Fault diagnosis system and method for hydropower oil filter based on blind source separation

Abstract

The present invention discloses a hydroelectric oil filter fault diagnosis system and method based on blind source separation, which relates to the field of fault diagnosis technology. The system includes a data acquisition module, an adaptive preprocessing module, a diagnostic engine module, a digital twin model library and an application module. The data acquisition module synchronously acquires multi-source heterogeneous observation signals; the adaptive preprocessing module uses adaptive variational modal decomposition based on an intelligent optimization algorithm for preprocessing; the diagnostic engine module includes a multi-physical quantity deep fusion unit, a dynamic source number estimation unit and an online blind source separation unit. The multi-physical quantity deep fusion unit deeply fuses heterogeneous data through a physical information autoencoder to generate a high-dimensional feature matrix. The dynamic source number estimation unit uses a three-layer hierarchical structure to estimate the number of source signals online. The online blind source separation unit uses an independent component analysis algorithm driven by the operating status to achieve a complete diagnostic process. It solves the problem of low diagnostic accuracy of traditional methods under strong noise, multi-source coupling and dynamic working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, in particular to a system and method for diagnosing faults of a hydroelectric oil filter based on blind source separation. Background Art

[0002] As a critical component of ensuring the proper functioning of the unit's lubrication system, the operating status of an online oil filter in a hydropower unit directly impacts the unit's safety and reliability. Traditional oil filter fault diagnosis relies primarily on periodic manual inspections and single-value monitoring, making it difficult to achieve early fault warning and precise fault location. As hydropower units become larger and more complex, the requirements for accurate and timely fault diagnosis in oil filter systems are increasing.

[0003] Existing oil filter fault diagnosis technologies primarily include mechanical fault diagnosis based on vibration analysis, lubrication system monitoring based on oil analysis, and thermodynamic fault identification based on temperature and pressure monitoring. While these methods have achieved some success in identifying specific fault types, they lack the ability to process multi-source, heterogeneous signals under complex operating conditions.

[0004] Some current research uses traditional signal processing techniques, such as wavelet transforms and empirical mode decomposition, to analyze oil filter operating signals, identifying faults by extracting time-frequency domain features. Others are attempting to apply machine learning methods, such as support vector machines and neural networks, to construct fault diagnosis models. Other research is also using multi-sensor information fusion to improve diagnostic accuracy, integrating multiple physical quantities such as vibration, pressure, and temperature for comprehensive analysis.

[0005] A certain degree of fault identification capability has been achieved through single or multiple traditional signal processing methods, but there are still certain limitations, such as: poor signal preprocessing effect in strong background noise environment, and inability to adaptively adjust key parameters; lack of physical constraints in multi-physical quantity information fusion, and the physical rationality of the fusion results is difficult to guarantee; low accuracy in source signal quantity estimation, especially large estimation errors under dynamic working conditions; fixed parameters of blind source separation algorithm, and insufficient adaptability to changes in working conditions; lack of effective cold start mechanism, and poor initial performance of online learning algorithm; lack of visualization and explainability of the diagnostic process, making it difficult to provide intuitive decision support for operation and maintenance personnel; single system architecture, poor scalability and maintainability, making it difficult to adapt to the diverse needs of complex industrial environments.

[0006] Therefore, there is an urgent need for an intelligent fault diagnosis technology solution for online oil filters of hydropower units that can achieve high-precision fault diagnosis under strong noise, multi-source coupling and dynamic working conditions, and has adaptive learning capabilities and good engineering applicability. Summary of the Invention

[0007] The purpose of the present invention is to overcome the deficiencies of the prior art and to propose a fault diagnosis system and method for a hydroelectric oil filter based on blind source separation to solve the above-mentioned problems.

[0008] The object of the present invention is achieved through the following technical solutions: a fault diagnosis system for a hydroelectric oil filter based on blind source separation, characterized in that it includes:

[0009] The data acquisition module is used to synchronously collect multi-source heterogeneous observation signals including vibration, pressure, temperature and oil quality when the online oil filter of the hydropower unit is running;

[0010] The adaptive preprocessing module uses adaptive variational mode decomposition based on an intelligent optimization algorithm to preprocess multi-source heterogeneous observation signals. It uses the particle swarm optimization algorithm or the gray wolf optimization algorithm to automatically optimize the mode number K and penalty factor α of the variational mode decomposition, adaptively decomposes the original mixed signal into multiple intrinsic mode functions, and achieves noise reduction and feature enhancement.

[0011] The diagnostic engine module receives the signal processed by the adaptive preprocessing module and performs diagnosis. The diagnostic engine module specifically includes:

[0012] The multi-physics deep fusion unit is configured to first extract the respective physical feature vectors of kurtosis, energy entropy, pressure pulsation amplitude, and temperature rise rate from the multi-source heterogeneous observation signals processed by the adaptive preprocessing module, then input the physical feature vectors into the physical information autoencoder. By adding physical law constraints to the loss function for learning, it finally generates a high-dimensional feature matrix that is deeply fused and contains physical meaning;

[0013] The dynamic source number estimation unit uses a three-layer hierarchical structure to estimate the number of independent source signals online. The first layer is the sentinel screening layer, which determines whether the system state has changed by monitoring the covariance matrix residual energy ratio of the high-dimensional feature matrix within the sliding time window. The second layer is the expert diagnosis layer, which is triggered when the state changes, runs multiple source number estimation algorithms in parallel, and outputs the candidate source number through the Bayesian evidence voting mechanism. The third layer is the physical constraint layer, which compares and verifies the candidate source number with the preset fault mode and effect analysis knowledge base to determine the final number of source signals.

[0014] The online blind source separation unit runs an online blind source separation algorithm based on independent component analysis. The learning rate or forgetting factor of the algorithm is dynamically adjusted in a state-driven manner based on the system state output by the dynamic source number estimation unit.

[0015] A digital twin model library that pre-stores the optimal separation matrix under various operating conditions. This library is used to provide the optimal initial separation matrix for the online blind source separation unit to achieve a hot start when the system is cold started or when there is a significant change in operating conditions.

[0016] The application module is used to receive and visually display the waveforms of each independent source signal decoupled by the online blind source separation unit and the final fault diagnosis results.

[0017] In the adaptive preprocessing module, the intelligent optimization algorithm uses the fitness function to evaluate the decomposition effect and determines the optimal parameter combination by minimizing the correlation between each modal function and maximizing the sparsity of each modal function.

[0018] The expert diagnosis layer of the dynamic source number estimation unit runs multiple source number estimation algorithms in parallel, including the improved Akaike information criterion algorithm, the Bayesian information criterion algorithm, and an algorithm based on random matrix theory. The Bayesian evidence voting mechanism assigns weights according to the historical accuracy of each algorithm; the physical constraint layer of the dynamic source number estimation unit, when performing comparison and verification, if the fault combination corresponding to the candidate source number does not exist in the fault mode and effect analysis knowledge base or is physically not valid, the confidence of the candidate source number will be downgraded or rejected, and the algorithm will be triggered to re-estimate.

[0019] The state-driven dynamic adjustment of the online blind source separation unit is as follows: when the dynamic source number estimation unit judges that the system state is stable, the forgetting factor is increased to close to 1 or the learning rate is reduced to enhance the stability of the model; when it is judged that the system state has undergone a sudden change, such as an increase in the number of source signals, the forgetting factor is reduced or the learning rate is increased to accelerate the learning of new fault features.

[0020] In the multi-physics deep fusion unit, the loss function L of the physical information autoencoder is composed of the reconstruction loss L rec and physical constraint loss L phy Weighted composition: L=L rec +λL phy , where λ is the weight coefficient, and the physical constraint loss is constructed based on the known physical relationship between pressure and vibration; the visualization methods of the application module include: on the same interface, simultaneously displaying the original mixed observation signal and the independent source signals decoupled by the online blind source separation unit in the form of dynamic time domain waveform diagram and spectrum waterfall diagram, and providing real-time visualization of the blind source separation process.

[0021] The system adopts a microservice architecture, decoupling data acquisition, preprocessing, diagnostic engine units, and visualization functions into independent services. The services communicate through RESTful APIs and provide a unified data access layer that supports OPC UA and Modbus-TCP industrial protocols.

[0022] The fault diagnosis method of the hydroelectric oil filter based on blind source separation includes the following steps:

[0023] S1. Data Collection Steps: Synchronously collect multi-source heterogeneous observation signals, including vibration, pressure, temperature, and oil quality, during operation of the hydropower unit's online oil filter.

[0024] S2. Adaptive preprocessing: Adaptive variational mode decomposition (VMD) based on an intelligent optimization algorithm is employed. Using either the particle swarm optimization algorithm or the gray wolf optimization algorithm, VMD key parameters are automatically optimized. This adaptively decomposes the multi-source heterogeneous observation signals into multiple intrinsic mode functions (IMFs), achieving noise reduction and feature enhancement.

[0025] S3. Deep fusion steps: First, extract the physical feature vectors of kurtosis, energy entropy, pressure pulsation amplitude, and temperature rise rate from the preprocessed multi-source heterogeneous observation signals. Then, using a physical information autoencoder, we learn the physical feature vectors by introducing physical laws as constraints in the loss function, generating a deeply fused high-dimensional feature matrix.

[0026] S4. Source number estimation step: A three-layer hierarchical algorithm is used to estimate the number of independent source signals online. First, a statistical indicator, the residual energy ratio of the covariance matrix, is monitored to determine whether the system state has changed. If so, multiple estimation algorithms are run in parallel and a Bayesian evidence voting mechanism is used to output candidate source numbers. Finally, this candidate source number is compared and verified with a pre-set FMEA knowledge base to determine the final number of source signals.

[0027] S5. Initial parameter loading step: When the system starts or the operating conditions change significantly, the optimal separation matrix that best matches the current operating conditions is matched and loaded from the pre-built digital twin model library. This matrix serves as the initial parameters for the subsequent online blind source separation algorithm.

[0028] S6. Online separation step: Based on the number of source signals and the loaded initial parameters, an online blind source separation algorithm based on independent component analysis is run to decouple the high-dimensional feature matrix and separate the individual source signals.

[0029] S7 dynamic update step: During the online separation step S6 operation, the system state outputted by the source number estimation step S4 is adjusted in real time and dynamically to adjust the learning rate or forgetting factor of the online blind source separation algorithm;

[0030] S8. Diagnosis output step: Based on the source signal, identify the specific failure mode and output the diagnosis result.

[0031] In the adaptive preprocessing step S2, the fitness function of the intelligent optimization algorithm evaluates the decomposition effect by minimizing the correlation between each modal function and maximizing the sparsity of each modal function, and determines the optimal modal number K and penalty factor α; in the deep fusion step S3, the loss function L of the physical information autoencoder is composed of the reconstruction loss L recand physical constraint loss L phy Weighted composition: L=L rec +λL phy , where λ is the weight coefficient and the physical constraint loss is constructed based on the known physical relationship between pressure and vibration.

[0032] In the source number estimation step S4, the covariance matrix residual energy ratio is the ratio of the covariance matrix residual energy of the high-dimensional feature matrix in the sliding time window to the total energy. When this ratio continuously exceeds the preset threshold, the expert diagnosis layer is triggered; the dynamic update step S7 is specifically as follows: when step S4 determines that the number of system source signals has increased, the forgetting factor of the online blind source separation algorithm is immediately reduced to accelerate the convergence of the signal characteristics of the newly emerging fault source; when the system state is stable, the forgetting factor is increased to close to 1 to enhance the stability of the model.

[0033] The method also includes preliminary steps for constructing a digital twin model library, which include: establishing a high-fidelity digital twin model of the online oil filter; simulating various single and compound faults in the digital twin model and collecting their simulation data; using the simulation data for offline training and optimization to obtain the optimal separation matrix corresponding to different fault combinations, and storing it in the digital twin model library.

[0034] The beneficial effects of the present invention are:

[0035] By employing adaptive variational modal decomposition (VMD) technology based on an intelligent optimization algorithm, the subjectivity and blindness inherent in manually setting parameters in traditional VMD methods are overcome. The intelligent optimization algorithm automatically optimizes key parameters such as the number of modes and penalty factor based on signal characteristics, significantly improving the adaptability and effectiveness of signal preprocessing. This maintains excellent decomposition results even in high-noise and complex signal environments, providing high-quality input data for subsequent diagnosis.

[0036] The deep fusion of multiple physical quantities using a physical information autoencoder ensures that the fusion results are not only mathematically optimal but also physically reasonable by introducing constraints based on the known physical relationship between pressure and vibration into the loss function. This physically constrained deep learning approach significantly improves the interpretability and credibility of diagnostic results compared to traditional mathematical fusion methods, making the integration of heterogeneous data more scientific and rigorous, and increasing diagnostic accuracy by over 30% compared to existing technologies.

[0037] The proposed three-layered dynamic source number estimation algorithm effectively addresses the technical challenges of low estimation accuracy and poor computational efficiency faced by existing source number estimation methods under dynamic conditions. The first sentinel screening layer rapidly determines the system state by monitoring the residual energy ratio of the covariance matrix, avoiding unnecessary computational overhead. The second expert diagnosis layer runs multiple estimation algorithms in parallel and integrates the advantages of multiple algorithms through a Bayesian evidence voting mechanism, significantly improving estimation accuracy. The third physical constraint layer ensures the physical rationality of the estimation results by comparing and verifying them with a fault mode knowledge base. This layered architecture ensures both computational efficiency and improved estimation accuracy, providing reliable prior information for the blind source separation algorithm.

[0038] A state-driven online blind source separation algorithm uses the system state information output by the dynamic source number estimation unit to adjust key parameters such as the learning rate and forgetting factor in real time. When the system state is stable, the algorithm increases the forgetting factor to enhance model stability. When a new fault source is detected, the algorithm decreases the forgetting factor to accelerate learning and convergence of the new fault signature. This adaptive parameter adjustment mechanism enables the algorithm to quickly respond to dynamically changing operating conditions while maintaining stability, significantly improving the system's adaptability to complex industrial environments.

[0039] The digital twin model library constructed by this invention uses offline simulation training to obtain optimal separation matrices under various operating conditions, effectively addressing the poor cold-start performance of online learning algorithms. Upon system startup or when significant operating conditions change, the most appropriate initial parameters can be quickly matched and loaded from the model library, significantly shortening the algorithm's convergence time and improving the system's response speed and initial diagnostic accuracy. This technical approach, combining offline training with online application, provides important performance guarantees for practical engineering applications.

[0040] The system utilizes a microservices architecture, breaking down traditional monolithic applications into multiple independent service modules that communicate via RESTful APIs. This architecture ensures excellent scalability and maintainability, enabling independent expansion and optimization of each functional module based on actual needs, adapting to application scenarios of varying scale and complexity. Furthermore, the unified data access layer supports mainstream industrial protocols such as OPC UA and Modbus-TCP, addressing compatibility issues with diverse industrial field equipment and reducing the complexity and cost of system deployment.

[0041] Comprehensive visualization capabilities, including dynamic time-domain waveforms and spectrum waterfall displays, make the complex blind source separation process intuitive and understandable, significantly improving the system's user acceptance and practicality. This real-time visualization of the separation process enhances the transparency of the diagnostic process, allowing operations and maintenance personnel to directly observe the algorithm's operating status and convergence, identify anomalies promptly, and initiate manual intervention, improving system controllability and reliability.

[0042] This system represents a fundamental shift from traditional reactive maintenance to intelligent predictive maintenance, enabling the detection of potential faults weeks or even months in advance, thus preventing sudden equipment damage. Precise fault tracing eliminates the need for traditional "blind" disassembly and repair, significantly shortening troubleshooting time and reducing repair costs. The system significantly enhances early fault warning capabilities, increasing the detection rate of subtle fault signatures by over 30% compared to existing technologies, effectively extending equipment lifespan and improving equipment availability and operational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The system architecture of the present invention Figure 1 ;

[0044] Figure 2 The system architecture of the present invention Figure 2 ;

[0045] Figure 3 A diagram showing the steps of the method of the present invention. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0047] It is to be noted that the directions of "left", "right", "up", "down", "front", "back", "inside" and "outside" in the following schemes are all relative directions and are not listed here one by one.

[0048] Example 1:

[0049] like Figures 1 to 3 As shown, this embodiment provides a core diagnosis engine implementation solution for an online oil filter fault diagnosis system for a hydropower unit based on blind source separation. The system achieves precise decoupling of complex mixed signals and accurate identification of fault sources through adaptive signal preprocessing, intelligent source number estimation, and state-driven online blind source separation.

[0050] The fault diagnosis system of this embodiment includes a data acquisition module, an adaptive preprocessing module, a diagnostic engine module, a digital twin model library, and an application module. The diagnostic engine module is the core of this embodiment and includes a multi-physics deep fusion unit, a dynamic source number estimation unit, and an online blind source separation unit.

[0051] The data acquisition module is configured to synchronously collect multi-source heterogeneous observation signals during the operation of the online oil filter of the hydropower unit. The module obtains vibration signals, pressure signals, temperature signals and oil quality signals in real time through a distributed sensor network. The vibration signal collects the vibration characteristics of the oil filter pump body, pipeline connections and key bearing positions through the acceleration sensor; the pressure signal monitors the inlet and outlet pressure difference of the oil filter, the system working pressure and the pressure pulsation characteristics through the pressure transmitter; the temperature signal collects the oil temperature, bearing temperature and the temperature rise of key components through the temperature sensor; the oil quality signal obtains the oil viscosity, acid value, moisture content and particle contamination parameters through the online oil monitoring device. All signal acquisitions are synchronized using a unified time base to ensure that the timing correspondence between different types of signals is accurate. The time synchronization accuracy of the acquisition system adopts a multi-level clock calibration algorithm:

[0052] ;

[0053] is the time synchronization error at time t, N is the number of sensors, is the importance weight of the i-th sensor, is the local timestamp of the i-th sensor, is the master clock timestamp, is the clock drift correction term, t is the current time;

[0054] in, Defined as: , is the drift coefficient, is the periodic drift frequency (rad / s), is the phase shift (rad);

[0055] This term compensates for systematic clock drift and periodic deviations caused by environmental factors.

[0056] The adaptive preprocessing module receives the multi-source heterogeneous observation signals output by the data acquisition module and preprocesses the original mixed signals using adaptive variational mode decomposition technology based on intelligent optimization algorithms. This module first constructs a fitness function and uses multi-objective optimization fitness design:

[0057] , is the value of the variational mode decomposition fitness function, is the weight coefficient of each item, O is the orthogonality index, S is the sparsity index, is the energy conservation indicator, is the modal discrimination index; where each term is defined as follows: the orthogonality index O adopts the improved correlation evaluation function:

[0058] ; is the mutual correlation coefficient between the ith and jth mode functions, are the center frequencies of the i-th and j-th modes (Hz), is the frequency sensitive parameter (Hz); the sparsity index S adopts the adaptive sparse measurement algorithm:

[0059] K is the total number of mode functions, is the i-th intrinsic mode function, norm (number of non-zero elements), norm (sum of absolute values), is the adaptive weight of the i-th mode;

[0060] in, Defined as:

[0061] is the variance of the ith mode function, is the mean variance, is the energy proportion of the i-th mode; energy conservation index Ensure the energy balance of the decomposition process:

[0062] is the total energy of the original signal, is the energy of the i-th mode, the modal discrimination index Evaluate the distinguishability of different modalities: , is the Kullback-Leibler divergence between the probability density functions of the i-th and j-th modes, is the probability density function of the i-th and j-th modes, is the frequency separation between modes i and j;

[0063] in, Defined as: , is the bandwidth (Hz) of the i-th and j-th modes; then, the particle swarm optimization algorithm or the gray wolf optimization algorithm is used as the intelligent optimization algorithm, and a hybrid optimization strategy is adopted:

[0064] is the position vector of the k+1th iteration, is the mixing weight coefficient, is the particle swarm optimization component, Optimize the weight for the gray wolf, is the variation component;

[0065] in, Optimize the component for the particle swarm:

[0066] w is the inertia weight, is the velocity vector of the kth iteration, is the learning factor, is a random number in the interval [0,1], is the optimal position of an individual, is the global optimal position, is the position vector of the kth iteration;

[0067] Optimize weights for gray wolves: , for The location of the wolf (optimal solution), for The location of the wolf (suboptimal solution), for The position of the wolf (third best solution);

[0068] is the variation component: , The mean is 0 and the variance is The normal distribution of is the adaptive factor;

[0069] in, , k is the current iteration number, is the maximum number of iterations, is the fitness ratio;

[0070] Using the fitness function as the objective function, the key parameters of variational mode decomposition are automatically optimized, including the mode number K and the penalty factor. During the optimization process, the algorithm searches for the optimal parameter combination within a preset parameter space. When the fitness function reaches convergence or the maximum number of iterations, the optimal parameters are output. Using the obtained optimal parameters, the original mixed signal is adaptively decomposed into multiple intrinsic mode functions (IMFs). Each IMF represents a component of the original signal within a specific frequency range, thereby achieving signal noise reduction and feature enhancement.

[0071] The diagnostic engine module is the core component of this embodiment, responsible for receiving the signals processed by the adaptive preprocessing module and performing intelligent diagnosis. This module contains three key units, which respectively implement different levels of signal processing and fault identification functions.

[0072] The multi-physics deep fusion unit is configured to perform feature extraction and deep fusion on pre-processed multi-source heterogeneous observation signals. The unit first extracts the physical feature vectors of each multi-source heterogeneous observation signal after it has been processed by the adaptive pre-processing module. For vibration signals, the kurtosis feature is extracted using an adaptive kurtosis calculation algorithm: , is the adaptive kurtosis value, N is the number of mode functions, is the time-varying weight coefficient of the i-th mode at time t, is the fourth-order central moment of the ith mode function, is the standard deviation of the ith mode function, is the time domain weight function; where, is the time-varying weight coefficient: , is the target frequency (Hz), is the regulating factor, is the relative energy of the i-th mode at time t; is the time domain weight function:

[0073] is the modulation depth, is the modulation frequency (Hz), is the phase shift of the ith mode (rad), is the enhancement factor based on the envelope change of the i-th modal signal;

[0074] For the pressure signal, the pressure pulsation amplitude characteristics are extracted and the multi-scale pulsation analysis algorithm is used:

[0075] ; is the multi-scale pulsation amplitude, S is the total number of scales, is the weight of the s-th scale, is the pressure signal of the sth scale at time t, is the variance of the s-th scale pressure signal, is the gradient weight factor, is the time derivative of the pressure signal at the sth scale;

[0076] in, is the scale weight: , is the optimal scale,

[0077] is the scale diffusion parameter, is the information content index of the s-th scale;

[0078] in, is the information content index: , is the probability of the jth feature at the sth scale, is the frequency-related factor of the sth scale. For the temperature signal, the temperature rise rate feature is extracted using the nonlinear temperature gradient algorithm: ; is the nonlinear temperature gradient, M is the number of temperature measurement points, is the importance coefficient of the kth measurement point, is the temperature of the kth measuring point (°C), is the reference temperature (°C), is the nonlinear correction coefficient of the kth measuring point, n is the nonlinear index, is the temperature activation function;

[0079] in, Is the activation function: ; is the temperature sensitivity parameter, is the threshold temperature (°C);

[0080] For multi-source signals, the energy entropy feature is also calculated using the time-frequency joint entropy algorithm: ; is the time-frequency joint entropy, is the probability density function of the time-frequency distribution, is the importance mapping function.

[0081] in, is the importance mapping function: , Time-frequency point The distance to the time-frequency ridge, is the distance diffusion parameter, is the local signal-to-noise ratio of the time-frequency point (t, f);

[0082] After the extraction is completed, these physical feature vectors are input into the physical information autoencoder for deep fusion. The physical information autoencoder adopts an encoder-decoder architecture, and its loss function adopts a multi-level physical constraint design: , is the total loss function of the physical information autoencoder, is the reconstruction loss, L is the number of physical constraint layers, is the weight coefficient of the physical constraint of the i-th layer, is the physical constraint loss of the i-th layer, is the regularization loss;

[0083] Among them, the reconstruction loss adopts an adaptive weighting method: , D is the feature dimension, is the adaptive weight of the j-th feature, is the jth input feature, is the j-th reconstruction feature, where , is the uncertainty of the j-th feature, For the importance of the j-th feature, the physical constraint loss is divided into multiple levels: , is the vibration-pressure physical mapping function, is the vibration eigenvector, is the pressure eigenvector, is the vibration-pressure coupling intensity coefficient; ;

[0084] is the temperature-pressure physical mapping function, is the temperature eigenvector, is the temperature-pressure coupling strength coefficient; , is the physical law constraint function, is the total eigenvector, is the physical law weight; where, is the coupling strength coefficient: , is the main frequency of the vibration signal (Hz), is the main frequency of the pressure signal (Hz), is the frequency diffusion parameter (Hz), is the coherence function of the vibration and pressure eigenvectors;

[0085] is the energy conservation constraint: , is the input energy, is the output energy, is the dissipated energy; the regularization loss contains sparsity and smoothness constraints: ; is the sparsity weight, is the latent space representation matrix, is the smoothness weight, is the gradient of the latent space representation, is the orthogonality weight, is the identity matrix, is the Frobenius norm;

[0086] Through this deep learning process, a high-dimensional feature matrix that is deeply integrated and contains physical meaning is finally generated. This matrix integrates information from multiple physical quantities and provides high-quality input data for subsequent source number estimation and blind source separation.

[0087] The dynamic source number estimation unit adopts a three-layer hierarchical structure to estimate the number of independent source signals online and provide key parameters for the blind source separation algorithm. The first layer is the sentinel screening layer, which uses the covariance matrix residual monitoring algorithm: , is the covariance matrix residual energy ratio, is the current covariance matrix, is the reference covariance matrix;

[0088] Among them, the calculation of the covariance matrix residual energy adopts the sliding window adaptive algorithm: , is the adaptive update coefficient, is the covariance matrix of the previous moment, is the new covariance matrix at the current moment; , is a dynamic factor based on the rate of change of the signal, is a stability indicator;

[0089] in, Is the stability index: , T is the time window length, is the time delay, For signal At time t and Correlation coefficient of

[0090] State change detection uses a multi-level decision algorithm: , is the state change detection result (Boolean value), is the covariance residual threshold, For trend analysis indicators, is the trend analysis threshold, is a continuity condition (Boolean), where Indicators for trend analysis: , W is the number of windows, is the time weight of the i-th window, is the residual ratio of the i-th window, is the mean of the residual ratios, is the standard deviation of the residual ratio;

[0091] The second layer is the expert diagnosis layer, which is triggered when the first layer detects a state change. The expert diagnosis layer runs multiple source number estimation algorithms in parallel and adopts the Bayesian evidence fusion mechanism: , is the number of estimated source signals, K is the number of source estimation algorithms, is the evidence weight of the kth algorithm, The number of source signals estimated by the kth algorithm;

[0092] Among them, the weight of evidence adopts an adaptive update algorithm: , is the weight of the kth algorithm at time t, is the comprehensive evaluation index of the kth algorithm, Based on the historical accuracy and current confidence, it is calculated as , is the weight coefficient, is the historical accuracy of the k-th algorithm, is the current confidence of the k-th algorithm, is the stability of the kth algorithm; where, For historical accuracy: , H is the number of historical time windows, is the time decay weight of the h-th historical time window, is the success rate of the kth algorithm in the hth historical time window;

[0093] is the time decay weight. is the time decay constant, is the current confidence, calculated based on the algorithm's internal consistency: ; is the bootstrap sample set of the kth algorithm, is the variance of the bootstrap sample, is the mean of the bootstrap samples, and the improved Akaike Information Criterion algorithm uses an adaptive penalty term: , is the modified Akaike information criterion value, L is the likelihood function value, k is the number of model parameters, is the adaptive penalty factor; where . is the penalty adjustment coefficient, n is the number of samples, is the noise standard deviation; the Bayesian Information Criterion algorithm uses a non-uniform prior: , is the extended Bayesian information criterion value, is the weight based on physical constraints, where Design based on physical constraints: , is the physically expected number of parameters, To account for the uncertainty of physical constraints, the algorithm based on random matrix theory uses an improved eigenvalue distribution test: , is the random matrix theory score, r is the number of eigenvalues, is the actual eigenvalue of the ith is the ith theoretical eigenvalue, is the weight of the i-th eigenvalue;

[0094] in, is the rank distance. is the rank distance of the i-th eigenvalue, the third layer is the physical constraint layer, and the fault pattern matching algorithm is used: , is the physical constraint score, M is the total number of failure modes, To estimate the similarity of the parameters to the mth failure mode, is the estimated parameter vector, is the parameter vector of the kth source and the mth fault mode, is the prior probability of the mth fault mode; the similarity measurement adopts the multidimensional distance algorithm: , is the weighted distance between the two parameter vectors is the similarity diffusion parameter , D is the parameter dimension, is the weight of the dth parameter, is the dth component of the two parameter vectors;

[0095] Prior probabilities are based on historical data and physical likelihood: , is the historical frequency of the mth failure mode, is the physical possibility of the mth failure mode;

[0096] The online blind source separation unit is the core computing component of the diagnosis engine and uses a state-driven parameter adaptation algorithm: , is the adaptive learning rate, is the basic learning rate, is the state factor, is the convergence factor, is the stability factor;

[0097] Among them, the state factor Dynamic adjustment based on system status:

[0098] is the state adjustment coefficient, is the state sensitivity parameter, is the norm of the source number change, , is the current source signal number, is the number of source signals at the previous moment;

[0099] Convergence Factor Monitor the algorithm convergence status: , is the convergence rate, is the convergence time constant; , is the objective function value at time t, For the moment The objective function value of

[0100] Stability Factor Evaluate separation matrix stability: , is the separation matrix, is the variance of the separation matrix, is the mean of the absolute values ​​of the separation matrix;

[0101] The forgetting factor adopts an adaptive adjustment strategy: , is the adaptive forgetting factor, is the basic forgetting factor, is the adaptive adjustment amount; , is the forgetting factor adjustment coefficient, is the system stability index, where is the signal stationarity index: , is the spectral centroid trajectory vector, and the cost function of online learning adopts multi-objective optimization design: , is the total cost function, is the independence cost function, is the time consistency weight, is the time consistency cost function, is the physical constraint weight, is the physical constraint cost function;

[0102] Among them, the independence cost function is: , Source signal Mutual information between Higher-order correlations between

[0103] Time consistency cost function: , T is the length of time, is the separation matrix at time t, is the time weight at time t;

[0104] Physical constraint cost function: , To separate the signal The physical model predicts that is the observed physical quantity;

[0105] The digital twin model library pre-stores the optimal separation matrix under various working conditions, using pattern matching and parameter interpolation algorithms: , is the optimal separation matrix, K is the number of stored working conditions, is the similarity weight of the kth working condition, is the separation matrix of the kth operating condition, is the interpolation correction term; the similarity weight is based on multi-dimensional working condition matching: , For the current working conditions and historical conditions distance, is the working condition similarity diffusion parameter, Calculate multi-dimensional working distance: , D is the working condition parameter dimension, is the weight of the dth operating condition parameter, is the dth parameter of the two working conditions;

[0106] The interpolation correction term handles the smooth transition between working conditions: , is the radial basis function interpolation, is the set of separation matrices of adjacent working conditions, It is an extrapolation factor to handle situations beyond the range of known operating conditions;

[0107] The application module is responsible for receiving the independent source signals decoupled by the online blind source separation unit and displaying the final fault diagnosis results in a visual manner. Fault identification uses a multi-level decision fusion algorithm: , is the fault decision value, N is the number of source signals, is the weight of the i-th source signal, is the failure probability of the i-th source signal, is the confidence of the i-th source signal;

[0108] The source signal weight is based on the signal quality assessment: , is the quality index of the i-th source signal, is the information content of the i-th source signal;

[0109] Failure probability calculation uses multi-feature fusion: , K is the number of fault types, is the weight of the Kth fault type, is the matching degree between the i-th source signal and the K-th fault mode, is the i-th source signal, is the mode parameter of the Kth fault;

[0110] Confidence assessment is based on multiple verifications: , is the temporal consistency confidence, is the spatial consistency confidence, is the probability of physical plausibility;

[0111] The working process of this embodiment is as follows: first, the data acquisition module synchronously collects multi-source heterogeneous observation signals when the online oil filter of the hydropower unit is running; then, the adaptive preprocessing module uses an intelligent optimization algorithm to automatically determine the optimal parameters of the variational modal decomposition, and decomposes the original mixed signal into multiple intrinsic modal functions; then, the multi-physical quantity deep fusion unit extracts the physical feature vectors from the preprocessed signal, and performs deep fusion through the physical information autoencoder to generate a high-dimensional feature matrix; then, the dynamic source number estimation unit intelligently estimates the number of independent source signals through a three-layer hierarchical structure; finally, the online blind source separation unit separates the high-dimensional feature matrix into independent source signals based on the estimated source number and state-driven parameter adjustment, thereby realizing accurate identification and positioning of faults.

[0112] Through intelligent parameter optimization of adaptive variational mode decomposition, the blindness and subjectivity of manually setting parameters in traditional methods are avoided, significantly improving the effect and adaptability of signal preprocessing. The three-layer hierarchical dynamic source number estimation algorithm combines the dual advantages of computational efficiency and estimation accuracy. It can not only monitor system state changes in real time, but also accurately estimate the number of source signals, providing reliable prior information for blind source separation. The multi-physical quantity deep fusion technology based on physical information autoencoders not only achieves effective integration of heterogeneous data, but also ensures the rationality of the fusion results through physical constraints, improving the accuracy and interpretability of diagnosis. The state-driven online blind source separation algorithm can dynamically adjust key parameters according to the system state, while ensuring the stability of the algorithm and improving the adaptability to dynamically changing working conditions. The introduction of the digital twin model library solves the problem of cold start of online learning, significantly improving the response speed and initial diagnostic accuracy of the system. Overall, this embodiment realizes the transition from passive maintenance to predictive maintenance, providing effective technical support for the intelligent operation and maintenance of online oil filters of hydropower units.

[0113] Example 2:

[0114] like Figures 1 to 3 As shown, this embodiment provides an implementation scheme for a multi-physical quantity deep fusion and visualization system based on a physical information autoencoder. The system realizes the intelligent fusion of heterogeneous data through deep learning technology with physical constraints, and adopts a microservice architecture to provide a comprehensive visual diagnostic interface and industrial protocol support, providing an efficient data processing and human-computer interaction platform for fault diagnosis of online oil filters of hydropower units.

[0115] This embodiment further improves the technical implementation of the multi-physics deep fusion unit based on the first embodiment, and focuses on building the visualization function of the application module and the system's microservice architecture. While maintaining the core diagnostic capabilities of the first embodiment, the system strengthens the physical constraint mechanism of data fusion, the intuitiveness of the visualization display, and the modernization and scalability of the system architecture.

[0116] The multi-physics deep fusion unit has been further enhanced and improved in this embodiment. This unit receives multi-source heterogeneous observation signals processed by the adaptive preprocessing module, including the intrinsic mode functions of vibration signals, pressure signals, temperature signals, and oil quality signals. For each type of signal, the unit executes a specialized feature extraction algorithm.

[0117] For the intrinsic mode function of the vibration signal, its kurtosis feature is calculated. This embodiment adopts the multi-scale kurtosis calculation formula: , For time The multi-scale kurtosis value at , N is the number of intrinsic mode functions, is the weight coefficient of the i-th mode, is the ith intrinsic mode function at time The fourth-order central moment of is the i-th mode at time The corresponding standard deviation, is the characteristic time scale of the ith mode, is the scale attenuation parameter;

[0118] This formula realizes multi-scale feature fusion through an exponential weight function, which can sensitively reflect the transient impact components in the signal and is a key feature for identifying mechanical faults such as bearing faults and gear collisions.

[0119] For the intrinsic mode function of the pressure signal, the pressure pulsation amplitude characteristics are extracted and the adaptive pulsation intensity algorithm is used: ; is the pressure pulsation intensity at time t, is the maximum pressure value in the sliding window, is the average pressure value in the sliding window, is the gradient weight factor, is the rate of change of pressure with time, is the modulation function of the dominant frequency, is the dominant frequency (Hz);

[0120] in, Defined as: , is the modulation depth, is the phase shift (rad);

[0121] This feature can effectively reflect fluid dynamics faults such as blockage of the internal flow channel of the oil filter and wear of the pump impeller.

[0122] For the intrinsic mode function of the temperature signal, the temperature rise rate characteristics are calculated using the nonlinear temperature rise gradient algorithm: , is the composite temperature rise rate, M is the number of temperature measurement points, is the importance coefficient of the kth measurement point, is the temperature of the kth measuring point (℃), is the nonlinear correction coefficient of the kth measuring point, is the reference temperature (°C), is the threshold temperature (°C), is the Heaviside step function. By introducing nonlinear terms and threshold functions, the algorithm can identify thermodynamic faults such as poor lubrication and bearing overheating at an early stage.

[0123] For the intrinsic mode functions of all signals, the energy entropy characteristics are also calculated using the weighted frequency domain entropy algorithm: , is the weighted energy entropy, F is the number of frequency points, is the normalized energy probability of the jth frequency point, is the frequency corresponding to the j-th frequency point (Hz), is the center frequency (Hz), is the frequency span (Hz), is the frequency weight factor; this feature can more accurately reflect the signal complexity and fault severity by introducing the frequency position weight.

[0124] These extracted physical feature vectors are then input into a physical information autoencoder for deep fusion processing. The physical information autoencoder uses a deep neural network architecture, consisting of two main components: an encoder network and a decoder network. The encoder network consists of multiple fully connected layers, each of which uses a nonlinear activation function to gradually map the input high-dimensional physical feature vectors to a low-dimensional latent space representation. The encoder design follows the principle of progressive dimensionality reduction, extracting the essential information and inherent correlations from the physical feature vectors through multiple layers of nonlinear transformations. The decoder network is symmetrical in structure to the encoder network, and reconstructs the latent space representation back to the original feature space through multiple fully connected layers to verify the integrity of the information maintained during the encoding process.

[0125] The core of the physical information autoencoder lies in the design of its loss function. This embodiment adopts a multi-constraint fusion loss function: ;

[0126] is the total loss function, is the reconstruction loss, is the physical constraint loss weight, is the physical constraint loss, is the orthogonality loss weight, is the orthogonality loss, is the sparsity loss weight, is the sparsity loss; where each term is defined as follows: the reconstruction loss uses the adaptive weighted mean square error: , D is the feature dimension, is the adaptive weight of the i-th feature at time t, is the i-th input feature, is the i-th reconstructed feature

[0127] The adaptive weight is defined as: , is the mean of the i-th feature, is the standard deviation of the i-th feature; the physical constraint loss is based on the fluid-structure coupling constraint:

[0128]

[0129] is the physical constraint loss function value, is the vibration-pressure physical coupling matrix, is the pressure-vibration physical coupling matrix, is the vibration eigenvector, is the pressure eigenvector, is the temperature-pressure coupling weight coefficient, is the temperature-pressure physical coupling matrix, is the pressure-temperature physical coupling matrix, is the temperature eigenvector;

[0130] Orthogonality loss ensures independence between features: , is the orthogonality loss function value, is the latent space representation matrix, is the transpose of the latent space representation matrix, is the identity matrix, is the Frobenius norm;

[0131] Sparsity loss facilitates feature selection: , is the sparsity loss function value, is the latent space representation matrix norm, is the sparse penalty coefficient, N is the potential space dimension, is the i-th element of the latent space representation matrix, is the sparse threshold;

[0132] Weight coefficient It is used to balance the contributions of various losses. By introducing these physical constraints and orthogonality and sparsity constraints into the loss function, the autoencoder not only pursues the mathematically optimal solution during the learning process, but also ensures that the fusion results conform to objective physical laws, improving the interpretability and credibility of feature fusion.

[0133] After training, the latent space representation of the encoder output of the physical information autoencoder forms a high-dimensional feature matrix. This matrix has moderate dimensionality, high information density, and clear physical meaning, providing high-quality input data for subsequent dynamic source number estimation and online blind source separation. Compared to the original multi-source heterogeneous signal, this high-dimensional feature matrix eliminates dimensional differences between different physical quantities, reduces redundant information, enhances useful signals, and significantly improves the processing performance and diagnostic accuracy of subsequent algorithms.

[0134] In this embodiment, the application module implements comprehensive visualization capabilities, providing operators with intuitive and rich diagnostic information. This module receives the individual source signals decoupled by the online blind source separation unit, as well as the intermediate results from the diagnostic process, and presents them comprehensively through various visualization methods. The core of this visualization method is the simultaneous display of the original mixed observation signal and the individual source signals decoupled by the online blind source separation unit in the form of dynamic time-domain waveforms and spectrum waterfall plots on the same user interface.

[0135] The dynamic time domain waveform is displayed in a multi-channel synchronous manner, and its visual quality assessment uses multi-dimensional signal clarity indicators: , is the comprehensive clarity index, K is the number of signal channels, is the weight of the kth channel, is the signal-to-noise ratio of the kth channel, is the gradient of the k-th channel signal, is the mutual information weight coefficient, is the mutual information between the k-th channel signal and the reference signal, is the kth channel signal, is the reference signal;

[0136] This metric comprehensively considers the signal-to-noise ratio, signal change rate, and information correlation to provide a quantitative assessment of visualization quality.

[0137] The spectrum waterfall diagram uses a three-dimensional display method, and its spectrum evolution characteristics use the time-frequency energy distribution algorithm: , is the time-frequency energy distribution at frequency f and time t, where f is frequency (Hz) and t is time (s). For signal The short-time Fourier transform of is the time domain signal, is the adaptive window function, is the enhancement function;

[0138] in, is the adaptive window function: , is the instantaneous frequency at time t (Hz), is the adaptive bandwidth at time t (Hz), For the enhancement function: , is the enhancement coefficient, is the steepness parameter, is the local energy at frequency f and time t, is the energy threshold;

[0139] The application module also provides a real-time visualization function for the blind source separation process. This function uses animation to display the iterative update process of the separation matrix. The separation effect is evaluated using a multi-index fusion algorithm:

[0140] , As a comprehensive performance indicator, is the signal improvement ratio weight coefficient, is the signal improvement ratio, is the crosstalk index weight coefficient, is the crosstalk indicator, is the convergence rate weight coefficient, is the convergence rate indicator, and the definitions are as follows: Signal Improvement Ratio Calculation formula: , is the power calculation function, is the signal after the ith separation, is the i-th mixed signal, crosstalk Calculation formula: , is the correlation calculation function, is the ith separation signal, is the jth separated signal, the convergence rate Calculation formula: , is the loss function value of the nth iteration, For the The loss function value of the iteration, is the convergence threshold, weight coefficient Adjusted to the application's needs, this comprehensive metric provides a real-time quantitative assessment of separation quality.

[0141] The microservice architecture is an important technical feature of this embodiment. This architecture decomposes the traditional monolithic application into multiple independent service modules. The communication efficiency between each microservice adopts an adaptive load balancing algorithm: ;

[0142] Score the load of the i-th service, is the CPU usage weight coefficient, is the CPU usage of the i-th service (normalized value), is the memory usage weight coefficient, is the memory usage of the i-th service (normalized value), is the network delay weight coefficient, is the network delay of the ith service (normalized value), is the queue length weight coefficient, The queue length for the i-th service (normalized value);

[0143] The service selection probability adopts the flexible maximum algorithm: , is the probability of selecting the i-th service, is the temperature parameter, which controls the balance between randomness and determinism in the selection. j is the service index, which is used to sum the communication between microservices through the RESTful API. The API response time prediction adopts the autoregressive model: , is the predicted value of the response time at the n+1th moment, P is the order of the autoregressive model, is the kth autoregressive parameter, is the actual value of the response time at the n-k+1th moment, Q is the order of the moving average model, is the kth moving average parameter, For the The white noise error term at time , As a correction item for the current load, the system provides a unified data access layer that supports access to multiple industrial communication protocols. The protocol adaptation efficiency uses multi-protocol fusion indicators: , is the protocol adaptation efficiency, M is the number of protocol types, is the usage frequency weight of the kth protocol, is the data throughput of the kth protocol, is the communication delay of the kth protocol, is the data quality factor of the kth protocol, where Defined as: , is the error rate of the kth protocol, is the packet loss rate of the kth protocol, is the reliability coefficient of the kth protocol

[0144] The working process of this embodiment is as follows: the unified data access layer collects multi-source heterogeneous observation signals from field equipment through the supported industrial protocol, and performs format standardization and time synchronization processing; the data acquisition service receives the standardized signal data and passes it to the preprocessing service through the RESTful API; the preprocessing service performs adaptive variational mode decomposition and passes the processing results to the multi-physical quantity fusion service through the API; the multi-physical quantity fusion service runs the physical information autoencoder, extracts physical feature vectors and performs deep fusion to generate a high-dimensional feature matrix; the feature matrix is ​​passed to the source number estimation service and visualization service at the same time through the API; the source number estimation service executes a three-layer hierarchical estimation algorithm and passes the source number estimation result to the blind source separation service; the blind source separation service executes an online separation algorithm based on the received feature matrix and source number information, and passes the separation result to the visualization service; the visualization service integrates all received data to generate a dynamic time domain waveform diagram, a spectrum waterfall diagram and a real-time separation process display, providing complete diagnostic information for operation and maintenance personnel.

[0145] The deep fusion technology based on physical constraints ensures the rationality and interpretability of the integration of multi-physical quantity information. Compared with traditional mathematical fusion methods, the introduction of a physical constraint loss function makes the fusion results more consistent with objective physical laws, improving the credibility and accuracy of the diagnostic results. Comprehensive visualization capabilities provide operators with an intuitive display of diagnostic information. The comparative display of dynamic time-domain waveforms and spectrum waterfall diagrams makes the complex blind source separation process understandable and verifiable, significantly improving the system's user acceptance and practicality. The real-time visualization of the separation process enhances the transparency of the diagnostic process. Operators can observe the working status and convergence of the algorithm, detect anomalies in a timely manner, and perform manual intervention, thereby improving the controllability and reliability of the system. The adoption of a microservices architecture makes the system scalable and maintainable. Each service module can be independently expanded and optimized according to actual needs, adapting to application scenarios of varying scale and complexity. The standardized communication method of the RESTful API facilitates integration and expansion with third-party systems, supporting the construction of a larger-scale intelligent operation and maintenance ecosystem. The unified data access layer's support for multiple industrial protocols solves the problem of diverse industrial field equipment and reduces the complexity and cost of system deployment.

[0146] Application case: A large hydropower station has an installed capacity of six 300-megawatt mixed-flow turbine generator sets, each equipped with an online oil filter system. The power station uses the multi-physical quantity fusion and visualization system of this embodiment to diagnose oil filter faults. The system is deployed in the industrial Ethernet environment of the power station and connects to the power station's existing distributed control system via the OPC UA protocol to obtain real-time data such as vibration, pressure, and temperature of the oil filter. At the same time, it connects to the online oil monitoring device via the Modbus-TCP protocol to obtain oil quality parameters. In actual operation, the system successfully detected an early wear fault in the oil filter pump bearing of Unit 2. In the early stage of the fault, the fault characteristics in the original mixed signal were weak and masked by normal operation noise, making it difficult to identify with traditional methods. The multi-physical quantity deep fusion unit of this system uses a physical information autoencoder to deeply fuse the kurtosis characteristics, pressure pulsation amplitude characteristics, temperature rise rate characteristics, and energy entropy characteristics of the vibration signal. The generated high-dimensional feature matrix effectively enhances the fault feature signal. The dynamic source number estimation unit accurately identified the presence of three independent source signals in the system: the normal pump operation signal, the bearing wear fault signal, and the background noise signal. The online blind source separation unit successfully separated the mixed signal into three independent source components, among which the bearing wear fault signal showed typical high-frequency impact characteristics and a gradually increasing trend. The visualization interface clearly showed the signal comparison before and after separation. The spectrum waterfall diagram of the original mixed signal showed a broadband noise distribution, and the separated fault source signal showed obvious energy concentration and time evolution pattern in the high frequency band. The operation and maintenance personnel arranged the maintenance plan in time according to the visualization results and carried out preventive replacement before the bearing was completely damaged, avoiding equipment downtime losses. This case shows that this embodiment can work effectively in complex industrial environments and provide reliable technical support for predictive maintenance of equipment.

[0147] This embodiment uses physically constrained deep learning technology, comprehensive visualization capabilities, and a modern microservice architecture to provide an efficient, intuitive, and scalable technical solution for fault diagnosis of online oil filters in hydropower units. It achieves a shift from traditional passive maintenance to intelligent predictive maintenance, laying a solid technical foundation for the intelligent operation and maintenance of industrial equipment.

[0148] Example 3:

[0149] This embodiment provides a complete implementation plan for a fault diagnosis method for an online oil filter of a hydropower unit based on blind source separation. This method implements intelligent diagnosis of the entire process from data acquisition to fault identification through a systematic diagnostic process of eight ordered steps combined with offline pre-training technology of a digital twin model library, thus providing a complete methodology and technical path for predictive maintenance of online oil filters of hydropower units.

[0150] This example builds upon Examples 1 and 2 to establish a comprehensive fault diagnosis methodology, particularly enhancing the systematic nature of the methodology, the scientific nature of parameter optimization, and the construction of a digital twin model library. This methodology encompasses not only the complete online diagnosis process but also the preliminary preparations for offline model training and knowledge base construction, forming a complete technical closed loop from offline to online, and from model training to practical application.

[0151] The fault diagnosis method for online oil filters for hydropower units based on blind source separation consists of eight core steps, each with clear technical content and implementation methods. The data acquisition step, as the initial step, is responsible for synchronously collecting multi-source heterogeneous observation signals, including vibration, pressure, temperature, and oil quality, during the operation of the online oil filter. This step utilizes a distributed sensor network architecture, deploying different types of sensors at key locations within the oil filter. Vibration sensors are located on the bearing seat, inlet and outlet flanges, and major pipeline connections of the oil filter pump body, using triaxial accelerometers to simultaneously monitor radial and axial vibration. Pressure sensors are installed at the oil filter's inlet, outlet, bypass line, and before and after the filter element to monitor system pressure distribution and pressure differential changes in real time. Temperature sensors are located on the oil tank, pump body, bearings, and key pipeline nodes to monitor the system's thermodynamic state. Oil quality sensors, integrated into the oil monitoring device, provide real-time monitoring of key parameters such as viscosity index, acid value, moisture content, and particulate contamination. The sampling frequencies of all sensors are optimized based on the dynamic characteristics of the monitored physical quantities. Vibration signals use high-frequency sampling to capture impact fault characteristics, pressure and temperature signals use medium-frequency sampling to monitor gradual faults, and oil quality signals use low-frequency sampling to track long-term degradation trends. The data acquisition system is equipped with a high-precision time synchronization device to ensure that the signals collected by different sensors have an accurate time sequence correspondence. A distributed time synchronization algorithm is used:

[0152] : synchronization time vector at time t, : the master clock time vector at time t, : Network delay compensation vector, : Time offset correction vector, where the time offset correction matrix uses the adaptive Kalman filter algorithm:

[0153] : the time-varying Kalman gain matrix at time t, : the observation vector at time t, : observation matrix, : The prediction vector of the state at time t based on the information at time t-1, where is the time-varying Kalman gain matrix:

[0154] : prediction error covariance matrix, : the transpose of the observation matrix, : The clock noise covariance matrix and the improved algorithm of the two-way time transfer protocol are used for network delay compensation to provide a reliable data basis for the subsequent multi-source signal fusion.

[0155] The adaptive preprocessing step receives the multi-source heterogeneous observation signals obtained in the data acquisition step and preprocesses the raw signals using adaptive variational mode decomposition (VMD) technology based on an intelligent optimization algorithm. The core of this step is to automatically optimize the key parameters of VMD, including the mode number K and the penalty factor α, using a particle swarm optimization algorithm or a gray wolf optimization algorithm.

[0156] In view of the characteristics of multi-source heterogeneous signals, a multi-signal collaborative fitness function is used to evaluate the parameter optimization effect of variational mode decomposition. This function comprehensively considers the characteristics of various signal types and uniformly evaluates key indicators such as orthogonality, sparsity, and consistency of each signal through weighted fusion. In the function design, each signal type is assigned a corresponding weight coefficient to ensure that the contribution of different physical quantity signals in the parameter optimization process is reasonably balanced. The orthogonality index is used to measure the degree of independence between the modal functions after decomposition, the sparsity index reflects the sparsity of the signal in a specific transformation domain, and the consistency index evaluates the degree of preservation of the decomposition result and the original signal characteristics.

[0157] Building on the multi-signal collaborative evaluation, the function also introduces an inter-signal interaction consistency term, which uses a multidimensional information entropy algorithm to quantify the correlation between different signal types. By calculating the joint probability density distribution between each signal pair and combining it with the corresponding physical correlation weights, a comprehensive evaluation of the signal interactions is formed. This design ensures that the optimization process not only considers the decomposition quality of individual signals but also fully utilizes the physical coupling relationships between multiple source signals.

[0158] The design of the physical correlation weight matrix is ​​based on the causal relationships and physical mechanisms between signals. This matrix analyzes the time delay characteristics between different signals, combines them with the desired physical delay parameters, and uses an exponential decay function to quantify the causal strength between signals. Furthermore, a coherence function is introduced to assess the correlation between signals at specific frequencies, ensuring that the weight assignment conforms to the actual physical coupling laws. This physics-based weight design method enables the fitness function to more accurately reflect the inherent correlations of multi-source signals, providing a scientific and reliable evaluation criterion for the automatic optimization of variational mode decomposition parameters.

[0159] The fitness function of the intelligent optimization algorithm evaluates the decomposition performance by minimizing the correlation between modal functions and maximizing the sparsity of each modal function. Correlation is assessed by calculating the Pearson correlation coefficient between different modal functions. A smaller absolute value of the correlation coefficient indicates greater independence between the modal functions. Sparsity is assessed by calculating the sparsity index for each modal function. Sparsity can be quantified by the ratio of the L1 norm to the L2 norm; a larger ratio indicates greater signal sparsity. The fitness function weights these two metrics to form a comprehensive evaluation metric. During parameter optimization, the particle swarm optimization algorithm simulates the foraging behavior of bird flocks to search for the global optimal solution in parameter space. Each particle represents a set of parameter combinations, and the solution is continuously approached through speed and position update mechanisms. The gray wolf optimization algorithm simulates the social hierarchy and hunting behavior of gray wolf packs and uses strategies such as siege, encirclement, and attack to conduct a global search in parameter space. The optimization algorithm automatically adjusts its search strategy based on feedback from the fitness function. When convergence conditions or the maximum number of iterations are reached, it outputs the optimal number of modes K and penalty factor α. Using these optimized parameters, the variational mode decomposition algorithm adaptively decomposes the original multi-source heterogeneous observation signal into multiple intrinsic mode functions, each of which represents the component of the signal within a specific frequency bandwidth, achieving effective signal noise reduction and feature enhancement.

[0160] The deep fusion step is a key component of the method, responsible for extracting valuable physical features from the preprocessed signals and performing intelligent fusion. This step first extracts the physical feature vectors from the preprocessed multi-source heterogeneous observation signals, including key features such as kurtosis, energy entropy, pressure pulsation amplitude, and temperature rise rate. The kurtosis feature is extracted by calculating the ratio of the fourth-order central moment to the square of the second-order central moment of the intrinsic mode function. This feature is highly sensitive to transient shocks and abnormal spikes in the signal and is an important indicator for identifying mechanical faults such as bearing faults and gear shock. The energy entropy feature is obtained by calculating the Shannon entropy of the signal's energy distribution in the frequency domain. It reflects the complexity and randomness of the signal's energy distribution and can effectively identify changes in the system's operating state. The pressure pulsation amplitude feature is extracted by analyzing the pulsation characteristics of the pressure signal, including amplitude, frequency, and phase information, and can reflect the dynamic changes within the fluid system. The temperature rise rate feature is obtained by calculating the time derivative of the temperature signal. It reflects the rate of change of the system's thermodynamic state and is an important indicator for early fault warning. After extraction, these physical feature vectors are input into a physical information autoencoder for deep fusion processing. The physical information autoencoder uses a deep neural network architecture and learns physical feature vectors by incorporating physical laws as constraints in the loss function. The loss function L is composed of a reconstruction loss Lreconstruction and a weighted physical constraint loss Lphysical constraint. L equals Lreconstruction plus Lphysical constraint times λ, where λ is a weight coefficient. The reconstruction loss ensures that the autoencoder accurately reconstructs the input features. The physical constraint loss is based on the known physical relationship between pressure and vibration, including physical laws such as fluid dynamics, structural dynamics, and heat transfer. This physically constrained deep learning process ultimately generates a deeply fused high-dimensional feature matrix that not only preserves the important information of the original features but also exhibits good physical interpretability.

[0161] The source number estimation step uses a three-layer hierarchical algorithm to estimate the number of independent source signals online, providing key parameters for subsequent blind source separation. This step first determines whether the system state has changed by monitoring the statistical indicators of the covariance matrix residual energy ratio. The covariance matrix residual energy ratio uses a time-space joint algorithm to , is the residual energy ratio of the covariance matrix at time t, is the current covariance matrix at time t, is the benchmark covariance matrix at time t, is the Frobenius norm, is the regularization parameter to prevent the denominator from being zero;

[0162] Among them, the adaptive benchmark covariance matrix uses the exponentially weighted moving average algorithm to be , is the exponentially weighted moving average coefficient, for The benchmark covariance matrix at time , is the confidence function at time t, and the confidence function is evaluated based on system stability: , is the system variability at time t, is the stability threshold, is the Heaviside step function, For the system stable operation time, is the minimum stabilization time requirement.

[0163] Among them, the system variability adopts the multidimensional stability index as , K is the number of monitoring variables, is the weight of the kth variable, is the value of the kth variable at time t, is the variance of the kth variable, is the mean of the absolute values ​​of the kth variable, is the reference variability of the k-th variable;

[0164] The state change detection uses a multi-level statistical test algorithm to , is the state change detection result at time t (Boolean value), is the covariance residual threshold, is the trend test statistic, is the trend test threshold, is the chi-square test statistic, is the critical value of the chi-square test, is the logical AND operator;

[0165] The trend test statistic uses the improved algorithm of the nonparametric Mann-Kendall test: , is the Mann-Kendall statistic, is the expected value of the Mann-Kendall statistic, is the variance of the Mann-Kendall statistic, is the adaptive weight at time t;

[0166] This ratio reflects the stability of the system's statistical characteristics. The calculation process begins by calculating the covariance matrix of the high-dimensional feature matrix within the current time window. This is then compared with the historical baseline covariance matrix, and the residual matrix is ​​calculated. The ratio of the residual energy to the total energy is then calculated. When this ratio continuously exceeds a preset threshold, it indicates a significant change in the system state, triggering in-depth analysis at the expert diagnostic layer.

[0167] The expert diagnosis layer runs multiple estimation algorithms in parallel and uses the Bayesian evidence fusion algorithm to , is the number of estimated source signals, M is the number of source estimation algorithms, is the weight of the ith algorithm at time t, is the number of source signals estimated by the ith algorithm, is the consistency correction item;

[0168] The weight update uses the adaptive Bayesian learning algorithm to ;

[0169] is the updated weight of the ith algorithm at time t+1, is the learning rate parameter, is the accuracy evaluation of the ith algorithm at time t, is the prior probability of the i-th algorithm;

[0170] The accuracy evaluation uses a sliding window weighted algorithm. , W is the sliding window size, is the time decay factor, is the time index, It is an indicator function that returns 1 if the condition is true, otherwise it returns 0. For the i-th algorithm The estimated value of the moment, The number of real source signals at the moment, is the i-th algorithm at time Confidence correction factor of ;

[0171] The confidence correction factor is calculated based on the internal consistency of the algorithm as ; is the i-th algorithm at time The bootstrap sample set, is the variance calculation function, is the mean calculation function, is the i-th algorithm at time The residual vector of is the residual threshold;

[0172] The consistency correction term handles the systematic deviation between algorithms as , is the median of all algorithm estimates, is the deviation correction factor;

[0173] These include an improved Akaike Information Criterion algorithm, a Bayesian Information Criterion algorithm, and an algorithm based on random matrix theory. Each algorithm is based on a different theoretical foundation and evaluates the number of information sources from different perspectives. The Akaike Information Criterion selects the optimal model order by balancing model fit and complexity. The Bayesian Information Criterion adds a stronger penalty term to the Akaike Criterion to avoid overfitting. The random matrix theory algorithm estimates the number of information sources by analyzing the distribution of eigenvalues. The estimation results of these algorithms are combined through a Bayesian evidence voting mechanism, and weights are assigned based on the historical accuracy of each algorithm to ultimately output the number of candidate sources. The third physical constraint layer compares and verifies the number of candidate sources with the preset fault mode and effect analysis knowledge base to ensure the physical rationality of the estimation results and ultimately determine the number of source signals.

[0174] The initial parameter loading step is executed when the system starts or when the operating conditions change significantly. The optimal separation matrix that best matches the current operating conditions is matched and loaded from the pre-built digital twin model library as the initial parameters for the subsequent online blind source separation algorithm. This step uses a multi-dimensional operating condition similarity matching algorithm.

[0175] , is the similarity between the current working condition and the working condition in the library, is the current working condition parameter vector, is the working condition parameter vector in the library, D is the working condition parameter dimension, is the importance weight of the dth parameter, is the dth parameter of the current working condition, is the dth parameter of the working condition in the library, is the similarity diffusion parameter of the dth parameter, is the influencing factor of the dth parameter;

[0176] The importance weight is calculated using the information gain algorithm. , is the unconditional entropy of the output, is the conditional entropy of the given d-th operating condition parameter, is the output vector, is the dth working condition parameter, is the correlation correction factor of the dth parameter, where the conditional entropy is calculated using the kernel density estimation algorithm. , Output y and parameters The joint probability density function of For the given parameters The conditional probability density function of the output y under the condition, y is the output variable, is the dth working condition parameter;

[0177] The physical constraints ensure that the solution is physically reasonable. , is the number of physical constraints, is the weight of the c-th physical constraint, is the coefficient matrix of the cth constraint, is the right-hand side vector of the c-th constraint, is the activation function of the cth constraint;

[0178] The constraint activation function is based on physical feasibility judgment:

[0179] ;

[0180] is the activation threshold of the c-th constraint, is the softening parameter of the cth constraint;

[0181] The optimal separation matrix corresponding to different fault combinations is obtained. The training process utilizes supervised learning, using known fault information in the simulation data as labels. An optimization algorithm is used to search for the optimal separation matrix parameters. The optimization objective function comprehensively considers multiple performance indicators, including separation accuracy, convergence speed, and robustness, and is solved using a multi-objective optimization algorithm. The optimal separation matrix obtained through training is categorized and stored according to operating condition characteristics, forming the core content of the digital twin model library. The model library also includes auxiliary content such as a fault mode knowledge base, a parameter optimization strategy library, and a performance evaluation index library, providing comprehensive knowledge support for online diagnosis.

[0182] This embodiment provides a complete technical solution for fault diagnosis of online oil filters in hydropower units through a systematic diagnostic method flow, a scientific parameter optimization strategy, and advanced digital twin model technology. It achieves a fundamental shift from traditional empirical maintenance to intelligent predictive maintenance, and lays a solid methodological foundation for equipment health management and intelligent operation and maintenance in the hydropower industry.

[0183] The foregoing description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein and should not be construed as excluding other embodiments. Rather, the present invention can be used in various other combinations, modifications, and environments and can be modified within the scope of the concept described herein through the above teachings or techniques or knowledge in the relevant field. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention are intended to be protected by the appended claims.

Claims

1. The fault diagnosis system of hydropower oil filter based on blind source separation is characterized by: include: The data acquisition module is used to synchronously collect multi-source heterogeneous observation signals including vibration, pressure, temperature and oil quality when the online oil filter of the hydropower unit is running; An adaptive preprocessing module preprocesses the multi-source heterogeneous observation signals using adaptive variational mode decomposition based on an intelligent optimization algorithm, automatically optimizes the mode number K and penalty factor α of the variational mode decomposition using a particle swarm optimization algorithm or a gray wolf optimization algorithm, and adaptively decomposes the original mixed signal into multiple intrinsic mode functions to achieve noise reduction and feature enhancement; A diagnosis engine module receives the signal processed by the adaptive preprocessing module and performs diagnosis, and the diagnosis engine module specifically includes: A multi-physical quantity deep fusion unit is configured to first extract the respective physical feature vectors of kurtosis, energy entropy, pressure pulsation amplitude, and temperature rise rate from the multi-source heterogeneous observation signals processed by the adaptive preprocessing module, then input the physical feature vectors into a physical information autoencoder, and learn by adding physical law constraints to the loss function to ultimately generate a high-dimensional feature matrix that is deeply fused and contains physical meaning; The dynamic source number estimation unit uses a three-layer hierarchical structure to estimate the number of independent source signals online. The first layer is the sentinel initial screening layer, which determines whether the system state has changed by monitoring the covariance matrix residual energy ratio of the high-dimensional feature matrix within the sliding time window; the second layer is the expert diagnosis layer, which is triggered when the state changes, runs multiple source number estimation algorithms in parallel, and outputs the candidate source number through the Bayesian evidence voting mechanism; the third layer is the physical constraint layer, which compares and verifies the candidate source number with the preset fault mode and effect analysis knowledge base to determine the final number of source signals; An online blind source separation unit, which internally runs an online blind source separation algorithm based on independent component analysis, wherein the learning rate or forgetting factor of the algorithm is dynamically adjusted in a state-driven manner according to the system state output by the dynamic source number estimation unit; A digital twin model library pre-stores optimal separation matrices under various operating conditions, which is used to provide the online blind source separation unit with an optimal initial separation matrix to achieve a hot start when the system is cold started or when the operating conditions change significantly; The application module is used to receive and visually display the waveforms of the independent source signals decoupled by the online blind source separation unit and the final fault diagnosis result.

2. The hydroelectric oil filter fault diagnosis system based on blind source separation according to claim 1 is characterized by: In the adaptive preprocessing module, the intelligent optimization algorithm uses a fitness function to evaluate the decomposition effect and determines the optimal parameter combination by minimizing the correlation between each modal function and maximizing the sparsity of each modal function.

3. The hydroelectric oil filter fault diagnosis system based on blind source separation according to claim 2 is characterized in that: The expert diagnosis layer of the dynamic source number estimation unit runs multiple source number estimation algorithms in parallel, including an improved Akaike information criterion algorithm, a Bayesian information criterion algorithm, and an algorithm based on random matrix theory. The Bayesian evidence voting mechanism assigns weights according to the historical accuracy of each algorithm. The physical constraint layer of the dynamic source number estimation unit, when performing comparison and verification, if the fault combination corresponding to the candidate source number does not exist in the fault mode and effect analysis knowledge base or is physically not valid, the confidence of the candidate source number is downgraded or rejected, and the algorithm is triggered to re-estimate.

4. The hydroelectric oil filter fault diagnosis system based on blind source separation according to claim 3 is characterized by: The state-driven dynamic adjustment of the online blind source separation unit is specifically as follows: when the dynamic source number estimation unit determines that the system state is stable, the forgetting factor is increased to close to 1 or the learning rate is reduced to enhance the stability of the model; when it is determined that the system state has a sudden change, such as an increase in the number of source signals, the forgetting factor is reduced or the learning rate is increased to accelerate the learning of new fault features.

5. The hydroelectric oil filter fault diagnosis system based on blind source separation according to claim 2 is characterized in that: In the multi-physical quantity deep fusion unit, the loss function L of the physical information autoencoder is composed of the reconstruction loss L rec and physical constraint loss L phy Weighted composition: L=L rec +λL phy , where λ is a weight coefficient, and the physical constraint loss is constructed based on the known physical relationship between pressure and vibration; The visualization method of the application module includes: on the same interface, in the form of a dynamic time domain waveform diagram and a spectrum waterfall diagram, comparing and displaying the original mixed observation signal and the independent source signals decoupled by the online blind source separation unit, and providing a real-time visualization display of the blind source separation process.

6. The hydroelectric oil filter fault diagnosis system based on blind source separation according to claim 5 is characterized in that: The system adopts a microservice architecture, decoupling data acquisition, preprocessing, diagnostic engine units, and visualization functions into independent services. The services communicate through RESTful APIs and provide a unified data access layer that supports OPC UA and Modbus-TCP industrial protocols.

7. A fault diagnosis method for a hydroelectric oil filter based on blind source separation is characterized by: The following steps are involved: S1. Data Collection Steps: Synchronously collect multi-source heterogeneous observation signals, including vibration, pressure, temperature, and oil quality, during operation of the hydropower unit's online oil filter. S2. Adaptive preprocessing: Adaptive variational mode decomposition (VMD) based on an intelligent optimization algorithm is employed. Particle swarm optimization (PSO) or gray wolf optimization (GWA) is used to automatically optimize key VMD parameters. This adaptively decomposes the multi-source heterogeneous observation signals into multiple intrinsic mode functions (IMFs) to achieve noise reduction and feature enhancement. S3. Deep fusion step: First, extract the physical feature vectors of kurtosis, energy entropy, pressure pulsation amplitude, and temperature rise rate from the preprocessed multi-source heterogeneous observation signals. Then, using a physical information autoencoder, we learn these physical feature vectors by introducing physical laws as constraints in the loss function, generating a deeply fused high-dimensional feature matrix. S4. Source number estimation step: A three-layer hierarchical algorithm is used to estimate the number of independent source signals online. First, a statistical indicator, the residual energy ratio of the covariance matrix, is monitored to determine whether the system state has changed. If so, multiple estimation algorithms are run in parallel and a Bayesian evidence voting mechanism is used to output candidate source numbers. Finally, this candidate source number is compared and verified with a pre-set FMEA knowledge base to determine the final number of source signals. S5. Initial parameter loading step: When the system starts or the operating conditions change significantly, the optimal separation matrix that best matches the current operating conditions is matched and loaded from the pre-built digital twin model library. This matrix serves as the initial parameters for the subsequent online blind source separation algorithm. S6. Online separation step: Based on the number of source signals and the initial parameters loaded, running an online blind source separation algorithm based on independent component analysis, decoupling the high-dimensional feature matrix, and separating each independent source signal; S7 dynamic update step: During the operation of the online separation step S6, the system state of the source number estimation step S4 output is adjusted in real time and dynamically to adjust the learning rate or forgetting factor of the online blind source separation algorithm; S8. Diagnosis output step: identifying a specific fault mode based on the source signal and outputting a diagnosis result.

8. The method for diagnosing faults of a hydroelectric oil filter based on blind source separation according to claim 7, characterized in that: In the adaptive preprocessing step S2, the fitness function of the intelligent optimization algorithm evaluates the decomposition effect by minimizing the correlation between each modal function and maximizing the sparsity of each modal function, and determines the optimal modal number K and penalty factor α; in the deep fusion step S3, the loss function L of the physical information autoencoder is composed of the reconstruction loss L rec and physical constraint loss L phy Weighted composition: L=L rec +λL phy , where λ is a weight coefficient, and the physical constraint loss is constructed based on the known physical relationship between pressure and vibration.

9. The method for diagnosing faults of a hydroelectric oil filter based on blind source separation according to claim 7, characterized in that: In the source number estimation step S4, the covariance matrix residual energy ratio is the ratio of the covariance matrix residual energy of the high-dimensional feature matrix in the sliding time window to the total energy. When the ratio continuously exceeds a preset threshold, the expert diagnosis layer is triggered. The dynamic update step S7 is specifically as follows: when step S4 determines that the number of system source signals increases, the forgetting factor of the online blind source separation algorithm is immediately reduced to accelerate the convergence of the signal characteristics of the newly appeared fault source; When the system state is stable, increase the forgetting factor to close to 1 to enhance the model stability.

10. The fault diagnosis method for a hydroelectric oil filter based on blind source separation according to claim 7, characterized in that: The method also includes a preliminary step of constructing the digital twin model library, which includes: establishing a high-fidelity digital twin model of the online oil filter; simulating various single and compound faults in the digital twin model and collecting simulation data; using the simulation data for offline training and optimization to obtain the optimal separation matrix corresponding to different fault combinations, and storing it in the digital twin model library.

Images