, avoiding misdiagnosis caused by fluctuations in operating conditions. By acquiring operating condition parameters such as load rate and winding temperature in real time through sensors, and simultaneously recording the current harmonic amplitude and bearing vibration velocity at corresponding times, a historical dataset containing operating condition parameters and signal characteristics is constructed. This historical dataset is then trained using a support vector machine to establish a mapping relationship between operating condition parameters and current harmonic amplitude and bearing vibration velocity. The mapping relationship outputs the influence factor of each operating condition parameter on current harmonic amplitude and bearing vibration velocity. This influence factor reflects the quantitative impact of operating condition changes on current harmonic amplitude and bearing vibration velocity. This clarifies the influence law of operating conditions on current harmonic amplitude and bearing vibration velocity, enabling the diagnostic threshold to be dynamically corrected according to changes in operating conditions. Specifically, under high load, the threshold is appropriately increased based on the influence factor to avoid misjudging signal fluctuations under normal operating conditions as faults.

[0079] By combining the adjustment weights and influencing factors of the diagnostic threshold, a final adjustment coefficient is obtained through weighted calculation, enabling real-time dynamic updating of the diagnostic threshold and allowing it to adapt to real-time changes in generator operating status. The adjustment coefficient is obtained by weighted summation of the initial diagnostic threshold's adjustment weights and influencing factors, and then combined with the initial diagnostic threshold to adjust it. This adjustment employs appropriate transformation methods, including linear transformations or nonlinear transformations based on historical data. This allows the diagnostic threshold to simultaneously respond to the characteristics of fault coupling modes and changes in operating conditions, improving its dynamic adaptability and ensuring accurate fault detection under different load and temperature conditions. The updated diagnostic threshold is used to screen for abnormal features in current harmonic amplitude and bearing vibration velocity, providing an accurate judgment benchmark for subsequent separation of electrical and mechanical faults.

[0080] Furthermore, such as Figure 3 As shown, the logic for separating electrical and mechanical faults includes:

[0081] Data points with current harmonic amplitudes greater than the diagnostic threshold are selected and marked as electrical candidate fault points, and data points with bearing vibration velocity greater than the diagnostic threshold are selected and marked as mechanical candidate fault points;

[0082] A dual-channel network consisting of physical and data channels is constructed. The physical channel enhances the features of electrical and mechanical candidate fault points to obtain electrical and mechanical fault features.

[0083] The data channel extracts the correlation features between the current harmonic amplitude and the bearing vibration velocity under the fault coupling mode from the coupling mode library based on the fault coupling identifier, and performs real-time correction of the correlation features in combination with the load rate and winding temperature to generate fault feature templates.

[0084] Calculate the cosine similarity between electrical fault features and mechanical fault features and fault feature templates to determine the electrical fault probability and mechanical fault probability, and configure decision thresholds to separate electrical faults and mechanical faults.

[0085] Based on dynamically updated diagnostic thresholds, anomalies in current harmonic amplitude and bearing vibration velocity are identified, providing candidate objects for subsequent fault isolation, thereby narrowing the scope of fault feature analysis and improving processing efficiency. The real-time acquired current harmonic amplitude and bearing vibration velocity are compared with the current diagnostic thresholds, filtering out data points exceeding the thresholds. Data points with current harmonic amplitudes exceeding the thresholds are marked as electrical candidate fault points, and data points with bearing vibration velocities exceeding the thresholds are marked as mechanical candidate fault points. Each candidate point is appended with contextual information such as timestamp, load rate, and winding temperature, forming a complete fault feature data record. Threshold filtering quickly locates abnormal features in the signal, reducing the computational load of subsequent feature analysis, while retaining the operating conditions at the time of the fault, providing auxiliary judgment criteria for fault classification.

[0086] The original features of candidate fault points contain noise or redundant information, requiring feature enhancement through a physical channel to extract key features that effectively distinguish electrical and mechanical faults, thereby improving the accuracy of fault separation. A physical channel incorporating a deep learning network is constructed, and different feature extraction paths are designed for electrical and mechanical candidate fault points. For electrical candidate fault points, parameters such as harmonic distortion rate and characteristic frequency amplitude ratio are extracted using frequency domain analysis methods, combined with time-frequency analysis to capture transient harmonic anomalies. For mechanical candidate fault points, the vibration signal is decomposed in the time-frequency domain to extract the energy distribution of bearing fault characteristic frequencies, and envelope analysis is used to enhance early fault features, thus obtaining both electrical and mechanical fault features. This process extracts highly discriminative fault features from candidate fault points, suppresses interference from noise and irrelevant features, and enhances the separability of electrical and mechanical fault features in the feature space, providing high-quality feature input for subsequent classification.

[0087] By combining prior knowledge of fault coupling modes with real-time operating conditions, a fault feature template reflecting the expected faults is generated as a reference standard for fault separation, ensuring that fault classification has a clear comparative benchmark. The data channel retrieves historical correlation features of the corresponding fault coupling mode from the coupling mode library based on the received fault coupling identifier, constructs a fault feature template, and simultaneously acquires load rate and winding temperature in real time. Interpolation methods are used to correct the fault feature template, compensating for the impact of current operating condition changes on fault features, so that the fault feature template can reflect the expected fault features under the current operating conditions. This generates a fault feature template that matches the current fault coupling mode and operating conditions, providing a dynamic reference benchmark for fault separation, avoiding classification bias caused by differences in operating conditions, and improving the accuracy and adaptability of fault separation.

[0088] By calculating the similarity between fault features and fault feature templates, the probability of electrical and mechanical faults is quantified. Combined with a preset decision threshold, automatic fault separation is achieved, providing clear fault type information for subsequent fault tracing. Using a cosine similarity calculation method, the similarity between electrical and mechanical fault features and fault feature templates is calculated separately. The similarity results are converted into fault probability values. A preset decision threshold is used: when the probability of an electrical fault is greater than that of a mechanical fault and also greater than the decision threshold, it is determined to be an electrical fault; otherwise, it is determined to be a mechanical fault. If the probability of both faults is less than the decision threshold, it is determined to be an unknown fault mode. This achieves quantitative separation of electrical and mechanical faults, reduces manual intervention, improves the automation level and classification accuracy of fault diagnosis, provides clear fault type input for the fault tracing module, narrows the tracing scope, and the separated fault type results drive the generation of the fault contribution matrix, providing key fault type clues for fault tracing and guiding the causal analysis direction during the tracing process.

[0089] Furthermore, the reconstruction sub-logic for the target current harmonics and target bearing vibration includes:

[0090] The generator's current harmonic amplitude and bearing vibration velocity are acquired in real time through the physical channel;

[0091] The historical current harmonic amplitude and historical bearing vibration velocity of the generator under normal operating conditions are retrieved from the data channel and classified and stored according to the combination of load rate and winding temperature to form a feature mapping relationship.

[0092] The current harmonic amplitude and bearing vibration velocity obtained from the physical channel are matched in the feature mapping relationship with the load rate and winding temperature to screen normal operating condition features, and the normal operating condition features are fed back to the physical channel to reconstruct the target current harmonics and target bearing vibration.

[0093] The reconstructed target current harmonics and target bearing vibrations are transmitted to the physical channel for error verification to determine whether to re-screen normal operating condition characteristics and reconstruct the target current harmonics and target bearing vibrations.

[0094] To ensure the reconstruction process closely reflects the actual operating status of the generator, it is necessary to obtain the current harmonic amplitude and bearing vibration velocity at the current time as the basis for reconstruction. Only based on real-time data can the operating characteristics of the generator under specific conditions be accurately reflected, providing a reliable basis for subsequent target data reconstruction. The current harmonic amplitude and bearing vibration velocity are obtained by current sensors and vibration sensors, respectively. The obtained current harmonic amplitude and bearing vibration velocity are preprocessed, including low-pass filtering to remove high-frequency noise and using the moving average method to eliminate random interference. Based on the preprocessed current harmonic amplitude and bearing vibration velocity, a time-series data stream including multiple frequency points is formed. This obtains high-quality real-time data, ensuring the authenticity and integrity of the data, providing a reliable data source for the subsequent reconstruction process, and reducing reconstruction errors caused by data quality issues.

[0095] To accurately identify normal operating conditions under complex operating conditions, it is necessary to establish a feature mapping relationship between different combinations of operating conditions and historical normal characteristics, serving as a reference standard for real-time data matching. Normal operating data of generators under various typical operating conditions is retrieved from historical databases, including historical current harmonic amplitudes and historical bearing vibration velocities under different load rate and winding temperature combinations. Then, feature extraction is performed on the historical data, calculating the statistical characteristics of current harmonic amplitudes and bearing vibration velocities at each frequency point, including mean, standard deviation, and skewness, forming a multi-dimensional feature vector. This multi-dimensional feature vector is organized based on a KD-tree index structure, using load rate and winding temperature as index keys to construct a feature mapping relationship between combined operating conditions and historical normal characteristics, enabling rapid operating condition matching queries. This results in a structured feature mapping relationship, facilitating rapid retrieval and matching of normal characteristics under the current combined operating conditions, improving reconstruction efficiency and accuracy, and providing a reliable comparison benchmark for real-time data.

[0096] By matching real-time data with feature mapping relationships, normal operating condition features are selected. These features are then used to reconstruct the target current harmonics and target bearing vibrations, thus separating the abnormal components caused by the fault. The real-time current harmonic amplitude and bearing vibration velocity obtained from the physical channel are combined with the current load rate and winding temperature. A nearest neighbor search is performed in the KD-tree index of the data channel to find one or more sets of normal operating condition features most similar to the current operating condition. The similarity score between the current operating condition and each set of normal operating condition features is calculated. A weighted fusion algorithm is used to assign weights based on the similarity score, and the selected normal operating condition features are linearly combined. An inverse Fourier transform is then performed on the fused normal operating condition features to reconstruct the target current harmonics and target bearing vibrations, ensuring that the target current harmonics and target bearing vibrations reflect the normal operating state under the current condition in both the frequency and time domains. This process separates normal operating characteristics from complex operating conditions and reconstructs the target current harmonics and target bearing vibrations reflecting the current condition, providing a clear comparative benchmark for subsequent fault contribution analysis and facilitating accurate identification of fault-induced anomalies.

[0097] The accuracy of the reconstructed target current harmonics and target bearing vibration needs to be verified to ensure that they accurately reflect the normal operating condition. If the error exceeds the acceptable range, the normal operating condition features need to be re-screened and iteratively reconstructed. Error indices between the reconstructed target current harmonics and target bearing vibration and the real-time measured values ​​need to be calculated, including time-domain error and frequency-domain error. The time-domain error includes mean square error, and the frequency-domain error includes spectral similarity. The error indices are compared with a preset error threshold. If the error exceeds the threshold, the screening conditions are adjusted, including expanding the operating condition matching range, increasing the number of candidate normal operating conditions, and changing the feature fusion weight allocation. Then, based on the adjustment... The adjusted screening criteria are used to re-screen normal operating condition features and reconstruct the target current harmonics and target bearing vibration. The error verification process is repeated until the error meets the requirements or the maximum number of iterations is reached. A closed-loop verification mechanism is used to ensure the reliability of the reconstructed waveform, improve the accuracy of fault feature extraction, avoid misjudgment of faults due to reconstruction errors, and ensure that the reconstruction results can accurately reflect the normal operating status. The verified reconstructed target current harmonics and target bearing vibration provide stable and reliable input for the generation of the fault contribution matrix, ensuring the accuracy of subsequent fault tracing. If the reconstruction waveform error is large, it will lead to distortion of the fault contribution matrix and affect the fault source location results.

[0098] Specifically, the logic for generating the fault contribution matrix includes:

[0099] The contribution weights of electrical and mechanical fault features to fault diagnosis are determined based on the attention mechanism.

[0100] The difference between the target current harmonic and the current harmonic amplitude, and the difference between the target bearing vibration and the bearing vibration velocity, are combined with contribution weights to determine the fault contribution of electrical fault characteristics and mechanical fault characteristics to fault diagnosis.

[0101] The fault contribution values ​​of electrical fault features and mechanical fault features are arranged in order to generate a fault contribution matrix.

[0102] Different fault features contribute differently to the final diagnostic results. An attention mechanism is needed to quantify the importance of each fault feature, providing a weighting basis for accurately calculating fault contribution and avoiding diagnostic bias caused by treating all fault features equally. A neural network based on the attention mechanism is constructed, taking electrical and mechanical fault features as input. The self-attention mechanism calculates the contribution weights of these features to fault diagnosis. During the training phase, historical fault data is used as samples, and known fault type labels are used as supervision signals. Backpropagation is employed to optimize the contribution weights assigned by the attention mechanism, enabling the neural network to automatically identify the fault features most discriminative for fault diagnosis. A regularization term is introduced to constrain the contribution weights, preventing overfitting and improving the generalization ability of the neural network. This ensures accurate allocation of contribution weights for fault features across different fault scenarios, thereby achieving adaptive weighting of fault features, highlighting the role of key fault features, suppressing interference from noise features, improving the accuracy and robustness of fault diagnosis, and making the diagnostic results more focused on the true fault features.

[0103] By comparing the differences between the reconstructed target current harmonics and target bearing vibration and the real-time measured values, and combining the contribution weights of fault features, the contribution of each fault feature to the overall fault is quantified, providing a quantitative basis for fault tracing and enabling accurate identification of the main fault causes. The difference between the target current harmonics and their amplitudes, and the difference between the target bearing vibration and its velocity, are calculated to obtain the residual vector of the original fault. The residual vector is then multiplied by the contribution weights determined based on an attention mechanism to obtain the fault contribution degree of each fault feature, highlighting the influence of important fault features. The fault contribution degrees are then normalized to a sum of 1, facilitating a direct comparison of the relative importance of different fault features and forming a standardized fault contribution degree vector. This quantifies the differences in fault features into specific contribution values, clearly reflecting the influence of each fault feature on the overall fault, providing a quantifiable basis for fault source location, and making the fault tracing process more scientific and objective.

[0104] The fault contribution of each fault feature is arranged according to a specific logic to form a structured matrix representation, which intuitively shows the correlation and relative importance between fault features, providing a systematic input for fault tracing and facilitating analysis and decision-making. The contribution of electrical and mechanical fault features is arranged according to preset rules to form a two-dimensional matrix structure. Rows of the matrix represent different types of fault features, and columns represent different fault dimensions, including time, frequency, and spatial dimensions. Matrix elements are the fault contribution of the corresponding fault feature in the corresponding dimension. Matrix operations can analyze the correlation and transmission relationships between different fault features, and determine the fault contribution. The contribution matrix is ​​visualized, including generating a heatmap and using color depth to represent the magnitude of contribution, enhancing the intuitive expression of fault feature relationships and facilitating manual analysis and understanding. This results in a structured and visualized fault contribution matrix that clearly shows the relative importance and interrelationships of each fault feature, providing systematic and hierarchical input information for the fault tracing module. This helps diagnostic personnel quickly locate the fault source. As input to fault tracing, the structure and content of the fault contribution matrix directly affect the efficiency and accuracy of the causal inference algorithm and determine the effectiveness of fault source localization. If the fault contribution matrix is ​​poorly designed, it will lead to a complex causal inference process and inaccurate results.

[0105] The fault tracing module receives the fault contribution matrix to analyze the fault causal chain and locate the fault source through a causal inference algorithm, and updates the diagnostic threshold and coupling pattern library in response.

[0106] Furthermore, such as Figure 4 As shown, the fault location sub-logic includes:

[0107] The fault contribution matrix is ​​received, and the causal relationship between current harmonic amplitude, bearing vibration velocity, load rate and winding temperature is automatically searched through the causal inference algorithm to generate a causal network.

[0108] The causal network is transformed into a Bayesian network, where each node represents a fault feature and each edge represents a causal relationship. The prior probability distribution of each node is set according to the fault contribution matrix.

[0109] By using Bayesian inference to determine the posterior probability of each fault feature as a fault source, potential fault sources are screened out, and the confidence level of each potential fault source is determined to locate the fault source.

[0110] Generator faults are typically caused by the interaction of multiple factors. A systematic approach is needed to reveal the causal relationships between variables such as current harmonic amplitude, bearing vibration velocity, load rate, and winding temperature. Causal networks can intuitively present the transmission paths between fault features, providing a logical framework for locating the fault source and avoiding blind troubleshooting. The system receives a fault contribution matrix output from the fault diagnosis module, using electrical and mechanical fault features, load rate, and winding temperature as nodes. Based on the fault contribution degree, the correlation strength between nodes is set; the higher the fault contribution degree, the higher the priority of the connection edge. Then, a PC algorithm is used to deduce the causal relationships between the nodes. This PC algorithm passes a conditional independence test, step by step... The process involves removing edges without causal relationships. First, the unconditional independence between all node pairs is calculated, and the connections between independent nodes are removed. Then, intermediate nodes are introduced as condition variables, and the independence is tested again. This process is repeated until no more edges can be removed, forming a causal network with a directed acyclic graph structure. Finally, the generated causal network is manually verified based on the generator's operating principles and expert experience. If the network shows that changes in winding temperature cause abnormal load rates, which is inconsistent with the actual physical logic, the incorrect connections are corrected to ensure that the causal network conforms to the generator's operating rules. This constructs a causal network that conforms to the generator's physical characteristics, clearly presenting the logical chain between fault characteristics, reducing the uncertainty of fault diagnosis, and improving fault location efficiency.

[0111] Causal networks only describe the logical relationships between variables and cannot quantify the probability of a fault occurring. Bayesian networks, through probabilistic models, incorporate the uncertainty of fault characteristics into the analysis, providing data support for assessing the likelihood of each node as a fault source. Causal networks are directly transformed into Bayesian networks, with each node corresponding to a random variable in the Bayesian network. Directed edges represent the dependencies of conditional probabilities. Prior probabilities are assigned to nodes based on the fault contribution matrix, with fault characteristics contributing more to the network receiving higher prior probabilities. Load rate and winding temperature are set based on the probability distribution of historical normal data. Current harmonic amplitude and bearing vibration velocity are discretized, with bearing vibration velocity divided into three intervals: "normal," "slightly abnormal," and "severely abnormal." A conditional probability table is constructed by statistically analyzing the frequency of occurrence of each interval in historical data. The Bayesian network is then trained using historical fault data, and the probability parameters in the conditional probability table are adjusted using maximum likelihood estimation or Bayesian estimation to make the Bayesian network more closely reflect actual fault scenarios. This allows for a comprehensive consideration of multi-factor uncertainties, providing a more scientific assessment of the probability of fault sources and reducing subjective judgment errors.

[0112] Based on Bayesian inference and current fault data, the posterior probability of each node as a fault source is calculated, thereby identifying fault sources and providing a clear direction for maintenance decisions. Real-time data of the current fault is input as observational evidence, including abnormal current harmonic amplitudes and bearing vibration speeds. The posterior probability of each node is calculated based on the Bayesian formula. All nodes are ranked according to their posterior probabilities, and the nodes with the highest posterior probabilities are selected as potential fault sources. A trust assessment mechanism is then introduced to refine the location results, comprehensively considering the historical frequency of fault characteristics, the reliability of detection equipment, and the mutual correlation between fault characteristics. The system establishes a causal relationship and applies a weighted correction to the posterior probability. If a fault feature has a high posterior probability but the detection sensor has recently exhibited anomalies, its confidence level is reduced, and a confidence threshold is set. Only nodes with confidence levels greater than the threshold are identified as the final fault source. If none of the nodes meet the threshold, the system returns to the causal network generation step to re-examine the causal relationship or supplement data. This allows for precise fault source localization, reducing the scope of troubleshooting for maintenance personnel, shortening fault handling time, and lowering maintenance costs. The identified fault source serves as the basis for updating diagnostic thresholds and optimizing the coupling pattern library. If the fault source is incorrectly identified, subsequent system parameter adjustments will fail.

[0113] Specifically, the logic for updating the diagnostic threshold includes:

[0114] Analyze the changing trends of current harmonic amplitude and bearing vibration velocity at the fault source, and determine the threshold offset of the changes in current harmonic amplitude and bearing vibration velocity caused by the fault source.

[0115] The diagnostic threshold is updated by using a reinforcement learning algorithm with the false positive rate and false negative rate as reward functions, combined with the threshold offset.

[0116] The effectiveness of the updated diagnostic thresholds was evaluated using cross-validation to determine whether further updates to the diagnostic thresholds were necessary.

[0117] Different fault sources cause varying signal variation amplitudes, and load rate and winding temperature affect the normal fluctuation range of the signal. Therefore, it is necessary to dynamically update the diagnostic threshold based on the characteristics of the fault source and the current operating conditions to avoid missed or false alarms due to fixed thresholds. Analyze the changing trends of current harmonic amplitude and bearing vibration velocity at the fault source during the fault development process. If the fault source is rotor eccentricity, extract characteristic parameters such as the abrupt change amplitude and rise rate of the first harmonic component amplitude in the vibration signal. Combine this with historical fault data to statistically determine the typical threshold offset caused by the fault source. If historical fault data is insufficient, refer to expert experience to set the initial range of the threshold offset, and combine it with load rate and winding temperature to correct the threshold offset, avoiding misjudging normal fluctuations as faults. This allows the diagnostic threshold to adapt to the characteristics of the fault source and changes in operating conditions, improving the sensitivity and specificity of fault detection and reducing diagnostic errors caused by unreasonable diagnostic thresholds.

[0118] Relying solely on experience or fixed rules to adjust thresholds is insufficient to adapt to complex and ever-changing operating scenarios. Reinforcement learning, by simulating a trial-and-error process and using false positive and false negative rates as feedback, automatically searches for the optimal threshold, enhancing the system's adaptability. The state space of reinforcement learning is defined to include information such as the current diagnostic threshold, load rate, winding temperature, historical fault detection results, and threshold offset. The action space sets the step size range for adjusting the diagnostic threshold; each action corresponds to an increase or decrease in the diagnostic threshold. The reward function aims to reduce the false positive and false negative rates, employing a weighted reward method and training with a Q-learning algorithm. Starting from an initial diagnostic threshold, the system selects actions based on the current state, observes the fault detection results under the new state, calculates the reward, and updates the Q-value table, gradually learning the optimal threshold adjustment strategy. This enables autonomous updating of the diagnostic threshold, allowing the system to automatically adjust to the best detection state under different operating conditions and fault modes, reducing manual intervention and improving diagnostic accuracy.

[0119] Reinforcement learning-derived thresholds can overfit or fail in specific scenarios, necessitating validation of their generalization ability across multiple datasets to ensure that adjusting the diagnostic threshold does not degrade overall system performance. In a specific application, k-fold cross-validation is used to divide historical fault data and normal operation data into five subsets. Four subsets are used each time to train the model and adjust the threshold, while the remaining subset is used for validation. This process is repeated five times to ensure each subset is used as the validation set. During validation, performance metrics for the updated diagnostic threshold are calculated, including false positive rate, false negative rate, and F1 score. The false positive rate represents the proportion of normal data misclassified as faulty, and the false negative rate represents the proportion of data misclassified as faulty. The proportion of fault data that is not detected is determined by the F1 score, which represents the overall accuracy and recall. If the F1 score of the updated diagnostic threshold is higher than the original diagnostic threshold by more than a certain threshold, and the false positive rate and false negative rate are both within an acceptable range, then the diagnostic threshold update is confirmed to be effective. Otherwise, the process returns to the reinforcement learning step, and the learning parameters are adjusted to re-optimize the diagnostic threshold. A strict verification mechanism ensures the reliability of the threshold adjustment, avoids the decline in diagnostic performance due to blind updates, and ensures system stability. The verified diagnostic threshold will be applied to subsequent fault detection. If it fails verification, it needs to be re-optimized to avoid affecting the system's response speed to new faults.

[0120] Specifically, the update logic for the coupling pattern library includes:

[0121] The DBSCAN algorithm is used to cluster abnormal data with unmatched fault coupling patterns, and the clustering regions of new fault coupling patterns are identified based on data density.

[0122] The mapping relationship between phase difference interval division and fault coupling identifier in the new fault coupling mode is dynamically adjusted by using the gradient descent algorithm to update the coupling mode library.

[0123] As generator operating conditions change or new faults emerge, existing coupling pattern libraries cannot cover all fault scenarios. Therefore, it is necessary to mine new fault coupling patterns from unmatched abnormal data to expand the system's fault identification capabilities. This involves collecting abnormal data not explained by existing patterns during fault diagnosis, including the coherence coefficient and phase difference between current harmonic components and bearing vibration frequency bands, as well as corresponding combined operating conditions, including load rate and winding temperature. The DBSCAN density clustering algorithm is used to process the abnormal data, setting a neighborhood radius and minimum sample size to group density-connected data points into the same cluster. If multiple phase differences exist within a certain region... Concentrated abnormal data points form a new cluster. Then, feature analysis is performed on each cluster to calculate the central phase difference, mean coherence coefficient, and distribution range of operating parameters within the cluster. This forms a preliminary description of the new fault coupling mode and identifies the clustering area of ​​the new fault coupling mode. If a cluster shows that the phase difference between the current harmonic amplitude and the bearing vibration velocity is concentrated at 180° and has a high probability of occurrence under high load rate, it is inferred to be a new fault coupling mode related to the load. This automatically discovers unknown fault coupling modes, fills the coverage blind spots of the coupling mode library, and improves the system's ability to diagnose new faults.

[0124] Newly identified fault coupling patterns suffer from issues such as ambiguous parameter boundaries and conflicts with existing patterns. Parameter optimization and mapping adjustments are needed to ensure the accuracy and compatibility of these new patterns. An optimization objective function is constructed, comprehensively considering cluster compactness, inter-cluster separation, and matching with historical fault data. Cluster compactness includes intra-cluster data dispersion, and inter-cluster separation includes inter-cluster distance. The boundaries of the phase difference interval and the coherence coefficient threshold are iteratively adjusted using a gradient descent algorithm to minimize the objective function value. If the phase difference interval of a new fault coupling pattern overlaps with that of an existing fault coupling pattern, the boundaries are fine-tuned to avoid confusion and to provide a basis for the new fault coupling pattern. Each fault coupling mode is assigned a unique fault coupling identifier, following existing identifier coding rules, including a letter prefix to distinguish mode type and a numeric suffix to represent the sequence number. This ensures that the new fault coupling identifier does not duplicate existing fault coupling modes, and the identifier mapping table is updated. Historical fault data is used to backtest and verify the new fault coupling mode. Specifically, historical fault data is substituted into the judgment logic of the new fault coupling mode to calculate the classification accuracy. If the classification accuracy is less than the classification threshold, the parameter optimization step is returned for readjustment. By optimizing the parameters and logic of the new fault coupling mode, seamless integration with the existing coupling mode library is ensured, improving the consistency and accuracy of fault diagnosis.

[0125] New, validated fault coupling modes are incorporated into the coupling mode library, enabling the system to directly invoke them during fault diagnosis. This facilitates continuous knowledge accumulation and iterative upgrades of system capabilities. New mode records are created in the coupling mode library, including mode name, phase difference range, coherence coefficient conditions, fault coupling identifier, and applicable operating conditions. Each update to the coupling mode library records the update time, update content, and operator information, and retains historical versions. If a new version encounters problems, it can be quickly rolled back to a stable version. This allows for timely integration of new knowledge into the system, enhancing its ability to diagnose complex faults, extending system lifespan, and reducing diagnostic blind spots caused by missing modes. The updated coupling mode library directly affects the fault coupling identifier marking in the coupling sensing module and the threshold adjustment in the fault diagnosis module. Errors during integration can trigger a chain reaction of diagnostic failures.

Claims

1. A general-purpose generator electrical monitoring system with fault self-diagnosis function, characterized in that, include: Coupled sensing module, fault diagnosis module, and fault tracing module; The coupling sensing module is used to acquire the current harmonic amplitude, bearing vibration velocity, load rate and winding temperature in real time, and simultaneously extract the current harmonic components and bearing vibration frequency band. In response to the judgment of the coherence coefficient and phase difference between the current harmonic components and the bearing vibration frequency band, the fault coupling identifier is marked by the coupling mode library. The fault diagnosis module is used to extract the diagnosis threshold. After receiving the fault coupling identifier, it combines the load rate and winding temperature to dynamically adjust the diagnosis threshold. During the adjustment process, it filters the current harmonic amplitude and bearing vibration velocity that are greater than the diagnosis threshold. It separates electrical faults from mechanical faults based on a dual-channel network. At the same time, it reconstructs the target current harmonics and target bearing vibrations through the dual-channel network, and generates a fault contribution matrix by comparing them with the current harmonic amplitude and bearing vibration velocity respectively. The logic for generating the fault contribution matrix includes: The contribution weights of electrical and mechanical fault features to fault diagnosis are determined based on the attention mechanism. The difference between the target current harmonic and the current harmonic amplitude, and the difference between the target bearing vibration and the bearing vibration velocity, are combined with contribution weights to determine the fault contribution of electrical fault characteristics and mechanical fault characteristics to fault diagnosis. The fault contribution rates of electrical fault features and mechanical fault features are arranged in order to generate a fault contribution matrix. The fault tracing module is used to receive the fault contribution matrix, analyze the fault causal chain and locate the fault source through the causal inference algorithm, and update the diagnostic threshold and coupling mode library in response.

2. The general-purpose generator electrical monitoring system with fault self-diagnosis function as described in claim 1, characterized in that, The marking logic for the fault coupling identifier includes: Configure a coherence coefficient threshold, filter associated frequency points with a coherence coefficient greater than the coherence coefficient threshold, and calculate the phase difference between the current harmonic component and the bearing vibration frequency band at the associated frequency points. Then, use a clustering algorithm to divide the frequency into multiple phase difference intervals based on the magnitude of the phase difference, with each phase difference interval representing a fault coupling mode. Establish a coupling mode library, map phase difference intervals to fault coupling modes, and assign fault coupling identifiers to each fault coupling mode to form an identifier mapping table; Based on the phase difference interval to which the current phase difference belongs, find the fault coupling identifier in the identifier mapping table and mark the associated frequency point.

3. The general-purpose generator electrical monitoring system with fault self-diagnosis function as described in claim 2, characterized in that, The sub-logic for determining the coherence coefficient includes: Real-time acquisition of current harmonic amplitude and bearing vibration velocity, and simultaneous extraction of current harmonic components and bearing vibration frequency band; The current harmonic components and bearing vibration frequency bands are divided into data segments according to time intervals. The dimensional features of the current harmonic components and bearing vibration frequency bands in each data segment are extracted by fast Fourier transform, and the dimensional features are reduced by principal component analysis. Based on the dimensionality features after dimensionality reduction, the coherence coefficients of the current harmonic components and the bearing vibration frequency band at each frequency point are calculated to form a coherence coefficient sequence. Then, outlier detection is performed on the coherence coefficient sequence using a clustering algorithm.

4. The general-purpose generator electrical monitoring system with fault self-diagnosis function as described in claim 3, characterized in that, The adjustment sub-logic for the diagnostic threshold includes: Historical operating data of the generator were obtained, and the data distribution of current harmonic amplitude and bearing vibration velocity under normal operating conditions were statistically analyzed to serve as the initial diagnostic threshold. After receiving the fault coupling identifier, the fault coupling mode corresponding to the fault coupling identifier is obtained from the coupling mode library, and the changing trends of current harmonic amplitude and bearing vibration velocity under the fault coupling mode are statistically analyzed to determine the adjustment weight of the initial diagnostic threshold. The load rate and winding temperature are acquired in real time, and the influence factors of load rate and winding temperature on current harmonic amplitude and bearing vibration velocity are analyzed by support vector regression algorithm. The adjustment coefficient is obtained by weighting and summing the adjustment weights of the initial diagnostic threshold and the influencing factors, and then combined with the initial diagnostic threshold to adjust the diagnostic threshold.

5. The general-purpose generator electrical monitoring system with fault self-diagnosis function as described in claim 4, characterized in that, The separation logic between electrical and mechanical faults includes: Data points with current harmonic amplitudes greater than the diagnostic threshold are selected and marked as electrical candidate fault points, and data points with bearing vibration velocity greater than the diagnostic threshold are selected and marked as mechanical candidate fault points; A dual-channel network consisting of physical and data channels is constructed. The physical channel enhances the features of electrical and mechanical candidate fault points to obtain electrical and mechanical fault features. The data channel extracts the correlation features between the current harmonic amplitude and the bearing vibration velocity under the fault coupling mode from the coupling mode library based on the fault coupling identifier, and performs real-time correction of the correlation features in combination with the load rate and winding temperature to generate fault feature templates. The cosine similarity between electrical fault features, mechanical fault features and fault feature templates is calculated separately to determine the probability of electrical faults and the probability of mechanical faults, and a decision threshold is configured to separate electrical faults and mechanical faults.

6. The general-purpose generator electrical monitoring system with fault self-diagnosis function as described in claim 5, characterized in that, The reconstruction sub-logic for the target current harmonics and target bearing vibration includes: The generator's current harmonic amplitude and bearing vibration velocity are acquired in real time through the physical channel; The historical current harmonic amplitude and historical bearing vibration velocity of the generator under normal operating conditions are retrieved from the data channel and classified and stored according to the combination of load rate and winding temperature to form a feature mapping relationship. The current harmonic amplitude and bearing vibration velocity obtained from the physical channel are matched in the feature mapping relationship with the load rate and winding temperature to screen normal operating condition features, and the normal operating condition features are fed back to the physical channel to reconstruct the target current harmonics and target bearing vibration. The reconstructed target current harmonics and target bearing vibrations are transmitted to the physical channel for error verification to determine whether to re-screen normal operating condition characteristics and reconstruct the target current harmonics and target bearing vibrations.

7. The general-purpose generator electrical monitoring system with fault self-diagnosis function as described in claim 6, characterized in that, The update logic for the diagnostic threshold includes: Analyze the changing trends of current harmonic amplitude and bearing vibration velocity at the fault source, and determine the threshold offset of the changes in current harmonic amplitude and bearing vibration velocity caused by the fault source. The diagnostic threshold is updated by using a reinforcement learning algorithm with the false positive rate and false negative rate as reward functions, combined with the threshold offset. The effectiveness of the updated diagnostic thresholds was evaluated using cross-validation to determine whether further updates to the diagnostic thresholds were necessary.

8. The general-purpose generator electrical monitoring system with fault self-diagnosis function as described in claim 7, characterized in that, The fault source location sub-logic includes: The fault contribution matrix is ​​received, and the causal relationship between current harmonic amplitude, bearing vibration velocity, load rate and winding temperature is automatically searched through the causal inference algorithm to generate a causal network. The causal network is transformed into a Bayesian network, where each node represents a fault feature and each edge represents a causal relationship. The prior probability distribution of each node is set according to the fault contribution matrix. By using Bayesian inference to determine the posterior probability of each fault feature as a fault source, potential fault sources are screened out, and the confidence level of each potential fault source is determined to locate the fault source.

9. The general-purpose generator electrical monitoring system with fault self-diagnosis function as described in claim 8, characterized in that, The update logic of the coupling mode library includes: The DBSCAN algorithm is used to cluster abnormal data with unmatched fault coupling patterns, and the clustering regions of new fault coupling patterns are identified based on data density. The mapping relationship between phase difference interval division and fault coupling identifier in the new fault coupling mode is dynamically adjusted by using the gradient descent algorithm to update the coupling mode library.

Images